diff --git a/UPDATING.md b/UPDATING.md index e1012be14d1b..3e0bca7d83ef 100644 --- a/UPDATING.md +++ b/UPDATING.md @@ -40,6 +40,143 @@ Importing a dataset now validates the `catalog` field against the target databas If you relied on importing datasets with a non-default catalog, enable "Allow changing catalogs" on the target connection, or set the dataset's catalog to the connection's default before importing. +### Entity version history for charts, dashboards, and datasets + +Saves of charts, dashboards, and datasets now automatically produce a version history — browsable and restorable via new API endpoints. No frontend UI in this release; the backend plumbing is the deliverable. + +**New endpoints** (per entity type — same pattern for `chart`, `dashboard`, and `dataset`): + +| Method | Path | Purpose | +|---|---|---| +| `GET` | `/api/v1/{resource}//versions/` | List the entity's version history (0-based `version_number`, `version_uuid`, `issued_at`, `changed_by`) | +| `GET` | `/api/v1/{resource}//versions//` | Get a single version snapshot (scalar fields at that version; plus `columns` / `metrics` for datasets) | +| `POST` | `/api/v1/{resource}//versions//restore` | Restore the entity to the state captured by that version | + +`` is a deterministic `UUIDv5` derived from the entity's UUID and the Continuum transaction id — stable across replicas and retention pruning. Authorisation reuses the resource's existing FAB permissions: list/get require `can_read`; restore requires `can_write`. Object-level access is enforced via `security_manager.raise_for_access`, so viewers who can see the entity can also see its history; only writers can restore. + +**Version response shape — `changes` array:** + +Each entry returned by `GET /api/v1/{resource}//versions/` and `GET .../versions//` includes a `changes` array describing what changed relative to the previous version: + +```json +"changes": [ + {"kind": "field", "path": "slice_name", "from_value": "Old", "to_value": "New"} +] +``` + +The array is empty for baseline (`operation_type=0`) transactions. `kind` enumerates structured record types (`field`, layout-walker records for dashboards, dataset child diffs for `TableColumn` / `SqlMetric`); `path` is a dotted JSON-pointer-style locator; `from_value` / `to_value` are JSON-safe scalars or compact records. + +**Save-response and ETag headers:** + +- Save responses (`PUT /api/v1/{resource}/`) include `old_version_uuid` and `new_version_uuid` body fields so the client can correlate a save with its resulting version row. +- All entity GETs (`GET /api/v1/{chart,dashboard,dataset}/`), version-list GETs, single-version GETs, and save responses emit an `ETag: ""` header reflecting the entity's current live version. The default `CORS_OPTIONS` now sets `expose_headers: ["ETag"]` so cross-origin browser clients can read the header. **No `If-Match` enforcement in v1** — `ETag` is informational; concurrent-edit detection is deferred to a follow-up SIP. +- **Operators overriding `CORS_OPTIONS` in `superset_config.py` MUST include `"expose_headers": ["ETag"]`** (or merge with the default) for cross-origin clients to read the ETag. A bare `CORS_OPTIONS = {"origins": [...]}` will silently drop the expose-headers default. + +**Behaviour changes on save:** + +- Every save of a chart, dashboard, or dataset produces one new version row. Rows preserve the full post-save state (scalar fields for all three entity types; `TableColumn` / `SqlMetric` children for datasets; `dashboard_slices` chart membership for dashboards — children versioned via SQLAlchemy-Continuum shadow tables `table_columns_version`, `sql_metrics_version`, and `dashboard_slices_version`). +- First save after an entity already exists in the DB creates a retroactive baseline version so the UI can show "what this looked like before I edited it." +- Tags, owners, and roles are **not** versioned in v1 (ADR-005). A restore leaves those at their live values. + +**New config keys:** + +| Key | Default | Purpose | +|---|---|---| +| `ENABLE_VERSIONING_CAPTURE` | `True` | Master switch for the two before-flush listeners that drive version capture. Default-on; set to `False` in `superset_config.py` (or via the env var of the same name) for an operational kill-switch — when a versioning-induced save-path regression needs a 30-second recovery (restart workers, capture stops) instead of revert + redeploy. Existing shadow tables stay; `/versions/` and `/activity/` endpoints continue to work read-only against captured history. New deployments leave it on. | +| `SUPERSET_VERSION_HISTORY_RETENTION_DAYS` | `30` | Versions older than this many days are pruned by a nightly Celery beat task (`superset.tasks.version_history_retention.prune_old_versions`). Each entity's live row (`end_transaction_id IS NULL`) is always preserved; closed historical rows including the baseline age out with the rest. Set to `0` to disable retention entirely. | + +**Impact on external integrations:** + +- New tables populated on every save — `dashboards_version`, `slices_version`, `tables_version` (parent shadow tables for the three entity types), `table_columns_version`, `sql_metrics_version`, `dashboard_slices_version` (child shadow tables), plus the shared `version_transaction` and `version_changes` tables. External tooling that queries Superset's DB directly will see writes to these tables proportional to save traffic. +- On MySQL, the large-payload shadow columns (`dashboards_version.{position_json,css,json_metadata}`, `slices_version.params`, `tables_version.sql`, `{table_columns,sql_metrics}_version.{description,expression}`) are declared `MEDIUMTEXT` to match their live counterparts (16 MB) — Postgres `TEXT` is unbounded and SQLite ignores the length. Operators inspecting the schema will see this dialect-specific type; no operator action is required for new deployments. +- Existing entity endpoints (`GET`/`PUT /api/v1/{chart,dashboard,dataset}/`) gain an `ETag` response header and the save response gains `old_version_uuid` / `new_version_uuid` body fields. No existing fields are removed or repurposed. +- Version capture is on by default but operationally disable-able via `ENABLE_VERSIONING_CAPTURE=False` — an escape hatch for capture-induced regressions, not a feature flag. The migrations and the endpoints are not gated; only the listeners that write new shadow rows on save. + +**Querying the shadow tables — audit columns are frozen at capture time:** + +The parent shadow tables (`dashboards_version`, `slices_version`, `tables_version`) deliberately exclude the audit columns `changed_on`, `created_on`, `changed_by_fk`, and `created_by_fk` from version capture. The "who changed this version, and when?" facts live on `version_transaction.user_id` and `version_transaction.issued_at` instead — every shadow row carries a `transaction_id` FK to that row. + +Consequence for external tooling: a query that joins a shadow table to `ab_user` via `changed_by_fk` (e.g. `SELECT u.username FROM dashboards_version v JOIN ab_user u ON v.changed_by_fk = u.id`) returns whatever audit metadata was captured at *baseline* time — typically stale or null — not the user who produced the version. The correct join is through the transaction row: + +```sql +SELECT u.username, t.issued_at, v.dashboard_title +FROM dashboards_version v +JOIN version_transaction t ON v.transaction_id = t.id +LEFT JOIN ab_user u ON t.user_id = u.id +``` + +The exclusion is deliberate (the audit columns would otherwise grow proportional to save count with redundant data) — but operators writing reports against the shadow tables need to know which join carries the version's authorship. + +**Behavior change — `ImportExportMixin.reset_ownership`:** + +The ownership-reset helper used by every import/clone/duplicate path was rewritten so that when a Flask user is present in `g.user`, `created_by` and `changed_by` are assigned to that user explicitly. Previously the helper left both fields `None` and relied on the FAB column default to backfill at flush time. The new shape was forced by the versioning capture path: when Continuum-attached relationships are present, the `None` propagates through to the FK and suppresses the column default, leaving the imported entity with no recorded author. + +The behavior change applies to **every** `ImportModelsCommand` / `CopyDashboardCommand` / `DuplicateDatasetCommand` invocation, not just versioning-adjacent ones. Operators who notice imported entities now consistently carry the importing user as `created_by` / `changed_by` (where previously some imports landed with `None` audit fields under specific FAB session configurations) are seeing this change. + +### Cross-entity activity stream for charts, dashboards, and datasets + +A read-only companion to the version-history endpoints (above). Each entity type gains an `/activity/` endpoint that returns a chronological stream of edits — the entity's own edits plus, for dashboards and charts, transitive edits to related entities during their association windows. + +**New endpoints** (per entity type): + +| Method | Path | Purpose | +|---|---|---| +| `GET` | `/api/v1/dashboard//activity/` | Dashboard own edits + edits to charts attached during their dashboard window + edits to datasets those charts pointed at during their chart window | +| `GET` | `/api/v1/chart//activity/` | Chart own edits + edits to datasets the chart pointed at during association | +| `GET` | `/api/v1/dataset//activity/` | Dataset own edits only (no transitive layer in V2) | + +**Query parameters** (all optional): + +| Param | Type | Default | Purpose | +|---|---|---|---| +| `since` | ISO 8601 datetime | — | Lower bound on `issued_at` | +| `until` | ISO 8601 datetime | — | Upper bound on `issued_at` | +| `include` | `self` \| `related` \| `all` | `all` | Filter to only the entity's own edits, only related edits, or both | +| `page` | integer ≥ 0 | `0` | 0-based page index | +| `page_size` | integer in `[1, 200]` | `25` | Records per page (clamped silently to 200) | + +**Response shape:** + +```json +{ + "result": [ + { + "version_uuid": "...", + "entity_kind": "chart", + "entity_uuid": "...", + "entity_name": "Top 10 Girls", + "entity_deleted": false, + "entity_deletion_state": null, + "source": "related", + "transaction_id": 1234, + "issued_at": "2026-05-26T12:00:00", + "changed_by": {"id": 5, "first_name": "Mike", "last_name": "Bridge"}, + "kind": "filter", + "path": ["params", "adhoc_filters", "country"], + "from_value": null, + "to_value": "US", + "summary": "Chart filter changed: Top 10 Girls", + "impact": null + } + ], + "count": 47 +} +``` + +`count` is the total record count *after* the silent permission filter (see below), not the raw query size. + +**Authorisation:** reuses the resource's existing `can_read` permission. The endpoint runs `security_manager.raise_for_access(=path_entity)` — users without read access to the path entity get `403`. Workspace admins can read any entity's activity stream. + +**Silent permission filter (AV-008):** records whose source entity the requesting user can't read are silently dropped — no placeholder, no count contribution. The frontend cannot distinguish "no activity" from "you can't see this activity." + +**Tombstones (AV-009 / D-15):** when an activity record references a hard-deleted source entity, the record still appears with `entity_deleted: true`, `entity_uuid: null`, and `entity_name` recovered from the last shadow row. + +**Impact on external integrations:** + +- Pure read-only. No new tables, no new columns, no migrations. Reads sc-103156's shadow tables and the `version_changes` table. +- No new save-path code paths — perf-validation gate confirms the activity-view branch does not regress sc-103156's SC-004 50ms-overhead budget. +- No feature flag; the endpoints are always available once sc-103156's version-history feature is enabled. + ### Granular Export Controls A new feature flag `GRANULAR_EXPORT_CONTROLS` introduces three fine-grained permissions that replace the legacy `can_csv` permission: @@ -326,6 +463,246 @@ See `superset/mcp_service/PRODUCTION.md` for deployment guides. } ``` +### Composite primary keys on many-to-many association tables + +The eight M:N association tables listed below have been changed from a synthetic surrogate `id INTEGER PRIMARY KEY` to a composite `PRIMARY KEY (fk1, fk2)` on the two foreign-key columns. The `id` column is dropped, and the two tables that previously carried a redundant `UNIQUE (fk1, fk2)` constraint have that constraint removed (it is now subsumed by the composite primary key). + +**Affected tables and their composite-PK column pairs:** + +| Table | Composite PK | +|---|---| +| `dashboard_roles` | `(dashboard_id, role_id)` | +| `dashboard_slices` | `(dashboard_id, slice_id)` | +| `dashboard_user` | `(user_id, dashboard_id)` | +| `report_schedule_user` | `(user_id, report_schedule_id)` | +| `rls_filter_roles` | `(role_id, rls_filter_id)` | +| `rls_filter_tables` | `(table_id, rls_filter_id)` | +| `slice_user` | `(user_id, slice_id)` | +| `sqlatable_user` | `(user_id, table_id)` | + +**Impact on external readers:** Any BI tool, custom report, backup script, or external integration that references these tables by their old surrogate `id` column (e.g., `SELECT id FROM dashboard_slices WHERE …`, `WHERE dashboard_slices.id IN (…)`) will break. Update such queries to project or filter on the FK pair (`dashboard_id, slice_id`) instead. The FK columns themselves are unchanged. + +**Pre-flight inventory queries.** Before applying the upgrade, operators are encouraged to run the queries below against their database to assess what the migration will change. Two classes of pre-existing data are not preserved by the migration: duplicate `(fk1, fk2)` rows (the migration keeps `MIN(id)` and deletes the rest) and rows with `NULL` in either FK column (the migration deletes them, since FK columns are promoted to `NOT NULL` for the composite PK). Compliance- or audit-sensitive operators should also `\copy` (Postgres) or `SELECT … INTO OUTFILE` (MySQL) the affected rows for their own records before upgrading. + +```sql +-- Duplicate (fk1, fk2) pairs (the migration will keep MIN(id) per group, delete the rest) +SELECT dashboard_id, role_id, COUNT(*) FROM dashboard_roles GROUP BY dashboard_id, role_id HAVING COUNT(*) > 1; +SELECT dashboard_id, slice_id, COUNT(*) FROM dashboard_slices GROUP BY dashboard_id, slice_id HAVING COUNT(*) > 1; +SELECT user_id, dashboard_id, COUNT(*) FROM dashboard_user GROUP BY user_id, dashboard_id HAVING COUNT(*) > 1; +SELECT user_id, report_schedule_id, COUNT(*) FROM report_schedule_user GROUP BY user_id, report_schedule_id HAVING COUNT(*) > 1; +SELECT role_id, rls_filter_id, COUNT(*) FROM rls_filter_roles GROUP BY role_id, rls_filter_id HAVING COUNT(*) > 1; +SELECT table_id, rls_filter_id, COUNT(*) FROM rls_filter_tables GROUP BY table_id, rls_filter_id HAVING COUNT(*) > 1; +SELECT user_id, slice_id, COUNT(*) FROM slice_user GROUP BY user_id, slice_id HAVING COUNT(*) > 1; +SELECT user_id, table_id, COUNT(*) FROM sqlatable_user GROUP BY user_id, table_id HAVING COUNT(*) > 1; + +-- Rows with a NULL in either FK (the migration will delete these) +SELECT COUNT(*) FROM dashboard_roles WHERE dashboard_id IS NULL OR role_id IS NULL; +SELECT COUNT(*) FROM dashboard_slices WHERE dashboard_id IS NULL OR slice_id IS NULL; +SELECT COUNT(*) FROM dashboard_user WHERE user_id IS NULL OR dashboard_id IS NULL; +SELECT COUNT(*) FROM report_schedule_user WHERE user_id IS NULL OR report_schedule_id IS NULL; +SELECT COUNT(*) FROM rls_filter_roles WHERE role_id IS NULL OR rls_filter_id IS NULL; +SELECT COUNT(*) FROM rls_filter_tables WHERE table_id IS NULL OR rls_filter_id IS NULL; +SELECT COUNT(*) FROM slice_user WHERE user_id IS NULL OR slice_id IS NULL; +SELECT COUNT(*) FROM sqlatable_user WHERE user_id IS NULL OR table_id IS NULL; +``` + +**Sizing the maintenance window on PostgreSQL.** The queries above are dialect-portable but only count rows. Operators on PostgreSQL can run the diagnostic queries below to characterize the migration's runtime cost ahead of time: per-table row count and on-disk size, an aggregated duplicate roll-up, the external-FK pre-flight check (the migration runs the same check and aborts if it returns rows), and a rewrite-time estimate for the two tables that go through the slower full-table-rebuild path. + +```sql +-- Per-table size, row count, and which migration path each will take. +-- Two tables ("dashboard_slices", "report_schedule_user") have a +-- redundant UNIQUE constraint that the migration drops via a full +-- table rewrite (op.batch_alter_table(recreate="always")). The other +-- six use direct ALTER TABLE, which is much cheaper. +WITH affected(name, has_unique) AS ( + VALUES + ('dashboard_roles', false), + ('dashboard_slices', true), + ('dashboard_user', false), + ('report_schedule_user', true), + ('rls_filter_roles', false), + ('rls_filter_tables', false), + ('slice_user', false), + ('sqlatable_user', false) +) +SELECT + a.name AS table_name, + CASE WHEN a.has_unique THEN 'recreate (full rewrite)' + ELSE 'direct ALTER' END AS migration_path, + c.reltuples::bigint AS estimated_rows, + pg_size_pretty(pg_total_relation_size(c.oid)) AS total_size, + pg_size_pretty(pg_relation_size(c.oid)) AS heap_size, + pg_size_pretty(pg_indexes_size(c.oid)) AS index_size +FROM affected a +JOIN pg_class c ON c.relname = a.name AND c.relkind = 'r' +ORDER BY pg_total_relation_size(c.oid) DESC; +``` + +```sql +-- Aggregated duplicate-row roll-up. +-- "dup_groups" is the number of (fk1, fk2) pairs that appear more +-- than once; "rows_dropped" is the total number of rows the +-- migration will delete during the dedupe pass (it keeps MIN(id) per +-- group and discards the rest). +SELECT 'dashboard_roles' AS t, COUNT(*) AS dup_groups, SUM(c) - COUNT(*) AS rows_dropped + FROM (SELECT COUNT(*) c FROM dashboard_roles GROUP BY dashboard_id, role_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'dashboard_slices', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM dashboard_slices GROUP BY dashboard_id, slice_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'dashboard_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM dashboard_user GROUP BY user_id, dashboard_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'report_schedule_user',COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM report_schedule_user GROUP BY user_id, report_schedule_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'rls_filter_roles', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM rls_filter_roles GROUP BY role_id, rls_filter_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'rls_filter_tables', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM rls_filter_tables GROUP BY table_id, rls_filter_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'slice_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM slice_user GROUP BY user_id, slice_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'sqlatable_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM sqlatable_user GROUP BY user_id, table_id HAVING COUNT(*) > 1) g +ORDER BY rows_dropped DESC NULLS LAST; +``` + +```sql +-- External-FK pre-flight check. +-- The migration runs the equivalent check at upgrade time and aborts +-- if any external FK references one of the soon-to-be-removed `id` +-- columns. Running it ahead of time lets you discover (and migrate) +-- any such reference before the maintenance window. On a stock +-- Superset install this should return zero rows. (Default schema +-- only; multi-schema deployments need to broaden the lookup.) +SELECT + rc.constraint_name, + kcu.table_schema || '.' || kcu.table_name AS referencing_table, + kcu.column_name AS referencing_column, + ccu.table_name AS referenced_table, + ccu.column_name AS referenced_column +FROM information_schema.referential_constraints rc +JOIN information_schema.key_column_usage kcu + ON kcu.constraint_name = rc.constraint_name + AND kcu.constraint_schema = rc.constraint_schema +JOIN information_schema.constraint_column_usage ccu + ON ccu.constraint_name = rc.constraint_name + AND ccu.constraint_schema = rc.constraint_schema +WHERE ccu.table_name IN ( + 'dashboard_roles','dashboard_slices','dashboard_user', + 'report_schedule_user','rls_filter_roles','rls_filter_tables', + 'slice_user','sqlatable_user') + AND ccu.column_name = 'id'; +``` + +```sql +-- Lock-window estimate for the two full-rewrite tables. +-- recreate="always" takes ACCESS EXCLUSIVE on the table for the full +-- rewrite. Use heap size combined with your hardware's effective +-- write throughput (~100-200 MB/s on commodity SSD; faster on NVMe) +-- to size the maintenance window. The other six tables use direct +-- ALTER and are dominated by composite-index build time, typically +-- seconds for tables in the low millions of rows. +SELECT + c.relname AS table_name, + pg_size_pretty(pg_relation_size(c.oid)) AS heap_size, + pg_relation_size(c.oid) / 1024 / 1024 AS heap_size_mb, + ROUND(pg_relation_size(c.oid) / 1024 / 1024 / 100.0, 1) AS est_rewrite_seconds_at_100mbs +FROM pg_class c +WHERE c.relname IN ('dashboard_slices', 'report_schedule_user'); +``` + +**Sizing the maintenance window on MySQL.** Equivalent diagnostic queries for MySQL/InnoDB. One important difference from PostgreSQL: InnoDB rebuilds the clustered index on every PK change, so *all eight* tables undergo a full table rebuild on MySQL — not just the two that go through the explicit `recreate="always"` path. The lock-window estimate query below therefore covers all eight tables. + +```sql +-- Per-table size, row count, and which migration path each will take. +-- TABLE_ROWS is an InnoDB estimate (analogous to PostgreSQL's reltuples); +-- run SELECT COUNT(*) per table for an exact count if needed. +SELECT + TABLE_NAME AS table_name, + CASE WHEN TABLE_NAME IN ('dashboard_slices', 'report_schedule_user') + THEN 'recreate (explicit, drops UNIQUE)' + ELSE 'direct ALTER (still rebuilds InnoDB clustered index)' + END AS migration_path, + TABLE_ROWS AS estimated_rows, + CONCAT(ROUND((DATA_LENGTH + INDEX_LENGTH) / 1024 / 1024, 1), ' MB') AS total_size, + CONCAT(ROUND(DATA_LENGTH / 1024 / 1024, 1), ' MB') AS heap_size, + CONCAT(ROUND(INDEX_LENGTH / 1024 / 1024, 1), ' MB') AS index_size +FROM information_schema.TABLES +WHERE TABLE_SCHEMA = DATABASE() + AND TABLE_NAME IN ( + 'dashboard_roles', 'dashboard_slices', 'dashboard_user', + 'report_schedule_user', 'rls_filter_roles', 'rls_filter_tables', + 'slice_user', 'sqlatable_user' + ) +ORDER BY (DATA_LENGTH + INDEX_LENGTH) DESC; +``` + +```sql +-- Aggregated duplicate-row roll-up. Same SQL as the PostgreSQL version +-- (standard SQL); included here for copy-paste convenience. +SELECT 'dashboard_roles' AS t, COUNT(*) AS dup_groups, SUM(c) - COUNT(*) AS rows_dropped + FROM (SELECT COUNT(*) c FROM dashboard_roles GROUP BY dashboard_id, role_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'dashboard_slices', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM dashboard_slices GROUP BY dashboard_id, slice_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'dashboard_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM dashboard_user GROUP BY user_id, dashboard_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'report_schedule_user',COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM report_schedule_user GROUP BY user_id, report_schedule_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'rls_filter_roles', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM rls_filter_roles GROUP BY role_id, rls_filter_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'rls_filter_tables', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM rls_filter_tables GROUP BY table_id, rls_filter_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'slice_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM slice_user GROUP BY user_id, slice_id HAVING COUNT(*) > 1) g +UNION ALL SELECT 'sqlatable_user', COUNT(*), SUM(c) - COUNT(*) + FROM (SELECT COUNT(*) c FROM sqlatable_user GROUP BY user_id, table_id HAVING COUNT(*) > 1) g +ORDER BY rows_dropped DESC; +``` + +```sql +-- External-FK pre-flight check. KEY_COLUMN_USAGE on MySQL carries +-- both sides of the FK in a single row, so this is simpler than the +-- PostgreSQL version. Should return zero rows on a stock install. +SELECT + CONSTRAINT_NAME, + CONCAT(TABLE_SCHEMA, '.', TABLE_NAME) AS referencing_table, + COLUMN_NAME AS referencing_column, + REFERENCED_TABLE_NAME AS referenced_table, + REFERENCED_COLUMN_NAME AS referenced_column +FROM information_schema.KEY_COLUMN_USAGE +WHERE TABLE_SCHEMA = DATABASE() + AND REFERENCED_TABLE_NAME IN ( + 'dashboard_roles', 'dashboard_slices', 'dashboard_user', + 'report_schedule_user', 'rls_filter_roles', 'rls_filter_tables', + 'slice_user', 'sqlatable_user' + ) + AND REFERENCED_COLUMN_NAME = 'id'; +``` + +```sql +-- Lock-window estimate for ALL EIGHT tables (InnoDB rebuilds the +-- clustered index on PK change, so even "direct ALTER" is a rewrite). +-- ADD PRIMARY KEY is INPLACE but not LOCK=NONE — it allows concurrent +-- reads but blocks writes. Use heap size combined with your effective +-- rebuild throughput (~100-200 MB/s on commodity SSD; higher on NVMe). +SELECT + TABLE_NAME AS table_name, + CONCAT(ROUND(DATA_LENGTH / 1024 / 1024, 1), ' MB') AS heap_size, + ROUND(DATA_LENGTH / 1024 / 1024, 1) AS heap_size_mb, + ROUND(DATA_LENGTH / 1024 / 1024 / 100.0, 1) AS est_rewrite_seconds_at_100mbs +FROM information_schema.TABLES +WHERE TABLE_SCHEMA = DATABASE() + AND TABLE_NAME IN ( + 'dashboard_roles', 'dashboard_slices', 'dashboard_user', + 'report_schedule_user', 'rls_filter_roles', 'rls_filter_tables', + 'slice_user', 'sqlatable_user' + ) +ORDER BY DATA_LENGTH DESC; +``` + +**Restoring an old `pg_dump` (or equivalent) against the new schema.** A dump taken before the migration includes `INSERT` statements that populate the now-removed `id` column. Restoring such a dump against the post-migration schema will fail. The supported workaround is to dump only the schema and reference data, then re-create the M:N associations from application data after restore — for example with `pg_dump --exclude-table-data` (or per-table `--exclude-table-data=dashboard_slices` etc.) for the eight junction tables, restore the rest, then run a one-shot script that re-INSERTs `(fk1, fk2)` pairs derived from your application export. Operators who need to restore an old dump verbatim should restore against a pre-migration Superset and then re-run the upgrade. + +**Intentional downgrade asymmetry.** The migration's `downgrade()` restores the surrogate `id` column and (for `dashboard_slices` and `report_schedule_user`) the original `UNIQUE (fk1, fk2)` constraint, but it does **not** restore the original `NULL`-allowed state on the FK columns — they remain `NOT NULL`. This is intentional: under SQLAlchemy's `secondary=` semantics, a `NULL` in either FK column of a junction table is meaningless (it cannot participate in either side of the relationship). Operators downgrading are not expected to need this restored. The asymmetry is documented for completeness so that round-trip schema diffs are not mistaken for migration bugs. + +**Constraint-name divergence between upgrade and downgrade.** The composite primary key created on upgrade is named `pk_` (Alembic's default for `op.create_primary_key("pk_
", ...)`), while the surrogate `id` primary key restored on downgrade is named `
_pkey` (PostgreSQL's default convention for `PrimaryKeyConstraint("id")`). The two names alternate so that a round-trip (upgrade → downgrade → upgrade) does not collide on a pre-existing constraint name. Operators using schema-comparison tools (e.g. `pg_diff`, `migra`) against a downgraded database may see this as drift versus a fresh-install schema. It is cosmetic — no application code references either constraint name. + ## 6.0.0 - [33055](https://github.com/apache/superset/pull/33055): Upgrades Flask-AppBuilder to 5.0.0. The AUTH_OID authentication type has been deprecated and is no longer available as an option in Flask-AppBuilder. OpenID (OID) is considered a deprecated authentication protocol - if you are using AUTH_OID, you will need to migrate to an alternative authentication method such as OAuth, LDAP, or database authentication before upgrading. - [34871](https://github.com/apache/superset/pull/34871): Fixed Jest test hanging issue from Ant Design v5 upgrade. MessageChannel is now mocked in test environment to prevent rc-overflow from causing Jest to hang. Test environment only - no production impact. diff --git a/docker-compose-mysql.yml b/docker-compose-mysql.yml new file mode 100644 index 000000000000..13f4c99236cb --- /dev/null +++ b/docker-compose-mysql.yml @@ -0,0 +1,117 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +# +# Compose override that swaps the default Postgres metadata DB for MySQL 8. +# Useful for evaluating dialect-specific behaviour (e.g., DDL-migration +# cost on a deployment whose production metadata DB is MySQL). +# +# Usage: +# docker compose -f docker-compose.yml -f docker-compose-mysql.yml up +# docker compose -f docker-compose.yml -f docker-compose-mysql.yml down +# +# To switch back to Postgres, just drop the second `-f` flag — the MySQL +# data lives in a separate volume (`db_home_mysql`) so neither side is +# corrupted by switching dialects. +# +# Notes: +# - Mirrors the connection settings used by CI's `test-mysql` shard: +# dialect ``mysql+mysqldb``, charset utf8mb4 with binary_prefix. +# - Host port 13306 (configurable via DATABASE_PORT_MYSQL) to avoid +# colliding with a native MySQL install on 3306. +# - The Postgres-specific init scripts under +# docker/docker-entrypoint-initdb.d/ are not mounted (they are +# postgres-only); examples / cypress fixtures still load via +# `superset-init`'s post-startup steps. + +# Shared environment override applied to every Superset-side service that +# connects to the metadata DB. ``environment:`` takes precedence over the +# values inherited from the env_file in docker-compose.yml. +x-mysql-env: &mysql-env + DATABASE_DIALECT: mysql+mysqldb + DATABASE_HOST: db + DATABASE_PORT: "3306" + DATABASE_DB: superset + DATABASE_USER: superset + DATABASE_PASSWORD: superset + SQLALCHEMY_DATABASE_URI: "mysql+mysqldb://superset:superset@db:3306/superset?charset=utf8mb4&binary_prefix=true" + # Override the analytics-examples DB connection too. ``EXAMPLES_PORT`` + # in docker/.env is hardcoded to 5432 (the Postgres port); without + # this override the examples connection would try MySQL on 5432 and + # fail. The examples user/DB are created by docker/mysql-init/ + # examples-init.sql on first MySQL boot. + EXAMPLES_HOST: db + EXAMPLES_PORT: "3306" + EXAMPLES_DB: examples + EXAMPLES_USER: examples + EXAMPLES_PASSWORD: examples + SUPERSET__SQLALCHEMY_EXAMPLES_URI: "mysql+mysqldb://examples:examples@db:3306/examples?charset=utf8mb4&binary_prefix=true" + +services: + db: + image: mysql:8.0 + environment: + MYSQL_DATABASE: superset + MYSQL_USER: superset + MYSQL_PASSWORD: superset + MYSQL_ROOT_PASSWORD: root + # The original 5432 port mapping is harmless on a MySQL container + # (nothing listens on 5432 inside it) but we add 13306->3306 so the + # MySQL port is reachable from the host without colliding with a + # native MySQL on 3306. Compose merges port lists. + ports: + - "127.0.0.1:${DATABASE_PORT_MYSQL:-13306}:3306" + # Override the init-scripts mount by re-binding the same target path + # to a MySQL-compatible directory. Compose merges volume lists by + # target path; later definitions win on conflict, so this displaces + # the Postgres-specific ``./docker/docker-entrypoint-initdb.d`` mount + # from docker-compose.yml. Without this, MySQL would try to run + # ``cypress-init.sh`` (which invokes ``psql``, not in the MySQL + # image), abort the init phase, and never create the ``examples`` + # database. Add the MySQL data volume separately. + volumes: + - db_home_mysql:/var/lib/mysql + - ./docker/mysql-init:/docker-entrypoint-initdb.d + command: + - --default-authentication-plugin=caching_sha2_password + - --character-set-server=utf8mb4 + - --collation-server=utf8mb4_0900_ai_ci + healthcheck: + test: ["CMD-SHELL", "mysqladmin ping -h localhost -uroot -proot --silent"] + interval: 5s + timeout: 5s + retries: 20 + + superset: + environment: *mysql-env + + superset-init: + environment: *mysql-env + + superset-worker: + environment: *mysql-env + + superset-worker-beat: + environment: *mysql-env + + superset-node: + environment: *mysql-env + + superset-tests-worker: + environment: *mysql-env + +volumes: + db_home_mysql: diff --git a/docker/mysql-init/examples-init.sql b/docker/mysql-init/examples-init.sql new file mode 100644 index 000000000000..68dabe38671d --- /dev/null +++ b/docker/mysql-init/examples-init.sql @@ -0,0 +1,32 @@ +-- Licensed to the Apache Software Foundation (ASF) under one +-- or more contributor license agreements. See the NOTICE file +-- distributed with this work for additional information +-- regarding copyright ownership. The ASF licenses this file +-- to you under the Apache License, Version 2.0 (the +-- "License"); you may not use this file except in compliance +-- with the License. You may obtain a copy of the License at +-- +-- http://www.apache.org/licenses/LICENSE-2.0 +-- +-- Unless required by applicable law or agreed to in writing, +-- software distributed under the License is distributed on an +-- "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +-- KIND, either express or implied. See the License for the +-- specific language governing permissions and limitations +-- under the License. + +-- MySQL counterpart to docker/docker-entrypoint-initdb.d/examples-init.sh. +-- Creates the analytics-examples database and user that Superset's +-- ``load-examples`` command writes to. Mounted by docker-compose-mysql.yml +-- at /docker-entrypoint-initdb.d/ so the MySQL image's first-boot +-- entrypoint runs it automatically. (The Postgres init scripts under +-- docker/docker-entrypoint-initdb.d/ are NOT mounted on the MySQL +-- service — they invoke psql, which doesn't exist in the MySQL image.) + +CREATE DATABASE IF NOT EXISTS examples + CHARACTER SET utf8mb4 + COLLATE utf8mb4_0900_ai_ci; + +CREATE USER IF NOT EXISTS 'examples'@'%' IDENTIFIED BY 'examples'; +GRANT ALL PRIVILEGES ON examples.* TO 'examples'@'%'; +FLUSH PRIVILEGES; diff --git a/pyproject.toml b/pyproject.toml index dcc4b4f8c84b..0eb7fd0b02dd 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -100,6 +100,7 @@ dependencies = [ "simplejson>=3.15.0", "slack_sdk>=3.19.0, <4", "sqlalchemy>=1.4, <2", + "sqlalchemy-continuum>=1.6.0, <2.0.0", "sqlalchemy-utils>=0.38.0, <0.43", # expanding lowerbound to work with pydoris "sqlglot>=30.8.0, <31", # newer pandas needs 0.9+ diff --git a/requirements/base.txt b/requirements/base.txt index 2a0af7d9d4c9..25d80239e747 100644 --- a/requirements/base.txt +++ b/requirements/base.txt @@ -409,7 +409,10 @@ sqlalchemy==1.4.54 # flask-sqlalchemy # marshmallow-sqlalchemy # shillelagh + # sqlalchemy-continuum # sqlalchemy-utils +sqlalchemy-continuum==1.6.0 + # via apache-superset (pyproject.toml) sqlalchemy-utils==0.42.0 # via # apache-superset (pyproject.toml) diff --git a/requirements/development.txt b/requirements/development.txt index 120ee1ae6f63..b721731c4fa6 100644 --- a/requirements/development.txt +++ b/requirements/development.txt @@ -979,9 +979,14 @@ sqlalchemy==1.4.54 # marshmallow-sqlalchemy # shillelagh # sqlalchemy-bigquery + # sqlalchemy-continuum # sqlalchemy-utils sqlalchemy-bigquery==1.15.0 # via apache-superset +sqlalchemy-continuum==1.6.0 + # via + # -c requirements/base-constraint.txt + # apache-superset sqlalchemy-utils==0.42.0 # via # -c requirements/base-constraint.txt diff --git a/scripts/seed_junction_load.py b/scripts/seed_junction_load.py new file mode 100644 index 000000000000..5d72fa9da18a --- /dev/null +++ b/scripts/seed_junction_load.py @@ -0,0 +1,682 @@ +#!/usr/bin/env python3 +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +# +# ---------------------------------------------------------------------- +# Stress-test data generator for the composite-PK migration (sc-105349). +# +# Bulk-inserts synthetic parent rows and many-to-many junction rows for +# the eight association tables that the composite-PK migration touches. +# Useful for measuring migration runtime at varying scales — run this at +# 100K / 1M / 5M / 10M rows and time the migration at each scale to +# verify the O(N log N) extrapolation. +# +# Idempotent: rerunning with the same target is a no-op; rerunning with +# a higher target adds rows up to the new total. Batched bulk INSERTs +# (10K rows per statement) make it fast on Postgres, MySQL, and SQLite. +# +# Usage (inside the Superset container): +# +# docker exec superset-superset-1 \\ +# /app/.venv/bin/python /app/scripts/seed_junction_load.py \\ +# --dashboard-slices 1000000 \\ +# --slice-user 100000 \\ +# --dashboard-user 100000 +# +# Run with no flags for the defaults shown below. Use ``--dry-run`` to +# print the planned inserts without writing anything. +# +# The script connects via Superset's standard ``DATABASE_*`` env vars +# (or ``SUPERSET__SQLALCHEMY_DATABASE_URI`` if set), so it works +# automatically inside the Superset container regardless of which +# metadata DB backend is in use. + +from __future__ import annotations + +import argparse +import logging +import os +import sys +import time +from collections.abc import Iterator +from contextlib import contextmanager +from uuid import uuid4 + +import sqlalchemy as sa +from sqlalchemy.engine import Connection, Engine + +logger = logging.getLogger("seed_junction_load") + +# Bulk INSERT batch size. Larger values = fewer statements but more memory. +BATCH = 10_000 + +# Default per-junction-table target row counts. Tuned to mimic the shape +# of a large multi-team Superset install. Override via CLI flags. +DEFAULTS: dict[str, int] = { + "dashboard_slices": 1_000_000, + "slice_user": 100_000, + "dashboard_user": 100_000, + "dashboard_roles": 10_000, +} + +# (junction_table, fk1_col, fk2_col, parent1_table, parent2_table) +# parents reference id columns; we generate (fk1, fk2) pairs by sampling +# from the parents' existing IDs. +JUNCTIONS: list[tuple[str, str, str, str, str]] = [ + ("dashboard_slices", "dashboard_id", "slice_id", "dashboards", "slices"), + ("slice_user", "user_id", "slice_id", "ab_user", "slices"), + ("dashboard_user", "user_id", "dashboard_id", "ab_user", "dashboards"), + ("dashboard_roles", "dashboard_id", "role_id", "dashboards", "ab_role"), +] + +# Junction tables that originally carried ``UNIQUE(fk1, fk2)`` and therefore +# cannot accept duplicate ``(fk1, fk2)`` pairs even on the pre-migration +# (downgrade) schema. The other JUNCTIONS allow duplicates pre-migration. +JUNCTIONS_WITH_UNIQUE: set[str] = {"dashboard_slices", "report_schedule_user"} + + +# ---------------------------------------------------------------------- +# Connection setup +# ---------------------------------------------------------------------- + + +def build_engine() -> Engine: + """Build a SQLAlchemy engine from Superset env vars.""" + if uri := os.environ.get("SUPERSET__SQLALCHEMY_DATABASE_URI"): + logger.info("Using SUPERSET__SQLALCHEMY_DATABASE_URI from env") + return sa.create_engine(uri) + + try: + dialect = os.environ["DATABASE_DIALECT"] + user = os.environ["DATABASE_USER"] + password = os.environ["DATABASE_PASSWORD"] + host = os.environ["DATABASE_HOST"] + port = os.environ["DATABASE_PORT"] + db = os.environ["DATABASE_DB"] + except KeyError as exc: + sys.exit( + f"Missing env var {exc}; either set DATABASE_DIALECT/USER/PASSWORD/" + f"HOST/PORT/DB or SUPERSET__SQLALCHEMY_DATABASE_URI before running." + ) + + uri = f"{dialect}://{user}:{password}@{host}:{port}/{db}" + logger.info( + "Built URI from DATABASE_* env vars (dialect=%s, host=%s)", dialect, host + ) + return sa.create_engine(uri) + + +# ---------------------------------------------------------------------- +# Helpers +# ---------------------------------------------------------------------- + + +def uuid_value(dialect_name: str) -> bytes | str: + """Return a UUID in the form the active dialect expects. + + MySQL stores UUIDs as ``BINARY(16)`` (16 raw bytes); Postgres has a + native ``UUID`` type that accepts strings; SQLite stores them as + BLOB/TEXT and accepts either. Branching here keeps the seed script + backend-agnostic without depending on Superset's custom column types. + """ + if dialect_name.startswith("mysql"): + return uuid4().bytes + return str(uuid4()) + + +@contextmanager +def time_phase(name: str) -> Iterator[None]: + """Log elapsed wall time for a named phase.""" + start = time.monotonic() + logger.info("[%s] starting", name) + try: + yield + finally: + elapsed = time.monotonic() - start + logger.info("[%s] done in %.2fs", name, elapsed) + + +def count_rows(conn: Connection, table: str) -> int: + return conn.scalar(sa.text(f"SELECT COUNT(*) FROM {table}")) or 0 # noqa: S608 + + +def existing_ids(conn: Connection, table: str, limit: int | None = None) -> list[int]: + sql = f"SELECT id FROM {table} ORDER BY id" # noqa: S608 + if limit is not None: + sql += f" LIMIT {limit}" + return [row[0] for row in conn.execute(sa.text(sql))] + + +# ---------------------------------------------------------------------- +# Parent seeders +# +# Each function ensures the named parent table has at least ``target`` +# rows by inserting synthetic ones with minimal-but-valid columns. +# Returns nothing; subsequent code reads back IDs via ``existing_ids``. +# ---------------------------------------------------------------------- + + +def seed_dashboards(conn: Connection, target: int, dry_run: bool) -> None: + current = count_rows(conn, "dashboards") + if current >= target: + logger.info( + "dashboards: %d rows (target %d) — no insert needed", current, target + ) + return + needed = target - current + logger.info("dashboards: %d → %d (+%d)", current, target, needed) + if dry_run: + return + + dialect = conn.engine.dialect.name + sql = sa.text( + "INSERT INTO dashboards (uuid, dashboard_title, slug, published) " + "VALUES (:uuid, :title, :slug, :published)" + ) + for batch_start in range(0, needed, BATCH): + rows = [ + { + "uuid": uuid_value(dialect), + "title": f"seed_dashboard_{current + i}", + "slug": f"seed-dashboard-{current + i}-{uuid4().hex[:8]}", + "published": False, + } + for i in range(batch_start, min(batch_start + BATCH, needed)) + ] + conn.execute(sql, rows) + logger.info(" dashboards: inserted %d / %d", batch_start + len(rows), needed) + + +def seed_dbs(conn: Connection, dry_run: bool) -> int: + """Ensure at least one row exists in ``dbs`` (parent of ``tables``). + Returns the id to use as ``database_id`` when seeding ``tables``.""" + ids = existing_ids(conn, "dbs", limit=1) + if ids: + return ids[0] + if dry_run: + return -1 # placeholder + dialect = conn.engine.dialect.name + logger.info("dbs: inserting one synthetic database (no rows present)") + conn.execute( + sa.text( + "INSERT INTO dbs (uuid, database_name, sqlalchemy_uri, expose_in_sqllab) " + "VALUES (:uuid, :name, :uri, :expose)" + ), + { + "uuid": uuid_value(dialect), + "name": f"seed_db_{uuid4().hex[:8]}", + "uri": "sqlite:///seed.db", + "expose": False, + }, + ) + return existing_ids(conn, "dbs", limit=1)[0] + + +def seed_tables(conn: Connection, target: int, dry_run: bool) -> None: + current = count_rows(conn, "tables") + if current >= target: + logger.info("tables: %d rows (target %d) — no insert needed", current, target) + return + needed = target - current + logger.info("tables: %d → %d (+%d)", current, target, needed) + if dry_run: + return + + database_id = seed_dbs(conn, dry_run=False) + dialect = conn.engine.dialect.name + sql = sa.text( + "INSERT INTO tables (uuid, table_name, database_id) " + "VALUES (:uuid, :name, :db_id)" + ) + for batch_start in range(0, needed, BATCH): + rows = [ + { + "uuid": uuid_value(dialect), + "name": f"seed_table_{current + i}", + "db_id": database_id, + } + for i in range(batch_start, min(batch_start + BATCH, needed)) + ] + conn.execute(sql, rows) + logger.info(" tables: inserted %d / %d", batch_start + len(rows), needed) + + +def seed_slices(conn: Connection, target: int, dry_run: bool) -> None: + current = count_rows(conn, "slices") + if current >= target: + logger.info("slices: %d rows (target %d) — no insert needed", current, target) + return + needed = target - current + logger.info("slices: %d → %d (+%d)", current, target, needed) + if dry_run: + return + + # Slices reference tables.id; ensure at least one ``tables`` row exists + # so the FK is satisfiable (datasource_id is nullable but we set it for + # realism). The migration test doesn't care, but a real Superset that + # re-renders these slices does. + seed_tables(conn, target=1, dry_run=False) + table_id = existing_ids(conn, "tables", limit=1)[0] + dialect = conn.engine.dialect.name + sql = sa.text( + "INSERT INTO slices " + "(uuid, slice_name, datasource_id, datasource_type, viz_type) " + "VALUES (:uuid, :name, :ds_id, :ds_type, :viz)" + ) + for batch_start in range(0, needed, BATCH): + rows = [ + { + "uuid": uuid_value(dialect), + "name": f"seed_slice_{current + i}", + "ds_id": table_id, + "ds_type": "table", + "viz": "table", + } + for i in range(batch_start, min(batch_start + BATCH, needed)) + ] + conn.execute(sql, rows) + logger.info(" slices: inserted %d / %d", batch_start + len(rows), needed) + + +def seed_users(conn: Connection, target: int, dry_run: bool) -> None: + current = count_rows(conn, "ab_user") + if current >= target: + logger.info("ab_user: %d rows (target %d) — no insert needed", current, target) + return + needed = target - current + logger.info("ab_user: %d → %d (+%d)", current, target, needed) + if dry_run: + return + + sql = sa.text( + "INSERT INTO ab_user (first_name, last_name, username, email, active) " + "VALUES (:first, :last, :username, :email, :active)" + ) + for batch_start in range(0, needed, BATCH): + rows = [ + { + "first": "seed", + "last": f"user_{current + i}", + "username": f"seed_user_{current + i}_{uuid4().hex[:8]}", + "email": f"seed_user_{current + i}_{uuid4().hex[:8]}@example.invalid", + "active": True, + } + for i in range(batch_start, min(batch_start + BATCH, needed)) + ] + conn.execute(sql, rows) + logger.info(" ab_user: inserted %d / %d", batch_start + len(rows), needed) + + +def seed_roles(conn: Connection, target: int, dry_run: bool) -> None: + current = count_rows(conn, "ab_role") + if current >= target: + logger.info("ab_role: %d rows (target %d) — no insert needed", current, target) + return + needed = target - current + logger.info("ab_role: %d → %d (+%d)", current, target, needed) + if dry_run: + return + + sql = sa.text("INSERT INTO ab_role (name) VALUES (:name)") + for batch_start in range(0, needed, BATCH): + rows = [ + {"name": f"seed_role_{current + i}_{uuid4().hex[:8]}"} + for i in range(batch_start, min(batch_start + BATCH, needed)) + ] + conn.execute(sql, rows) + logger.info(" ab_role: inserted %d / %d", batch_start + len(rows), needed) + + +# ---------------------------------------------------------------------- +# Junction seeder +# ---------------------------------------------------------------------- + + +def _load_existing_pairs( + conn: Connection, junction: str, fk1_col: str, fk2_col: str +) -> set[tuple[int, int]]: + """Load existing ``(fk1, fk2)`` pairs from a junction table into a set. + + Used so the seeder can skip them when generating new pairs (junction + tables enforce uniqueness on the FK pair). Memory is ~32 bytes/tuple + on CPython, so 10M existing pairs is ~320MB — acceptable for a dev + machine. The junction / column names come from ``JUNCTIONS``, not + user input, so the f-string interpolation is safe. + """ + sql_text = f"SELECT {fk1_col}, {fk2_col} FROM {junction}" # noqa: S608 + return {(row[0], row[1]) for row in conn.execute(sa.text(sql_text))} + + +def _generate_new_pairs( + p1_ids: list[int], + p2_ids: list[int], + existing_pairs: set[tuple[int, int]], +) -> Iterator[tuple[int, int]]: + """Yield ``(fk1, fk2)`` pairs from the parent1 × parent2 cross-product + that are not already in ``existing_pairs``.""" + for fk1 in p1_ids: + for fk2 in p2_ids: + if (fk1, fk2) not in existing_pairs: + yield (fk1, fk2) + + +def seed_junction( + conn: Connection, + junction: str, + fk1_col: str, + fk2_col: str, + parent1: str, + parent2: str, + target: int, + dry_run: bool, +) -> None: + """Bulk-insert junction rows up to ``target`` rows total. + + Generates ``(fk1, fk2)`` pairs by walking the cross-product of + parent1 IDs × parent2 IDs in row-major order, skipping pairs that + already exist. Walking the cross-product deterministically keeps + the script replayable: re-running with the same target is a no-op, + and re-running with a higher target appends new pairs in a stable + order regardless of how many runs preceded. + """ + current = count_rows(conn, junction) + if current >= target: + logger.info( + "%s: %d rows (target %d) — no insert needed", junction, current, target + ) + return + needed = target - current + logger.info("%s: %d → %d (+%d)", junction, current, target, needed) + if dry_run: + return + + p1_ids = existing_ids(conn, parent1) + p2_ids = existing_ids(conn, parent2) + max_pairs = len(p1_ids) * len(p2_ids) + if max_pairs < target: + sys.exit( + f"Cannot reach {target} rows in {junction}: " + f"only {max_pairs} unique pairs available " + f"({len(p1_ids)} × {len(p2_ids)}). " + f"Increase parent targets and rerun." + ) + + existing_pairs: set[tuple[int, int]] = ( + _load_existing_pairs(conn, junction, fk1_col, fk2_col) if current > 0 else set() + ) + if existing_pairs: + logger.info( + " %s: loaded %d existing pairs into avoidance set", + junction, + len(existing_pairs), + ) + + insert_sql = sa.text( + f"INSERT INTO {junction} ({fk1_col}, {fk2_col}) " # noqa: S608 + f"VALUES (:fk1, :fk2)" + ) + + inserted = 0 + batch: list[dict[str, int]] = [] + for fk1, fk2 in _generate_new_pairs(p1_ids, p2_ids, existing_pairs): + batch.append({"fk1": fk1, "fk2": fk2}) + inserted += 1 + if len(batch) == BATCH or inserted == needed: + conn.execute(insert_sql, batch) + logger.info(" %s: inserted %d / %d", junction, inserted, needed) + batch = [] + if inserted == needed: + return + if inserted < needed: + sys.exit( + f"Ran out of unique pairs at {inserted}/{needed} for {junction} " + f"(parents have {len(p1_ids)} × {len(p2_ids)} = {max_pairs} pairs, " + f"{len(existing_pairs)} already present)" + ) + + +# ---------------------------------------------------------------------- +# Orchestration +# ---------------------------------------------------------------------- + + +def required_parent_count(target_pairs: int, other_parent: int) -> int: + """How many rows we need in this parent so that + (this_parent × other_parent) ≥ target_pairs.""" + if other_parent == 0: + # Bootstrapping: assume we'll create at least 1 + other_parent = 1 + return -(-target_pairs // other_parent) # ceil(target_pairs / other_parent) + + +def _compute_parent_requirements(targets: dict[str, int]) -> dict[str, int]: + """For each parent table, return the minimum row count needed so that + parent1 × parent2 ≥ target for every junction it participates in. + + Allocates ceil(sqrt(target)) rows per parent, balanced across the two + parents of each junction. The actual junction seeder will then walk + the cross-product to produce the target number of unique pairs. + """ + parent_req: dict[str, int] = {} + for junction, _, _, p1, p2 in JUNCTIONS: + target = targets.get(junction, 0) + if target == 0: + continue + sqrt_n = int(target**0.5) + 1 + parent_req[p1] = max(parent_req.get(p1, 0), sqrt_n) + parent_req[p2] = max(parent_req.get(p2, 0), sqrt_n) + return parent_req + + +def _seed_parents(conn: Connection, parent_req: dict[str, int], dry_run: bool) -> None: + """Seed parent tables in dependency order: + independent parents (ab_user, ab_role) first, then dashboards / slices / + tables (which transitively depend on dbs, seeded inside seed_tables).""" + if "ab_user" in parent_req: + seed_users(conn, parent_req["ab_user"], dry_run) + if "ab_role" in parent_req: + seed_roles(conn, parent_req["ab_role"], dry_run) + if "dashboards" in parent_req: + seed_dashboards(conn, parent_req["dashboards"], dry_run) + if "slices" in parent_req: + seed_slices(conn, parent_req["slices"], dry_run) + if "tables" in parent_req: + seed_tables(conn, parent_req["tables"], dry_run) + + +def _seed_all_junctions( + conn: Connection, targets: dict[str, int], dry_run: bool +) -> None: + for junction, fk1, fk2, p1, p2 in JUNCTIONS: + target = targets.get(junction, 0) + if target == 0: + continue + with time_phase(f"junction:{junction}"): + seed_junction(conn, junction, fk1, fk2, p1, p2, target, dry_run) + + +def inject_duplicates( + conn: Connection, + junction: str, + fk1_col: str, + fk2_col: str, + pct: float, + dry_run: bool, +) -> None: + """Insert duplicate ``(fk1, fk2)`` rows on a non-UNIQUE junction table. + + Used to stress-test the migration's ``_dedupe_by_min_id`` phase, which + is otherwise a no-op on cleanly-seeded data. Computes ``count = + current_rows * pct / 100`` and inserts that many rows by re-sampling + existing ``(fk1, fk2)`` pairs in row-major order. The synthetic + duplicates land on top of distinct existing pairs (one duplicate per + distinct pair, then wraps), so the migration's dedupe finds and + deletes them. + + **Pre-condition: the table must NOT have UNIQUE on (fk1, fk2)**, i.e., + the schema must be the pre-migration shape (after running + ``superset db downgrade``). On the post-migration schema the composite + PK rejects duplicates and this function will error. + """ + if pct == 0: + return + current = count_rows(conn, junction) + count = int(current * pct / 100) + if count == 0: + logger.info( + "%s: 0 duplicates to inject (current=%d, pct=%g)", + junction, + current, + pct, + ) + return + logger.info( + "%s: injecting %d duplicate rows (%g%% of %d existing)", + junction, + count, + pct, + current, + ) + if dry_run: + return + + select_sql = sa.text( + f"SELECT {fk1_col}, {fk2_col} FROM {junction} ORDER BY id LIMIT :n" # noqa: S608 + ) + sample = conn.execute(select_sql, {"n": count}).fetchall() + if not sample: + logger.warning("%s: no rows to duplicate (table is empty)", junction) + return + + insert_sql = sa.text( + f"INSERT INTO {junction} ({fk1_col}, {fk2_col}) " # noqa: S608 + f"VALUES (:fk1, :fk2)" + ) + inserted = 0 + while inserted < count: + batch: list[dict[str, int]] = [] + while len(batch) < BATCH and inserted < count: + row = sample[inserted % len(sample)] + batch.append({"fk1": row[0], "fk2": row[1]}) + inserted += 1 + conn.execute(insert_sql, batch) + logger.info(" %s: injected %d / %d duplicates", junction, inserted, count) + + +def _inject_dirty_data(conn: Connection, dirty_pct: float, dry_run: bool) -> None: + """Inject duplicate rows on every non-UNIQUE seeded junction. + + The two tables that originally carried ``UNIQUE(fk1, fk2)`` are + skipped because their composite-PK successor (and their pre-migration + UNIQUE constraint) both reject duplicate inserts. + """ + if dirty_pct == 0: + return + for junction, fk1, fk2, _, _ in JUNCTIONS: + if junction in JUNCTIONS_WITH_UNIQUE: + logger.info( + "%s: skipping duplicate injection (table has UNIQUE on FK pair)", + junction, + ) + continue + with time_phase(f"dirty:{junction}"): + inject_duplicates(conn, junction, fk1, fk2, dirty_pct, dry_run) + + +def run(targets: dict[str, int], dry_run: bool, dirty_duplicates_pct: float) -> None: + engine = build_engine() + with engine.begin() as conn: + parent_req = _compute_parent_requirements(targets) + logger.info("Required parent row counts: %s", parent_req) + + with time_phase("parents"): + _seed_parents(conn, parent_req, dry_run) + + with time_phase("junctions"): + _seed_all_junctions(conn, targets, dry_run) + + if dirty_duplicates_pct > 0: + with time_phase("dirty-duplicates"): + _inject_dirty_data(conn, dirty_duplicates_pct, dry_run) + + +# ---------------------------------------------------------------------- +# CLI +# ---------------------------------------------------------------------- + + +def main() -> None: + parser = argparse.ArgumentParser( + description=__doc__, + formatter_class=argparse.RawDescriptionHelpFormatter, + ) + for table, default in DEFAULTS.items(): + parser.add_argument( + f"--{table.replace('_', '-')}", + type=int, + default=default, + help=f"target row count for {table} (default: {default:,})", + ) + parser.add_argument( + "--dry-run", + "-n", + action="store_true", + help="print planned inserts without writing to the DB", + ) + parser.add_argument( + "--dirty-duplicates-pct", + type=float, + default=0, + help=( + "after seeding distinct pairs, inject this percentage of duplicate " + "rows on each non-UNIQUE junction (slice_user, dashboard_user, " + "dashboard_roles). Stress-tests the migration's _dedupe_by_min_id " + "phase. Requires the DB to be at the pre-migration revision " + "(33d7e0e21daa) — the post-migration composite PK rejects " + "duplicates and this will error. Default: 0 (no duplicates)." + ), + ) + parser.add_argument( + "--verbose", + "-v", + action="store_true", + help="increase log verbosity", + ) + args = parser.parse_args() + + logging.basicConfig( + level=logging.DEBUG if args.verbose else logging.INFO, + format="%(asctime)s [%(levelname)s] %(message)s", + datefmt="%H:%M:%S", + ) + + targets = {table: getattr(args, table) for table in DEFAULTS} + + logger.info("Targets: %s", targets) + logger.info("Dry run: %s", args.dry_run) + logger.info("Dirty duplicates pct: %g", args.dirty_duplicates_pct) + + with time_phase("total"): + run( + targets, + dry_run=args.dry_run, + dirty_duplicates_pct=args.dirty_duplicates_pct, + ) + + +if __name__ == "__main__": + main() diff --git a/superset-frontend/package-lock.json b/superset-frontend/package-lock.json index 80443db377db..e8b6cb05ff0a 100644 --- a/superset-frontend/package-lock.json +++ b/superset-frontend/package-lock.json @@ -49297,6 +49297,7 @@ "@ant-design/icons": "^6.2.3", "@apache-superset/core": "*", "@babel/runtime": "^7.29.7", + "@braintree/sanitize-url": "^7.1.2", "@types/json-bigint": "^1.0.4", "@visx/responsive": "^3.12.0", "ace-builds": "^1.44.0", diff --git a/superset-frontend/packages/superset-ui-core/src/utils/featureFlags.ts b/superset-frontend/packages/superset-ui-core/src/utils/featureFlags.ts index 35c0a37d562e..98a5aeff492b 100644 --- a/superset-frontend/packages/superset-ui-core/src/utils/featureFlags.ts +++ b/superset-frontend/packages/superset-ui-core/src/utils/featureFlags.ts @@ -68,6 +68,7 @@ export enum FeatureFlag { TaggingSystem = 'TAGGING_SYSTEM', Thumbnails = 'THUMBNAILS', UseAnalogousColors = 'USE_ANALOGOUS_COLORS', + VersionHistory = 'VERSION_HISTORY', ForceSqlLabRunAsync = 'SQLLAB_FORCE_RUN_ASYNC', SlackEnableAvatars = 'SLACK_ENABLE_AVATARS', EnableDashboardScreenshotEndpoints = 'ENABLE_DASHBOARD_SCREENSHOT_ENDPOINTS', diff --git a/superset-frontend/spec/helpers/reducerIndex.ts b/superset-frontend/spec/helpers/reducerIndex.ts index a4e0021839a8..034305ca5a05 100644 --- a/superset-frontend/spec/helpers/reducerIndex.ts +++ b/superset-frontend/spec/helpers/reducerIndex.ts @@ -30,6 +30,7 @@ import saveModal from 'src/explore/reducers/saveModalReducer'; import explore from 'src/explore/reducers/exploreReducer'; import sqlLab from 'src/SqlLab/reducers/sqlLab'; import reports from 'src/features/reports/ReportModal/reducer'; +import versionHistory from 'src/features/versionHistory/reducer'; import getBootstrapData from 'src/utils/getBootstrapData'; const impressionId = (state = '') => state; @@ -60,6 +61,7 @@ export default { sqlLab, localStorageUsageInKilobytes: noopReducer(0), reports, + versionHistory, common: noopReducer(common), user: noopReducer(user), }; diff --git a/superset-frontend/src/constants.ts b/superset-frontend/src/constants.ts index efd4a516e4c6..dceb75535443 100644 --- a/superset-frontend/src/constants.ts +++ b/superset-frontend/src/constants.ts @@ -107,6 +107,10 @@ export const URL_PARAMS = { name: 'edit', type: 'boolean', }, + versionHistory: { + name: 'version_history', + type: 'boolean', + }, } as const; export const RESERVED_CHART_URL_PARAMS: string[] = [ diff --git a/superset-frontend/src/dashboard/actions/hydrate.ts b/superset-frontend/src/dashboard/actions/hydrate.ts index f28803a0dec2..3997d703d2e1 100644 --- a/superset-frontend/src/dashboard/actions/hydrate.ts +++ b/superset-frontend/src/dashboard/actions/hydrate.ts @@ -73,7 +73,7 @@ import { export const HYDRATE_DASHBOARD = 'HYDRATE_DASHBOARD'; type AppDispatch = ThunkDispatch; -interface HydrateChartData { +export interface HydrateChartData { slice_id: number; slice_url: string; slice_name: string; @@ -85,7 +85,7 @@ interface HydrateChartData { changed_on: string; } -interface HydrateDashboardData extends Dashboard { +export interface HydrateDashboardData extends Dashboard { metadata: JsonObject; position_data: Record | null; [key: string]: unknown; diff --git a/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.test.tsx b/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.test.tsx index 726aab0a0ef3..72f28efbf4b4 100644 --- a/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.test.tsx +++ b/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.test.tsx @@ -567,3 +567,72 @@ test('should maintain layout when switching between tabs', async () => { expect(gridContainer).toBeInTheDocument(); expect(tabPanels.length).toBeGreaterThan(0); }); + +const dashboardPreviewState = { + isPanelOpen: true, + entityType: 'dashboard', + include: 'all', + preview: { + entityUuid: 'dash-uuid', + versionUuid: 'version-uuid', + transactionId: 7, + headline: 'Dec 5, 2025, 12:18 PM', + issuedAt: '2025-12-05T17:18:00', + }, + sessionLog: [], + restoreCount: 0, +}; + +test('gates the dashboard grid while a version preview is active', async () => { + (useStoredSidebarWidth as jest.Mock).mockImplementation(() => [ + 100, + jest.fn(), + ]); + (fetchFaveStar as jest.Mock).mockReturnValue({ type: 'mock-action' }); + (setActiveTab as jest.Mock).mockReturnValue({ type: 'mock-action' }); + + const { findByTestId } = render(, { + useRedux: true, + store: storeWithState({ + ...mockState, + dashboardLayout: undoableDashboardLayout, + versionHistory: dashboardPreviewState, + }), + useDnd: true, + useRouter: true, + useTheme: true, + }); + + const gate = await findByTestId('dashboard-grid-gate'); + expect(gate).toHaveAttribute('aria-disabled', 'true'); + expect(gate).toHaveStyleRule('pointer-events', 'none'); + // Tab navigation is deliberately carved out of the gate so tabbed + // dashboards remain navigable while previewing. + expect(gate).toHaveStyleRule('pointer-events', 'auto', { + target: '.ant-tabs-nav', + }); +}); + +test('does not gate the dashboard grid without an active preview', async () => { + (useStoredSidebarWidth as jest.Mock).mockImplementation(() => [ + 100, + jest.fn(), + ]); + (fetchFaveStar as jest.Mock).mockReturnValue({ type: 'mock-action' }); + (setActiveTab as jest.Mock).mockReturnValue({ type: 'mock-action' }); + + const { findByTestId } = render(, { + useRedux: true, + store: storeWithState({ + ...mockState, + dashboardLayout: undoableDashboardLayout, + }), + useDnd: true, + useRouter: true, + useTheme: true, + }); + + const gate = await findByTestId('dashboard-grid-gate'); + expect(gate).toHaveAttribute('aria-disabled', 'false'); + expect(gate).not.toHaveStyleRule('pointer-events', 'none'); +}); diff --git a/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.tsx b/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.tsx index 3fbfc5d10a55..dc3e353d2958 100644 --- a/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.tsx +++ b/superset-frontend/src/dashboard/components/DashboardBuilder/DashboardBuilder.tsx @@ -18,9 +18,24 @@ */ /* eslint-env browser */ import cx from 'classnames'; -import { memo, useCallback, useEffect, useMemo, useRef, useState } from 'react'; +import { + Suspense, + lazy, + memo, + useCallback, + useEffect, + useMemo, + useRef, + useState, +} from 'react'; import { t } from '@apache-superset/core/translation'; -import { addAlpha, JsonObject, useElementOnScreen } from '@superset-ui/core'; +import { + addAlpha, + isFeatureEnabled, + FeatureFlag, + JsonObject, + useElementOnScreen, +} from '@superset-ui/core'; import { css, styled, useTheme } from '@apache-superset/core/theme'; import { useDispatch, useSelector } from 'react-redux'; import { EmptyState, Loading } from '@superset-ui/core/components'; @@ -68,11 +83,21 @@ import { OPEN_FILTER_BAR_WIDTH, EMPTY_CONTAINER_Z_INDEX, } from 'src/dashboard/constants'; +import { selectIsDashboardVersionPreviewActive } from 'src/features/versionHistory/reducer'; import { getRootLevelTabsComponent, shouldFocusTabs } from './utils'; import DashboardContainer from './DashboardContainer'; import { useNativeFilters } from './state'; import DashboardWrapper from './DashboardWrapper'; +// Lazy-loaded so deployments with the VersionHistory flag off never pay +// the bundle cost of the feature's component graph. +const DashboardVersionHistory = lazy( + () => import('src/features/versionHistory/DashboardVersionHistory'), +); +const PreviewBanner = lazy( + () => import('src/features/versionHistory/PreviewBanner'), +); + // @z-index-above-dashboard-charts + 1 = 11 const FiltersPanel = styled.div<{ width: number; hidden: boolean }>` background-color: ${({ theme }) => theme.colorBgContainer}; @@ -128,6 +153,17 @@ const StyledContent = styled.div<{ ${({ fullSizeChartId }) => fullSizeChartId && `z-index: 101;`} `; +// Sticks alongside the page scroll so the panel stays fully visible. +const VersionHistoryColumn = styled.div` + grid-column: 3; + grid-row: 1 / span 2; + position: sticky; + top: 0; + align-self: start; + height: 100vh; + z-index: 99; +`; + const DashboardContentWrapper = styled.div` ${({ theme }) => css` &.dashboard { @@ -273,8 +309,9 @@ const DashboardContentWrapper = styled.div` const StyledDashboardContent = styled.div<{ editMode: boolean; marginLeft: number; + previewGated: boolean; }>` - ${({ theme, editMode, marginLeft }) => css` + ${({ theme, editMode, marginLeft, previewGated }) => css` background-color: ${theme.colorBgLayout}; display: flex; flex-direction: row; @@ -282,6 +319,20 @@ const StyledDashboardContent = styled.div<{ height: auto; flex: 1; + ${previewGated && + ` + /* Block chart interactions (context menus, cross-filters, drills) + while previewing a historical version. Tab bars deliberately stay + clickable: previewing a tabbed dashboard requires navigating its + tabs, and tab state is ephemeral — exiting the preview re-hydrates + and wipes it. Page scrolling is unaffected because the page, not + this subtree, owns the scroll. */ + pointer-events: none; + .ant-tabs-nav { + pointer-events: auto; + } + `} + .grid-container .dashboard-component-tabs { box-shadow: none; padding-left: 0; @@ -387,6 +438,9 @@ const DashboardBuilder = () => { const filterBarOrientation = useSelector( ({ dashboardInfo }) => dashboardInfo.filterBarOrientation, ); + const isVersionPreviewActive = useSelector( + selectIsDashboardVersionPreviewActive, + ); const handleChangeTab = useCallback( ({ pathToTabIndex }: { pathToTabIndex: string[] }) => { @@ -515,14 +569,35 @@ const DashboardBuilder = () => { {!hideDashboardHeader && } {showFilterBar && filterBarOrientation === FilterBarOrientation.Horizontal && ( -
__key`` on + Postgres, ``
__`` or the explicit ``uq_*`` we may have + given it on MySQL), so they're reflectable by name. SQLite is + handled separately via ``recreate="always"`` + ``copy_from`` because + it reflects unnamed UNIQUEs with ``name=None``. + + No-op if no matching UNIQUE is found (defensive — re-runs after a + partial application should not error). + """ + for uc in insp.get_unique_constraints(t.name): + if set(uc.get("column_names", [])) == {t.fk1, t.fk2} and uc.get("name"): + op.drop_constraint(uc["name"], t.name, type_="unique") + return + + +# MySQL ON DELETE actions that the downgrade re-create loop is allowed +# to interpolate into raw SQL. The reflected value comes from MySQL's +# information_schema (so not user input), but a whitelist eliminates +# the "what if an unexpected value appears" question entirely. The +# four entries are the SQL-standard set; SET DEFAULT is intentionally +# excluded because InnoDB silently downgrades it to NO ACTION. +_VALID_ONDELETE_ACTIONS: frozenset[str] = frozenset( + {"CASCADE", "SET NULL", "RESTRICT", "NO ACTION"} +) + + +def _enforce_not_null_for_sqlite( + batch_op: BatchOperations, t: AssociationTable, conn: Connection +) -> None: + """Force ``NOT NULL`` on the FK columns post-PK-promotion on SQLite only. + + SQLite has a long-standing quirk: composite ``PRIMARY KEY`` does not + promote constituent columns to ``NOT NULL`` (only ``INTEGER PRIMARY KEY`` + does). PostgreSQL and MySQL implicitly promote the PK columns to + ``NOT NULL`` when the constraint is added, making the explicit + ``alter_column`` redundant there. + + Skipping the ``alter_column`` on MySQL is also functionally required: + MySQL 8 rejects ``ALTER COLUMN`` on a column that participates in a + foreign key constraint with ``ERROR 1832 (HY000): Cannot change column + 'X': used in a foreign key constraint 'Y'`` whenever the table has + data — even when the only change is ``NULL`` → ``NOT NULL`` and the + column is already part of a freshly-added composite primary key (which + InnoDB has just made implicitly ``NOT NULL`` anyway). The error fires + on populated tables but not on empty ones, which is why CI's + ``test-mysql`` shard (fresh schema) didn't catch this and a real + production-shaped install does. + + Only SQLite still needs the explicit step, and SQLite has no FK + enforcement objection. + """ + if conn.dialect.name == "sqlite": + batch_op.alter_column(t.fk1, existing_type=sa.Integer, nullable=False) + batch_op.alter_column(t.fk2, existing_type=sa.Integer, nullable=False) + + +def upgrade() -> None: + conn = op.get_bind() + _check_no_external_fks_to_id(conn) + insp = inspect(conn) + + for t in AFFECTED_TABLES: + # Run NULL-FK cleanup unconditionally: it is a no-op DELETE on tables + # whose FK columns are already NOT NULL (cheap), and skipping it on a + # table whose FK was nullable would leave the PK-add to fail with a + # cryptic constraint violation. Cf. ``TABLES_WITH_NULLABLE_FKS`` above + # for documentation of which tables are known to have nullable FKs. + _delete_null_fk_rows(conn, t) + _dedupe_by_min_id(conn, t) + _assert_no_duplicates(conn, t) + + # Two tables (``dashboard_slices``, ``report_schedule_user``) + # carry a redundant ``UNIQUE(fk1, fk2)`` that the composite PK + # subsumes. Three dialect-specific paths: + # + # * **PostgreSQL** — the UNIQUE constraint has a stable + # reflected name (Postgres default convention), so we + # ``DROP CONSTRAINT`` by name and then run the structural + # change as direct ALTER. This avoids the full-table copy + # that ``recreate="always"`` would trigger + # (``CREATE TABLE AS SELECT → DROP → RENAME``), holding + # ``ACCESS EXCLUSIVE`` only for the (much shorter) PK + # index build instead of the full copy duration. + # + # * **MySQL** — InnoDB binds the FK constraints to the + # redundant UNIQUE's underlying index for back-reference, + # so a direct ``DROP CONSTRAINT`` of the UNIQUE raises + # ``ERROR 1553``. Use ``recreate="always"`` to rebuild the + # table without the UNIQUE; drop the FKs first to dodge + # the ``ERROR 1826`` (duplicate FK constraint name) that + # the temp-table phase would otherwise provoke. The FKs + # are re-created automatically as part of ``copy_from``. + # + # * **SQLite** — unnamed UNIQUE constraints reflect with + # ``name=None`` and can't be dropped by name. Use + # ``recreate="always"`` + ``copy_from`` (omits UNIQUE). + # SQLite always rebuilds for PK changes anyway, so the + # recreate isn't extra cost there. + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + if conn.dialect.name == "postgresql": + _drop_redundant_unique_by_name(conn, insp, t) + with op.batch_alter_table(t.name) as batch_op: + batch_op.drop_column("id") + batch_op.create_primary_key(f"pk_{t.name}", [t.fk1, t.fk2]) + _enforce_not_null_for_sqlite(batch_op, t, conn) + else: + if conn.dialect.name == "mysql": + for fk in insp.get_foreign_keys(t.name): + if fk_name := fk.get("name"): + op.drop_constraint(fk_name, t.name, type_="foreignkey") + with op.batch_alter_table( + t.name, + recreate="always", + copy_from=_build_pre_upgrade_table(insp, t), + ) as batch_op: + batch_op.drop_column("id") + batch_op.create_primary_key(f"pk_{t.name}", [t.fk1, t.fk2]) + _enforce_not_null_for_sqlite(batch_op, t, conn) + else: + with op.batch_alter_table(t.name) as batch_op: + batch_op.drop_column("id") + batch_op.create_primary_key(f"pk_{t.name}", [t.fk1, t.fk2]) + _enforce_not_null_for_sqlite(batch_op, t, conn) + + +def downgrade() -> None: + # Inverse order: undo upgrade transformations from last-applied to + # first-applied. Within each table, drop the composite PK, restore the + # surrogate ``id`` column, and re-add the original ``UNIQUE`` constraint + # on the two tables that previously carried one. + # + # Note: FK columns remain NOT NULL after downgrade (intentional asymmetry + # — see UPDATING.md). Restoring the original nullable state would require + # an explicit ``alter_column`` per FK per table for no operator value; + # junction-table NULL FKs were always meaningless under ``secondary=`` + # semantics. + # The downgrade names the restored PK ``
_pkey`` (matching Postgres' + # default constraint-naming convention, which was the original constraint + # name before this migration ran) so a downgrade-then-upgrade round-trip + # doesn't collide on the upgrade's ``pk_
`` name. + # + # Adding a NOT NULL ``id`` column to a table with existing rows requires + # a default that fires on the existing rows. ``sa.Identity()`` (Postgres + # 10+ / MySQL 8+) and ``sa.Sequence`` (with explicit nextval) both + # backfill existing rows during ALTER TABLE; bare ``autoincrement=True`` + # does not. ``Identity`` is the modern portable choice. + conn = op.get_bind() + insp = inspect(conn) + is_mysql = conn.dialect.name == "mysql" + for t in reversed(AFFECTED_TABLES): + if is_mysql: + _downgrade_mysql_table(insp, t) + else: + with op.batch_alter_table(t.name) as batch_op: + batch_op.drop_constraint(f"pk_{t.name}", type_="primary") + batch_op.add_column( + sa.Column( + "id", + sa.Integer, + sa.Identity(always=False), + nullable=False, + ) + ) + batch_op.create_primary_key(f"{t.name}_pkey", ["id"]) + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + batch_op.create_unique_constraint( + f"uq_{t.name}_{t.fk1}_{t.fk2}", [t.fk1, t.fk2] + ) + + +def _downgrade_mysql_table( + insp: sa.engine.reflection.Inspector, t: AssociationTable +) -> None: + """MySQL-specific downgrade for one table. + + Two MySQL quirks force a dialect-specific path here: + + 1. **ERROR 1553 — ``Cannot drop index 'PRIMARY': needed in a foreign + key constraint``**. InnoDB uses the composite PK index to back the + FK on the leftmost column. Dropping the PK before the FKs orphans + that backing index. PostgreSQL and SQLite create separate indexes + for FK columns and don't need this dance. We drop the FKs first + and re-add them after the structural change. + + 2. **``Identity(always=False)`` on a non-PK column add does not emit + ``AUTO_INCREMENT`` on MySQL.** SQLAlchemy 1.4 only emits + ``AUTO_INCREMENT`` when the column has both ``Identity()`` and + ``primary_key=True`` at create time. Our portable path adds the + column first, then creates the PK separately — which works on + Postgres (the column gets ``GENERATED BY DEFAULT AS IDENTITY``) + and SQLite (``INTEGER PRIMARY KEY`` becomes a rowid alias) but + leaves MySQL without auto-generation, so existing rows can't be + backfilled and future ``INSERT`` statements fail with + ``Field 'id' doesn't have a default value``. The combined + ``DROP PRIMARY KEY, ADD COLUMN AUTO_INCREMENT, ADD PRIMARY KEY`` + in a single ALTER statement is the canonical MySQL idiom: MySQL + backfills existing rows with sequential values and the column + remains auto-incrementing for future inserts. + + Raw SQL is unavoidable here — there is no SQLAlchemy core equivalent + for the combined-ALTER form, and the constitution allows raw SQL for + dialect-specific DDL with no programmatic equivalent (preferring + triple-quoted strings for legibility). + + Belt-and-braces guard: ``t.name`` is interpolated as a backtick-quoted + identifier in the ALTER statements below. The value comes from + ``AFFECTED_TABLES`` (a module-level literal), so SQL injection is + structurally precluded. The explicit ``allowed`` check here makes + that invariant load-bearing rather than implicit, so a future + refactor that loosens the call-site can't slip past review. + """ + allowed = {a.name for a in AFFECTED_TABLES} + if t.name not in allowed: + raise RuntimeError( + f"Refusing to ALTER unknown table {t.name!r}: " + f"only AFFECTED_TABLES entries may flow through this path." + ) + + fks = insp.get_foreign_keys(t.name) + + for fk in fks: + if fk_name := fk.get("name"): + op.execute(f"ALTER TABLE `{t.name}` DROP FOREIGN KEY `{fk_name}`") + + op.execute( + f""" + ALTER TABLE `{t.name}` + DROP PRIMARY KEY, + ADD COLUMN id INT NOT NULL AUTO_INCREMENT, + ADD PRIMARY KEY (id) + """ + ) + + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + op.execute( + f""" + ALTER TABLE `{t.name}` + ADD UNIQUE INDEX `uq_{t.name}_{t.fk1}_{t.fk2}` + (`{t.fk1}`, `{t.fk2}`) + """ + ) + + for fk in fks: + # Guard the FK name for symmetry with the drop loop above. + # MySQL/InnoDB always reflects a name for FK constraints + # (auto-assigning ``
_ibfk_`` if none was specified), + # so this branch is defensive rather than reachable in practice. + fk_name = fk.get("name") + if not fk_name: + continue + ondelete = fk.get("options", {}).get("ondelete") + # Defensive whitelist: ``ondelete`` is reflected from MySQL's + # information_schema (not user input), but interpolating it + # into raw SQL without a check leaves a "what if an + # unexpected value appears" footgun. The SQL standard defines + # exactly four actions; reject anything else loudly. + if ondelete and ondelete.upper() not in _VALID_ONDELETE_ACTIONS: + raise RuntimeError( + f"Unexpected ON DELETE action {ondelete!r} reflected from " + f"{t.name}.{fk_name}; refusing to interpolate into raw SQL." + ) + ondelete_clause = f" ON DELETE {ondelete}" if ondelete else "" + local_cols = ", ".join(f"`{c}`" for c in fk["constrained_columns"]) + ref_cols = ", ".join(f"`{c}`" for c in fk["referred_columns"]) + op.execute( + f""" + ALTER TABLE `{t.name}` + ADD CONSTRAINT `{fk_name}` + FOREIGN KEY ({local_cols}) + REFERENCES `{fk["referred_table"]}` ({ref_cols}) + {ondelete_clause} + """ + ) diff --git a/superset/migrations/versions/2026-05-28_19-50_56cd24c07170_add_versioning_tables.py b/superset/migrations/versions/2026-05-28_19-50_56cd24c07170_add_versioning_tables.py new file mode 100644 index 000000000000..2c036dba94b4 --- /dev/null +++ b/superset/migrations/versions/2026-05-28_19-50_56cd24c07170_add_versioning_tables.py @@ -0,0 +1,588 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""add_versioning_tables + +Creates the full schema backing sc-103156 entity versioning in a single +migration: + +1. ``version_transaction`` — audit log keyed by Continuum's per-flush + transaction id (plus a Postgres-specific id sequence). +2. **Parent shadow tables** mirroring each versioned entity's columns: + ``dashboards_version`` / ``slices_version`` / ``tables_version``. +3. ``version_changes`` — field-level diff log keyed to a + ``(transaction, entity)`` pair; each row describes one atomic change + (one field or one child-collection element) that occurred during a + save. +4. **Child shadow tables** for the collections Continuum auto-registers + when ``__versioned__`` is applied to ``TableColumn`` / ``SqlMetric`` + and the ``slices`` exclude is removed from + ``Dashboard.__versioned__``: ``table_columns_version`` / + ``sql_metrics_version`` / ``dashboard_slices_version``. + +All shadow tables follow the validity-strategy shape (mirrored columns ++ ``transaction_id`` / ``end_transaction_id`` / ``operation_type`` +bookkeeping with FKs to ``version_transaction.id``). The current +version row has ``end_transaction_id = NULL``. + +This migration replaces three iterative migrations from the spike phase +(``56cd24c07170``, ``e1f3c5a7b9d0``, ``f7a2b3c4d5e6``) that captured the +same schema in three steps as the feature was developed. Compacting +gives downstream operators one migration to apply / reverse and one +review surface. The ``revision`` hash is reused from the original first +migration so anyone still tracking the chain by that hash lands on the +same logical change set. + +Generated by hand because the current Continuum + Alembic-autogenerate +interaction trips on the renamed ``transaction`` -> ``version_transaction`` +table key (``KeyError`` lookups in ``table_key_to_table``). Column +inventories were sourced from the live model ``__table__`` definitions +and ``version_class(...).__table__`` / Continuum association metadata. + +Primary key choice. Both ``version_transaction.id`` and +``version_changes.id`` are ``BigInteger`` autoincrement — a deliberate +carveout from the project's UUID-PK convention for new models (see +``CLAUDE.md`` §"UUID Migration"). ``version_transaction`` is keyed +externally by SQLAlchemy-Continuum via +``nextval('version_transaction_id_seq')`` on every INSERT; matching +that contract is required for ``versioning_manager`` to function. +``version_changes`` follows the same shape because the user-facing +identity is the ``(transaction_id, entity_kind, entity_id, sequence)`` +composite unique key, not the row id; the API surfaces a deterministic +UUIDv5 ``version_uuid`` derived from ``entity.uuid`` and +``transaction_id`` for stable external references. + +See spec FR-016..FR-021, data-model.md §``version_changes`` / +§"Storage architecture", and the spike notes in +``spike-continuum-restore.md``. + +Revision ID: 56cd24c07170 +Revises: 2bee73611e32 +Create Date: 2026-05-28 19:50:00.000000 + +""" + +from __future__ import annotations + +import sqlalchemy as sa +from alembic import op +from sqlalchemy_utils import UUIDType + +from superset.utils.core import MediumText + +revision = "56cd24c07170" +# Stacked on sc-105349-composite-association-pks (2bee73611e32) so the +# Continuum shadow tables this migration creates can mirror the +# composite-PK shape of the live association tables. If sc-105349 +# is removed from the stack, this should be reverted to "ce6bd21901ab". +down_revision = "2bee73611e32" + + +def upgrade() -> None: + bind = op.get_bind() + + # ------------------------------------------------------------------ + # version_transaction + # + # Audit log for each versioning event. Continuum emits + # ``nextval('version_transaction_id_seq')`` on every INSERT, so the + # sequence must exist before the table on Postgres. SQLite/MySQL + # ignore the explicit CREATE SEQUENCE (they auto-increment natively). + # ------------------------------------------------------------------ + if bind.dialect.name == "postgresql": + op.execute("CREATE SEQUENCE IF NOT EXISTS version_transaction_id_seq") + + op.create_table( + "version_transaction", + sa.Column( + "id", + sa.BigInteger(), + sa.Sequence("version_transaction_id_seq"), + primary_key=True, + autoincrement=True, + nullable=False, + ), + sa.Column("issued_at", sa.DateTime(), nullable=True), + sa.Column("remote_addr", sa.String(50), nullable=True), + sa.Column("user_id", sa.Integer(), nullable=True), + # ``action_kind`` carries the high-level avenue that produced + # this transaction (``restore`` / ``import`` / ``clone``). + # ``NULL`` is the default "ordinary save" — most rows leave + # this empty. Commands set + # ``session.info["_versioning_action_kind"]`` before commit; + # the change-record listener stamps the value here. Parallel + # to ``version_changes.entity_kind`` and ``version_changes.kind`` + # — the schema's third ``*_kind`` column, at transaction scope. + sa.Column("action_kind", sa.String(32), nullable=True), + ) + + if bind.dialect.name == "postgresql": + op.execute( + "ALTER SEQUENCE version_transaction_id_seq OWNED BY version_transaction.id" + ) + + # ------------------------------------------------------------------ + # dashboards_version + # ------------------------------------------------------------------ + op.create_table( + "dashboards_version", + sa.Column("uuid", UUIDType(binary=True), nullable=True), + sa.Column("created_on", sa.DateTime(), nullable=True), + sa.Column("changed_on", sa.DateTime(), nullable=True), + sa.Column("id", sa.Integer(), nullable=False), + sa.Column("dashboard_title", sa.String(500), nullable=True), + # ``MediumText()`` mirrors the live column type — on MySQL plain + # ``TEXT`` caps at 64 KB, which large dashboards exceed; an + # oversized live write would then fail the shadow INSERT under + # ``STRICT_TRANS_TABLES`` (or silently truncate without it) and + # corrupt the history. Postgres ``TEXT`` is unbounded and SQLite + # ignores the length annotation so this is MySQL-driven. + sa.Column("position_json", MediumText(), nullable=True), + sa.Column("description", sa.Text(), nullable=True), + sa.Column("css", MediumText(), nullable=True), + sa.Column("theme_id", sa.Integer(), nullable=True), + sa.Column("certified_by", sa.Text(), nullable=True), + sa.Column("certification_details", sa.Text(), nullable=True), + sa.Column("json_metadata", MediumText(), nullable=True), + sa.Column("slug", sa.String(255), nullable=True), + sa.Column("published", sa.Boolean(), nullable=True), + sa.Column("is_managed_externally", sa.Boolean(), nullable=True), + sa.Column("external_url", sa.Text(), nullable=True), + sa.Column("created_by_fk", sa.Integer(), nullable=True), + sa.Column("changed_by_fk", sa.Integer(), nullable=True), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint("id", "transaction_id"), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_dashboards_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_dashboards_version_end_transaction_id", + ), + ) + op.create_index( + "ix_dashboards_version_end_transaction_id", + "dashboards_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_dashboards_version_operation_type", + "dashboards_version", + ["operation_type"], + ) + op.create_index( + "ix_dashboards_version_transaction_id", + "dashboards_version", + ["transaction_id"], + ) + + # ------------------------------------------------------------------ + # slices_version (Charts) + # ------------------------------------------------------------------ + op.create_table( + "slices_version", + sa.Column("uuid", UUIDType(binary=True), nullable=True), + sa.Column("created_on", sa.DateTime(), nullable=True), + sa.Column("changed_on", sa.DateTime(), nullable=True), + sa.Column("id", sa.Integer(), nullable=False), + sa.Column("slice_name", sa.String(250), nullable=True), + sa.Column("datasource_id", sa.Integer(), nullable=True), + sa.Column("datasource_type", sa.String(200), nullable=True), + sa.Column("datasource_name", sa.String(2000), nullable=True), + sa.Column("viz_type", sa.String(250), nullable=True), + sa.Column("params", MediumText(), nullable=True), + sa.Column("description", sa.Text(), nullable=True), + sa.Column("cache_timeout", sa.Integer(), nullable=True), + sa.Column("perm", sa.String(1000), nullable=True), + sa.Column("schema_perm", sa.String(1000), nullable=True), + sa.Column("catalog_perm", sa.String(1000), nullable=True), + sa.Column("last_saved_at", sa.DateTime(), nullable=True), + sa.Column("last_saved_by_fk", sa.Integer(), nullable=True), + sa.Column("certified_by", sa.Text(), nullable=True), + sa.Column("certification_details", sa.Text(), nullable=True), + sa.Column("is_managed_externally", sa.Boolean(), nullable=True), + sa.Column("external_url", sa.Text(), nullable=True), + sa.Column("created_by_fk", sa.Integer(), nullable=True), + sa.Column("changed_by_fk", sa.Integer(), nullable=True), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint("id", "transaction_id"), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_slices_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_slices_version_end_transaction_id", + ), + ) + op.create_index( + "ix_slices_version_end_transaction_id", + "slices_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_slices_version_operation_type", + "slices_version", + ["operation_type"], + ) + op.create_index( + "ix_slices_version_transaction_id", + "slices_version", + ["transaction_id"], + ) + + # ------------------------------------------------------------------ + # tables_version (SqlaTable / Datasets) + # ------------------------------------------------------------------ + op.create_table( + "tables_version", + sa.Column("uuid", UUIDType(binary=True), nullable=True), + sa.Column("created_on", sa.DateTime(), nullable=True), + sa.Column("changed_on", sa.DateTime(), nullable=True), + sa.Column("id", sa.Integer(), nullable=False), + sa.Column("description", sa.Text(), nullable=True), + sa.Column("default_endpoint", sa.Text(), nullable=True), + sa.Column("is_featured", sa.Boolean(), nullable=True), + sa.Column("filter_select_enabled", sa.Boolean(), nullable=True), + sa.Column("offset", sa.Integer(), nullable=True), + sa.Column("cache_timeout", sa.Integer(), nullable=True), + sa.Column("params", sa.String(1000), nullable=True), + sa.Column("perm", sa.String(1000), nullable=True), + sa.Column("schema_perm", sa.String(1000), nullable=True), + sa.Column("catalog_perm", sa.String(1000), nullable=True), + sa.Column("is_managed_externally", sa.Boolean(), nullable=True), + sa.Column("external_url", sa.Text(), nullable=True), + sa.Column("table_name", sa.String(250), nullable=True), + sa.Column("main_dttm_col", sa.String(250), nullable=True), + sa.Column("currency_code_column", sa.String(250), nullable=True), + sa.Column("database_id", sa.Integer(), nullable=True), + sa.Column("fetch_values_predicate", sa.Text(), nullable=True), + sa.Column("schema", sa.String(255), nullable=True), + sa.Column("catalog", sa.String(256), nullable=True), + sa.Column("sql", MediumText(), nullable=True), + sa.Column("is_sqllab_view", sa.Boolean(), nullable=True), + sa.Column("template_params", sa.Text(), nullable=True), + sa.Column("extra", sa.Text(), nullable=True), + sa.Column("normalize_columns", sa.Boolean(), nullable=True), + sa.Column("always_filter_main_dttm", sa.Boolean(), nullable=True), + sa.Column("folders", sa.JSON(), nullable=True), + sa.Column("created_by_fk", sa.Integer(), nullable=True), + sa.Column("changed_by_fk", sa.Integer(), nullable=True), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint("id", "transaction_id"), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_tables_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_tables_version_end_transaction_id", + ), + ) + op.create_index( + "ix_tables_version_end_transaction_id", + "tables_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_tables_version_operation_type", + "tables_version", + ["operation_type"], + ) + op.create_index( + "ix_tables_version_transaction_id", + "tables_version", + ["transaction_id"], + ) + + # ------------------------------------------------------------------ + # version_changes + # + # Field-level diff log keyed to a (transaction, entity) pair. Each + # row describes one atomic change (one field or one child-collection + # element) that occurred to one entity during a save. See spec + # FR-016..FR-021 and data-model.md §version_changes. + # + # ``(entity_kind, entity_id)`` is a polymorphic reference: depending + # on ``entity_kind`` (``"chart"`` / ``"dashboard"`` / ``"dataset"``) + # the ``entity_id`` is the integer PK on ``slices`` / ``dashboards`` / + # ``tables`` respectively. SQL has no native polymorphic FK, so the + # constraint is intentionally omitted — cleanup relies on the + # ``CASCADE`` from ``version_transaction.id`` plus command-layer + # ordering for entity deletes (the command that hard-deletes the + # entity runs inside the same transaction that prunes its history). + # A bare ``DELETE FROM WHERE id = X`` outside that + # transactional boundary leaves orphan ``version_changes`` rows + # whose ``entity_id`` references a vanished row — the read-side + # tombstone-state lookup handles this gracefully. + # ------------------------------------------------------------------ + op.create_table( + "version_changes", + sa.Column( + "id", + sa.BigInteger(), + primary_key=True, + autoincrement=True, + nullable=False, + ), + sa.Column( + "transaction_id", + sa.BigInteger(), + sa.ForeignKey("version_transaction.id", ondelete="CASCADE"), + nullable=False, + ), + sa.Column( + "entity_kind", + sa.String(length=32), + nullable=False, + ), + sa.Column( + "entity_id", + sa.Integer(), + nullable=False, + ), + sa.Column( + "sequence", + sa.SmallInteger(), + nullable=False, + ), + sa.Column( + "kind", + sa.String(length=32), + nullable=False, + ), + # ``operation`` is the per-record verb: ``add`` / ``remove`` / + # ``move`` / ``edit``. ``move`` only fires for layout records; + # the other three apply across every emit site. Made explicit + # so consumers don't have to infer the verb from ``from_value`` + # / ``to_value`` null-tests or from ``path[0]`` for layout records. + sa.Column( + "operation", + sa.String(length=16), + nullable=False, + ), + sa.Column("path", sa.JSON(), nullable=False), + sa.Column("from_value", sa.JSON(), nullable=True), + sa.Column("to_value", sa.JSON(), nullable=True), + sa.UniqueConstraint( + "transaction_id", + "entity_kind", + "entity_id", + "sequence", + name="uq_version_changes_tx_entity_sequence", + ), + ) + op.create_index( + "ix_version_changes_kind", + "version_changes", + ["kind"], + ) + op.create_index( + "ix_version_changes_transaction_id", + "version_changes", + ["transaction_id"], + ) + op.create_index( + "ix_version_changes_entity", + "version_changes", + ["entity_kind", "entity_id"], + ) + + # ------------------------------------------------------------------ + # table_columns_version + # ------------------------------------------------------------------ + op.create_table( + "table_columns_version", + sa.Column("uuid", UUIDType(binary=True), nullable=True), + sa.Column("id", sa.Integer(), nullable=False), + sa.Column("column_name", sa.String(255), nullable=True), + sa.Column("verbose_name", sa.String(1024), nullable=True), + sa.Column("is_active", sa.Boolean(), nullable=True), + sa.Column("type", sa.Text(), nullable=True), + sa.Column("advanced_data_type", sa.String(255), nullable=True), + sa.Column("groupby", sa.Boolean(), nullable=True), + sa.Column("filterable", sa.Boolean(), nullable=True), + sa.Column("description", MediumText(), nullable=True), + sa.Column("table_id", sa.Integer(), nullable=True), + sa.Column("is_dttm", sa.Boolean(), nullable=True), + sa.Column("expression", MediumText(), nullable=True), + sa.Column("python_date_format", sa.String(255), nullable=True), + sa.Column("datetime_format", sa.String(100), nullable=True), + sa.Column("extra", sa.Text(), nullable=True), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint("id", "transaction_id"), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_table_columns_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_table_columns_version_end_transaction_id", + ), + ) + op.create_index( + "ix_table_columns_version_end_transaction_id", + "table_columns_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_table_columns_version_operation_type", + "table_columns_version", + ["operation_type"], + ) + op.create_index( + "ix_table_columns_version_transaction_id", + "table_columns_version", + ["transaction_id"], + ) + + # ------------------------------------------------------------------ + # sql_metrics_version + # ------------------------------------------------------------------ + op.create_table( + "sql_metrics_version", + sa.Column("uuid", UUIDType(binary=True), nullable=True), + sa.Column("id", sa.Integer(), nullable=False), + sa.Column("metric_name", sa.String(255), nullable=True), + sa.Column("verbose_name", sa.String(1024), nullable=True), + sa.Column("metric_type", sa.String(32), nullable=True), + sa.Column("description", MediumText(), nullable=True), + sa.Column("d3format", sa.String(128), nullable=True), + sa.Column("currency", sa.JSON(), nullable=True), + sa.Column("warning_text", sa.Text(), nullable=True), + sa.Column("table_id", sa.Integer(), nullable=True), + sa.Column("expression", MediumText(), nullable=True), + sa.Column("extra", sa.Text(), nullable=True), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint("id", "transaction_id"), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_sql_metrics_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_sql_metrics_version_end_transaction_id", + ), + ) + op.create_index( + "ix_sql_metrics_version_end_transaction_id", + "sql_metrics_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_sql_metrics_version_operation_type", + "sql_metrics_version", + ["operation_type"], + ) + op.create_index( + "ix_sql_metrics_version_transaction_id", + "sql_metrics_version", + ["transaction_id"], + ) + + # ------------------------------------------------------------------ + # dashboard_slices_version (M2M association) + # + # The live ``dashboard_slices`` table is reshaped by sc-105349 to a + # composite PK on ``(dashboard_id, slice_id)`` — no surrogate ``id``. + # Continuum auto-mirrors the live columns into the shadow Table at + # ``make_versioned()`` time, so the shadow's SQLAlchemy metadata + # also has no ``id``. The DB shadow PK is the natural composite key + # plus Continuum's bookkeeping (``transaction_id``, ``operation_type``); + # ``operation_type`` is included because a single transaction can in + # principle produce both INSERT and DELETE shadows for the same + # ``(dashboard_id, slice_id)`` pair (slice removed and re-added in + # one save). + # + # If sc-105349 is removed from the stack, the live table reverts to + # carrying its surrogate ``id`` and this migration would need to + # match — see ``spike-continuum-restore.md`` "Branch maintenance". + # ------------------------------------------------------------------ + op.create_table( + "dashboard_slices_version", + sa.Column("dashboard_id", sa.Integer(), nullable=False), + sa.Column("slice_id", sa.Integer(), nullable=False), + sa.Column("transaction_id", sa.BigInteger(), nullable=False), + sa.Column("end_transaction_id", sa.BigInteger(), nullable=True), + sa.Column("operation_type", sa.SmallInteger(), nullable=False), + sa.PrimaryKeyConstraint( + "dashboard_id", "slice_id", "transaction_id", "operation_type" + ), + sa.ForeignKeyConstraint( + ["transaction_id"], + ["version_transaction.id"], + name="fk_dashboard_slices_version_transaction_id", + ), + sa.ForeignKeyConstraint( + ["end_transaction_id"], + ["version_transaction.id"], + name="fk_dashboard_slices_version_end_transaction_id", + ), + ) + op.create_index( + "ix_dashboard_slices_version_end_transaction_id", + "dashboard_slices_version", + ["end_transaction_id"], + ) + op.create_index( + "ix_dashboard_slices_version_operation_type", + "dashboard_slices_version", + ["operation_type"], + ) + op.create_index( + "ix_dashboard_slices_version_transaction_id", + "dashboard_slices_version", + ["transaction_id"], + ) + + +def downgrade() -> None: + # Drop in reverse dependency order: children with FKs to + # ``version_transaction`` drop first; ``version_transaction`` and its + # sequence drop last. + op.drop_table("dashboard_slices_version") + op.drop_table("sql_metrics_version") + op.drop_table("table_columns_version") + op.drop_table("version_changes") + op.drop_table("tables_version") + op.drop_table("slices_version") + op.drop_table("dashboards_version") + op.drop_table("version_transaction") + + bind = op.get_bind() + if bind.dialect.name == "postgresql": + op.execute("DROP SEQUENCE IF EXISTS version_transaction_id_seq") diff --git a/superset/migrations/versions/2026-06-03_12-00_8f3a1b2c4d5e_shadow_live_row_indexes.py b/superset/migrations/versions/2026-06-03_12-00_8f3a1b2c4d5e_shadow_live_row_indexes.py new file mode 100644 index 000000000000..72c047d2ffb0 --- /dev/null +++ b/superset/migrations/versions/2026-06-03_12-00_8f3a1b2c4d5e_shadow_live_row_indexes.py @@ -0,0 +1,166 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""shadow_live_row_indexes + +Adds per-shadow-table indexes covering the canonical "current live row +of entity X" lookup that ``find_active_by_uuid`` / ``list_versions`` / +``get_version`` / restore validation / activity-view all funnel +through: + + SELECT ... FROM _version + WHERE id = ? AND end_transaction_id IS NULL + +The base migration (``56cd24c07170_add_versioning_tables``) created +single-column indexes on ``transaction_id``, ``end_transaction_id``, +and ``operation_type``, but nothing covering the predicate combination +that actually runs in hot paths. + +Index choice is dialect-specific: + +* **PostgreSQL / SQLite** — partial index over the entity ``id`` with + ``WHERE end_transaction_id IS NULL``. Cuts the index size to one row + per live entity (vs. one row per historical version) and turns the + hot lookup into a single index probe. +* **MySQL** — partial indexes aren't supported; use a plain composite + ``(id, end_transaction_id)``. MySQL's optimizer handles the + ``IS NULL`` predicate against the composite efficiently. + +It also adds a composite ``(table_id, transaction_id)`` index on the two +child shadow tables (``table_columns_version`` / ``sql_metrics_version``). +The dataset child-diff path queries these by parent ``table_id`` plus a +transaction-range bound, neither of which the base migration's +single-column indexes nor the ``id``-leading PK can serve: + + SELECT ... FROM table_columns_version + WHERE table_id = ? AND transaction_id <= ? AND ... (shadow_rows_valid_at) + + SELECT max(transaction_id) FROM table_columns_version + WHERE table_id = ? AND transaction_id < ? (prior-tx probe) + +A plain composite leading with ``table_id`` serves both on every dialect, +so no partial-index split is needed here. + +Surfaced by sqlalchemy-review pass W-NEW-4 (live-row lookup) and a +Codex sqlalchemy-review pass (child-diff ``table_id`` lookup). + +Revision ID: 8f3a1b2c4d5e +Revises: 56cd24c07170 +Create Date: 2026-06-03 12:00:00.000000 + +""" + +from __future__ import annotations + +import sqlalchemy as sa +from alembic import op + +revision = "8f3a1b2c4d5e" +down_revision = "56cd24c07170" + + +# The parent + child shadow tables, all of which carry an ``id`` +# column (mirroring the live entity's integer PK). ``dashboard_slices_version`` +# is intentionally excluded: it's the M2M association shadow with a +# composite PK ``(dashboard_id, slice_id, transaction_id, operation_type)`` +# and no ``id`` column. The canonical "live row" lookup doesn't apply to +# the M2M shadow — readers query it by ``transaction_id`` (already +# indexed by the base migration) when reconstructing per-tx changes. +SHADOW_TABLES: tuple[str, ...] = ( + "dashboards_version", + "slices_version", + "tables_version", + "table_columns_version", + "sql_metrics_version", +) + + +# Child shadow tables whose rows are looked up by parent ``table_id`` plus a +# transaction-range bound on the dataset child-diff path. Both carry a +# nullable ``table_id`` mirroring the live row's FK to ``tables.id``. +CHILD_SHADOW_TABLES: tuple[str, ...] = ( + "table_columns_version", + "sql_metrics_version", +) + + +def _index_name(table: str) -> str: + return f"ix_{table}_live_id" + + +def _child_index_name(table: str) -> str: + return f"ix_{table}_table_id_transaction_id" + + +def upgrade() -> None: + bind = op.get_bind() + dialect = bind.dialect.name + + where_clause = sa.text("end_transaction_id IS NULL") + + for table in SHADOW_TABLES: + index_name = _index_name(table) + if dialect == "postgresql": + op.create_index( + index_name, + table, + ["id"], + unique=False, + postgresql_where=where_clause, + ) + elif dialect == "sqlite": + op.create_index( + index_name, + table, + ["id"], + unique=False, + sqlite_where=where_clause, + ) + else: + # MySQL (and any unknown dialect): partial indexes aren't + # supported, so use a plain composite. MySQL's optimizer + # handles ``id = ? AND end_transaction_id IS NULL`` against + # the composite efficiently. + op.create_index( + index_name, + table, + ["id", "end_transaction_id"], + unique=False, + ) + + # Child-diff access pattern: filter by parent ``table_id`` plus a + # transaction-range bound. A plain composite serves this on every + # dialect, so no partial-index split is needed. + for table in CHILD_SHADOW_TABLES: + op.create_index( + _child_index_name(table), + table, + ["table_id", "transaction_id"], + unique=False, + ) + + +def downgrade() -> None: + # ``if_exists=True`` makes the downgrade robust against a + # partial-application failure on upgrade (e.g. the first ``op.create_index`` + # succeeded under Postgres' transactional DDL but a later one failed + # and rolled back the rest — repeated downgrade should not raise on + # the missing indexes). Postgres + SQLite + MySQL all accept the + # IF EXISTS clause. + for table in SHADOW_TABLES: + op.drop_index(_index_name(table), table_name=table, if_exists=True) + for table in CHILD_SHADOW_TABLES: + op.drop_index(_child_index_name(table), table_name=table, if_exists=True) diff --git a/superset/models/dashboard.py b/superset/models/dashboard.py index 4653272fcbf3..b26eeb5cfd94 100644 --- a/superset/models/dashboard.py +++ b/superset/models/dashboard.py @@ -35,7 +35,6 @@ String, Table, Text, - UniqueConstraint, ) from sqlalchemy.engine.base import Connection from sqlalchemy.orm import relationship, subqueryload @@ -93,37 +92,53 @@ def copy_dashboard(_mapper: Mapper, _connection: Connection, target: Dashboard) dashboard_slices = Table( "dashboard_slices", metadata, - Column("id", Integer, primary_key=True), - Column("dashboard_id", Integer, ForeignKey("dashboards.id", ondelete="CASCADE")), - Column("slice_id", Integer, ForeignKey("slices.id", ondelete="CASCADE")), - UniqueConstraint("dashboard_id", "slice_id"), + Column( + "dashboard_id", + Integer, + ForeignKey("dashboards.id", ondelete="CASCADE"), + primary_key=True, + ), + Column( + "slice_id", + Integer, + ForeignKey("slices.id", ondelete="CASCADE"), + primary_key=True, + ), ) dashboard_user = Table( "dashboard_user", metadata, - Column("id", Integer, primary_key=True), - Column("user_id", Integer, ForeignKey("ab_user.id", ondelete="CASCADE")), - Column("dashboard_id", Integer, ForeignKey("dashboards.id", ondelete="CASCADE")), + Column( + "user_id", + Integer, + ForeignKey("ab_user.id", ondelete="CASCADE"), + primary_key=True, + ), + Column( + "dashboard_id", + Integer, + ForeignKey("dashboards.id", ondelete="CASCADE"), + primary_key=True, + ), ) DashboardRoles = Table( "dashboard_roles", metadata, - Column("id", Integer, primary_key=True), Column( "dashboard_id", Integer, ForeignKey("dashboards.id", ondelete="CASCADE"), - nullable=False, + primary_key=True, ), Column( "role_id", Integer, ForeignKey("ab_role.id", ondelete="CASCADE"), - nullable=False, + primary_key=True, ), ) @@ -132,6 +147,27 @@ class Dashboard(CoreDashboard, AuditMixinNullable, ImportExportMixin): """The dashboard object!""" __tablename__ = "dashboards" + # deleted_at exclusion will be added when sc-103157 (soft delete) is merged (T043). + # SPIKE (sc-103156-versioning-full-continuum-spike): ``slices`` removed from + # the exclude list so Continuum auto-creates an association version table + # for ``dashboard_slices`` and ``Reverter(relations=["slices"])`` can + # restore chart membership. Owners / roles stay excluded — access metadata, + # not user-authored content (ADR-005). + # Audit columns (changed_on/created_on/changed_by_fk/created_by_fk) are + # auto-bumped by AuditMixin on every save; excluding them lets Continuum's + # is_modified() return False on no-op saves (e.g. owners-only edits) so we + # don't create empty version rows. version_transaction.user_id / + # issued_at preserve "who/when" without per-row duplication. + __versioned__: dict[str, Any] = { + "exclude": [ + "owners", + "roles", + "changed_on", + "created_on", + "changed_by_fk", + "created_by_fk", + ] + } id = Column(Integer, primary_key=True) dashboard_title = Column(String(500)) position_json = Column(utils.MediumText()) diff --git a/superset/models/helpers.py b/superset/models/helpers.py index c82e27135522..024194549f58 100644 --- a/superset/models/helpers.py +++ b/superset/models/helpers.py @@ -264,6 +264,53 @@ class UUIDMixin: # pylint: disable=too-few-public-methods UUIDType(binary=True), primary_key=False, unique=True, default=uuid.uuid4 ) + @validates("uuid") + def _coerce_uuid(self, key: str, value: Any) -> Any: # noqa: ARG002 + """Coerce well-formed UUID strings to ``uuid.UUID`` on assignment; + pass everything else through untouched. + + **Why coerce.** ``UUIDType`` only converts at SQL bind / SQL + result time. Importers and ad-hoc construction + (``SqlMetric(uuid="…string…")``) leave the in-memory attribute + as a ``str`` until the next DB round-trip refreshes it. With + SQLAlchemy-Continuum versioning attached to a child mapper + (``TableColumn`` / ``SqlMetric``), the post-INSERT attribute- + expire behaviour changes enough that the refresh doesn't happen + before the caller reads the attribute — breaking equality + assertions like ``test_import_dataset``'s + ``metric.uuid == uuid.UUID(...)`` because str ≠ UUID. Coercing + defensively here makes the in-memory attribute always a UUID + regardless of provenance. + + **Why the non-UUID-string escape hatch.** Tightening this + validator to raise on non-UUID strings would break a small set + of existing unit tests that use human-readable placeholder + strings as fixture uuids (e.g. + ``test_dashboard_schemas.py``'s ``"dashboard-uuid-7"`` and + analogous placeholders in importer tests). The fixtures use + these placeholders for legibility — they're only ever compared + by string equality, never written to a real database. Letting + them through unchanged keeps the fixtures working at the cost + of deferring "real" UUID malformation to the SQL bind layer, + which raises a clearer "invalid input syntax for type uuid" + error keyed to the actual column. + + **Tightening path** (if amin M1 is ever revisited): replace + the ``return value`` in the ``except`` branch with + ``raise ValueError(f"Invalid UUID: {value!r}")``, then run the + unit test suite and migrate any remaining placeholder fixtures + to ``uuid.uuid4()`` (use + ``rg '''SqlMetric\\(uuid="[^"]*"|"dashboard-uuid|"slice-uuid'''`` + to find them). The full migration touches ~5–10 fixture files + and is non-breaking outside tests. + """ + if isinstance(value, str): + try: + return uuid.UUID(value) + except ValueError: + return value + return value + @property def short_uuid(self) -> str: return str(self.uuid)[:8] @@ -546,14 +593,23 @@ def remove_params(self, param_to_remove: str) -> None: def reset_ownership(self) -> None: """object will belong to the user the current user""" - # make sure the object doesn't have relations to a user - # it will be filled by appbuilder on save - self.created_by = None - self.changed_by = None - # flask global context might not exist (in cli or tests for example) + # Reset the audit pointers. When a Flask request context is + # available we explicitly stamp the current user, otherwise we + # leave the attributes unset so Flask-AppBuilder's column + # defaults fill them in on save. An explicit assignment is + # required because once the ``created_by`` / ``changed_by`` + # relationships are configured (which happens eagerly on models + # registered with SQLAlchemy-Continuum), setting them to + # ``None`` propagates to the FK column and suppresses the + # ``default=`` callable. self.owners = [] - if g and hasattr(g, "user"): + if g and hasattr(g, "user") and g.user: + self.created_by = g.user + self.changed_by = g.user self.owners = [g.user] + else: + self.created_by = None + self.changed_by = None @property def params_dict(self) -> dict[Any, Any]: diff --git a/superset/models/slice.py b/superset/models/slice.py index e10b373d945c..f4678126550a 100644 --- a/superset/models/slice.py +++ b/superset/models/slice.py @@ -58,9 +58,18 @@ slice_user = Table( "slice_user", metadata, - Column("id", Integer, primary_key=True), - Column("user_id", Integer, ForeignKey("ab_user.id", ondelete="CASCADE")), - Column("slice_id", Integer, ForeignKey("slices.id", ondelete="CASCADE")), + Column( + "user_id", + Integer, + ForeignKey("ab_user.id", ondelete="CASCADE"), + primary_key=True, + ), + Column( + "slice_id", + Integer, + ForeignKey("slices.id", ondelete="CASCADE"), + primary_key=True, + ), ) logger = logging.getLogger(__name__) @@ -73,6 +82,28 @@ class Slice( # pylint: disable=too-many-public-methods query_context_factory: QueryContextFactory | None = None __tablename__ = "slices" + # query_context is excluded: it is a cached/regenerated field, not user-authored. + # deleted_at exclusion will be added when sc-103157 (soft delete) is merged (T043). + # Exclude M2M association relationships: Continuum only captures FK columns on + # association INSERTs (not the auto-increment id), which breaks the NOT NULL PK. + # Ownership changes are administrative metadata, not user-authored content. + # Audit / save-marker columns are auto-bumped on every save. Excluding + # them lets Continuum's is_modified() return False on no-op saves + # (e.g. owners-only edits) so we don't create empty version rows. + # version_transaction.user_id / issued_at preserve "who/when". + __versioned__: dict[str, Any] = { + "exclude": [ + "query_context", + "owners", + "dashboards", + "changed_on", + "created_on", + "changed_by_fk", + "created_by_fk", + "last_saved_at", + "last_saved_by_fk", + ] + } id = Column(Integer, primary_key=True) slice_name = Column(String(250)) datasource_id = Column(Integer) @@ -322,7 +353,11 @@ def chart(self) -> str: @property def slice_link(self) -> Markup: name = escape(self.chart) - return Markup(f'{name}') + # ``self.url`` is ``/explore/?slice_id=``; the only + # interpolation is the integer primary key, so the URL has no + # user-controlled segment to escape (unlike ``Dashboard.url`` + # which embeds the user-set slug). ``noqa: S704`` is safe. + return Markup(f'{name}') # noqa: S704 @property def icons(self) -> str: diff --git a/superset/reports/models.py b/superset/reports/models.py index f0abda8a9216..7564336ae11d 100644 --- a/superset/reports/models.py +++ b/superset/reports/models.py @@ -101,20 +101,18 @@ class ReportSourceFormat(StrEnum): report_schedule_user = Table( "report_schedule_user", metadata, - Column("id", Integer, primary_key=True), Column( "user_id", Integer, ForeignKey("ab_user.id", ondelete="CASCADE"), - nullable=False, + primary_key=True, ), Column( "report_schedule_id", Integer, ForeignKey("report_schedule.id", ondelete="CASCADE"), - nullable=False, + primary_key=True, ), - UniqueConstraint("user_id", "report_schedule_id"), ) diff --git a/superset/tasks/celery_app.py b/superset/tasks/celery_app.py index 2049246f0428..6267aa3f3186 100644 --- a/superset/tasks/celery_app.py +++ b/superset/tasks/celery_app.py @@ -34,7 +34,7 @@ # Need to import late, as the celery_app will have been setup by "create_app()" # ruff: noqa: E402, F401 # pylint: disable=wrong-import-position, unused-import -from . import cache, scheduler +from . import cache, scheduler, version_history_retention # Export the celery app globally for Celery (as run on the cmd line) to find app = celery_app diff --git a/superset/tasks/version_history_retention.py b/superset/tasks/version_history_retention.py new file mode 100644 index 000000000000..a62bce79c3ae --- /dev/null +++ b/superset/tasks/version_history_retention.py @@ -0,0 +1,413 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Celery task: prune old entity-version history. + +Retention is time-based. The task deletes parent + child shadow rows +owned by ``version_transaction`` rows whose ``issued_at`` is older +than ``SUPERSET_VERSION_HISTORY_RETENTION_DAYS`` (default 30, env +overridable, ``0`` to disable). + +One preservation rule, applied per parent shadow: + +* **Live** (``end_transaction_id IS NULL``) — never pruned. + +Baseline rows (``operation_type = 0``) and any closed historical row +are subject to the same retention window as everything else. An +entity that hasn't been edited within the window has only its live +row remaining; the historical chain (including the synthetic +baseline) ages out. + +If a transaction's parent shadow includes the live row, the whole +transaction is preserved (along with its child shadows and +``version_changes`` rows). Otherwise, all of the transaction's shadow +rows are deleted and the ``version_transaction`` row itself is +dropped — its ``version_changes`` rows cascade via the FK. + +Registered via ``CELERYBEAT_SCHEDULE`` in ``superset/config.py``. +Idempotent: a second run prunes nothing. +""" + +from __future__ import annotations + +import logging +import time +from collections.abc import Iterator +from dataclasses import dataclass +from datetime import datetime, timedelta +from typing import Any + +import sqlalchemy as sa +from flask import current_app +from sqlalchemy.exc import OperationalError + +from superset.extensions import celery_app, db, stats_logger_manager + +logger = logging.getLogger(__name__) + + +@dataclass(frozen=True) +class ShadowTables: + """The four Continuum-managed Table objects the prune walks. + + Bundled here so the prune helper's signature stays at two arguments + instead of five. The shape is set once at task entry by + ``_resolve_shadow_tables`` and threaded through the retry loop. + """ + + parent: list[sa.Table] + child: list[sa.Table] + m2m: sa.Table | None + transaction: sa.Table + + +def _resolve_shadow_tables(tx_table: sa.Table) -> ShadowTables: + """Resolve the parent / child / m2m shadow Tables from Continuum's + mapper registry and bundle them with the transaction Table. + + ``dashboard_slices_version`` is M2M-tracked by Continuum and lives + in metadata under that name (Continuum auto-creates the Table; it + isn't registered as a versioned class). Carried separately on the + ``ShadowTables`` dataclass because it doesn't follow the parent / + child class shape. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + from sqlalchemy_continuum.exc import ClassNotVersioned + + from superset.connectors.sqla.models import SqlaTable, SqlMetric, TableColumn + from superset.models.dashboard import Dashboard + from superset.models.slice import Slice + + # ``ClassNotVersioned`` is the only expected failure here — versioning + # init runs at startup; if it didn't, every class lookup raises this. + # Narrowing the catch keeps a real underlying failure (e.g. a metadata + # inconsistency after ``make_versioned``) from being silently swallowed + # into a no-op retention pass. + parent_tables: list[sa.Table] = [] + for cls in (Dashboard, Slice, SqlaTable): + try: + parent_tables.append(version_class(cls).__table__) + except ClassNotVersioned: + logger.warning( + "retention: %s is not versioned; skipping shadow", cls.__name__ + ) + + child_tables: list[sa.Table] = [] + for cls in (TableColumn, SqlMetric): + try: + child_tables.append(version_class(cls).__table__) + except ClassNotVersioned: + logger.warning( + "retention: %s is not versioned; skipping shadow", cls.__name__ + ) + + metadata = parent_tables[0].metadata if parent_tables else None + m2m_table = ( + metadata.tables.get("dashboard_slices_version") + if metadata is not None + else None + ) + + return ShadowTables( + parent=parent_tables, + child=child_tables, + m2m=m2m_table, + transaction=tx_table, + ) + + +def _candidate_transaction_ids( + conn: sa.engine.Connection, + cutoff: datetime, + parent_tables: list[sa.Table], +) -> list[int]: + """Find ``version_transaction.id`` values that are eligible to + prune: ``issued_at < cutoff`` AND not currently the live row of + any versioned entity. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + tx_table = versioning_manager.transaction_cls.__table__ + candidate_ids = [ + row[0] + for row in conn.execute( + sa.select(tx_table.c.id).where(tx_table.c.issued_at < cutoff) + ) + ] + if not candidate_ids: + return [] + + # Build the set of transaction ids whose parent shadow includes a + # live row (``end_transaction_id IS NULL``). Those transactions + # represent the current state of an entity and must be preserved + # regardless of age. Chunked over candidate_ids to keep the bind- + # parameter count inside SQLite's ``SQLITE_MAX_VARIABLE_NUMBER`` + # floor (see ``_TX_ID_CHUNK_SIZE`` below). + preserved_ids: set[int] = set() + for ptbl in parent_tables: + for chunk in _chunked(candidate_ids, _TX_ID_CHUNK_SIZE): + for row in conn.execute( + sa.select(ptbl.c.transaction_id) + .where(ptbl.c.transaction_id.in_(chunk)) + .where(ptbl.c.end_transaction_id.is_(None)) + .distinct() + ): + preserved_ids.add(row[0]) + + return [tx_id for tx_id in candidate_ids if tx_id not in preserved_ids] + + +# SQLite's ``SQLITE_MAX_VARIABLE_NUMBER`` defaults to 999 (lifted to +# 32766 in 3.32+ but the older limit can still apply in shipped +# builds). Postgres + MySQL handle tens of thousands of bind params +# without complaint, so the chunk size is dictated by the SQLite floor. +# 500 leaves headroom for the ``transaction_id`` + ``end_transaction_id`` +# OR-pair (each ``tx_id`` is bound twice in the DELETE) plus a margin +# for any other bound params in the surrounding statement. +_TX_ID_CHUNK_SIZE = 500 + + +def _delete_for_transactions( + conn: sa.engine.Connection, + tables: list[sa.Table], + tx_ids: list[int], +) -> int: + """Delete shadow rows in *tables* whose lifespan touches a pruned + transaction — either ``transaction_id`` (created at) or + ``end_transaction_id`` (closed at) is in *tx_ids*. Returns total + rowcount across all tables. + + The ``end_transaction_id`` predicate is required to keep referential + integrity when transactions span multiple entities. A flush that + saves dashboard + slice + dataset at the same ``tx=X`` produces + three shadow rows sharing that tx. If only the dashboard is later + edited at ``tx=Y``, the dashboard row at ``tx=X`` is closed + (``end_tx=Y``) while the slice/dataset rows stay live at + ``tx=X``. Retention preserves ``tx=X`` (slice/dataset are live + there) and prunes ``tx=Y``. Without the ``end_tx`` predicate, the + dashboard's closed row at ``tx=X`` survives step 1 — its + ``end_transaction_id=Y`` then violates the FK when step 2 deletes + ``version_transaction`` row ``Y``. + + Live rows are never matched by either predicate + (``end_transaction_id IS NULL`` is not ``IN`` anything; live rows' + ``transaction_id`` is preserved by construction in + :func:`_candidate_transaction_ids`). + + ``tx_ids`` is chunked into batches of ``_TX_ID_CHUNK_SIZE`` so the + bind-parameter count stays inside SQLite's ``SQLITE_MAX_VARIABLE_ + NUMBER`` limit. Postgres and MySQL would happily accept the full + list, but the floor is dialect-agnostic since the retention task is + the only path that accumulates open-ended id batches. + """ + if not tx_ids: + return 0 + total = 0 + for tbl in tables: + for chunk in _chunked(tx_ids, _TX_ID_CHUNK_SIZE): + result = conn.execute( + sa.delete(tbl).where( + sa.or_( + tbl.c.transaction_id.in_(chunk), + tbl.c.end_transaction_id.in_(chunk), + ) + ) + ) + total += result.rowcount or 0 + return total + + +def _chunked(items: list[int], size: int) -> Iterator[list[int]]: + """Yield *items* in fixed-size lists. Final chunk may be smaller.""" + for i in range(0, len(items), size): + yield items[i : i + size] + + +#: Maximum number of attempts the prune will make before giving up. +#: A daily Celery beat schedule means the next chance is 24h out, so +#: a small inline retry materially improves the recovery time for the +#: serialization-conflict path. +_MAX_RETRY_ATTEMPTS = 3 + +#: Base for exponential backoff between retries (seconds). Worst-case +#: extra latency with the 3-attempt cap above and the factor below is +#: ``BASE + BASE * FACTOR`` = ~0.5s — well inside the prune's own +#: typical runtime. +_RETRY_BACKOFF_BASE_SECONDS = 0.1 + +#: Exponential-backoff multiplier between successive retry attempts. +#: Backoff for attempt N is ``BASE * (FACTOR ** (N - 1))``. +_RETRY_BACKOFF_FACTOR = 4 + +#: Statsd metric prefix for retention emissions. Mirrors the activity-view +#: orchestrator's ``superset.activity_view.*`` namespace so a single +#: Grafana filter (``superset.versioning.*``) catches both sides of the +#: feature. The pruned-count gauge fires every run; the skipped counter +#: fires for the "retention disabled" and "no versioned classes" cases; +#: the retried counter fires when the SERIALIZABLE block tripped at +#: least one conflict before settling. +_METRIC_PREFIX = "superset.versioning.retention" + + +def _run_prune_pass(cutoff: datetime, tables: ShadowTables) -> dict[str, Any]: + """One SERIALIZABLE pass of the prune. Caller wraps in the retry + loop so a serialization conflict re-opens a fresh connection + + transaction from a clean snapshot.""" + # The Celery task runs outside the request-bound DB session, so we + # use a fresh connection rather than ``db.session`` to avoid stepping + # on web-request state. + with ( + db.engine.connect().execution_options(isolation_level="SERIALIZABLE") as conn, + conn.begin(), + ): + tx_ids = _candidate_transaction_ids(conn, cutoff, tables.parent) + if not tx_ids: + return {"pruned_transactions": 0, "cutoff": cutoff.isoformat()} + + parent_rows = _delete_for_transactions(conn, tables.parent, tx_ids) + child_rows = _delete_for_transactions(conn, tables.child, tx_ids) + m2m_rows = ( + _delete_for_transactions(conn, [tables.m2m], tx_ids) + if tables.m2m is not None + else 0 + ) + + # Drop the version_transaction rows themselves. ON DELETE + # CASCADE on version_changes.transaction_id removes the + # associated change records automatically. Same SQLite bind- + # parameter chunking applies as the shadow deletes above. + tx_rows = 0 + for chunk in _chunked(tx_ids, _TX_ID_CHUNK_SIZE): + tx_rows += ( + conn.execute( + sa.delete(tables.transaction).where( + tables.transaction.c.id.in_(chunk) + ) + ).rowcount + or 0 + ) + + return { + "cutoff": cutoff.isoformat(), + "pruned_transactions": tx_rows, + "pruned_parent_shadows": parent_rows, + "pruned_child_shadows": child_rows, + "pruned_m2m_shadows": m2m_rows, + } + + +def _prune_old_versions_impl(retention_days: int) -> dict[str, Any]: + """Pure-Python implementation of the prune. Split out from the + Celery task wrapper so unit tests can call it directly without the + Celery harness. + + Returns a stats dict for logging / test assertions. + + Isolation level: SERIALIZABLE. The prune is logically a multi-step + read-then-write (candidate-vs-preserved SELECTs feeding the shadow + DELETEs). At READ COMMITTED there is a TOCTOU window — a save + committing between the preserved-ids snapshot and the DELETEs can + leave a stale view of which transaction ids are still serving as + the live row of some entity, and a shadow row that became live + mid-task can be silently dropped. SERIALIZABLE makes the prune + atomic against concurrent writers. SQLite is single-writer so + SERIALIZABLE is the only level available; MySQL InnoDB and Postgres + both support it natively. + + Postgres surfaces conflicts as ``SerializationFailure`` (a subclass + of ``sqlalchemy.exc.OperationalError``). The prune retries up to + ``_MAX_RETRY_ATTEMPTS`` with exponential backoff before giving up + and letting the outer Celery wrapper log + return ``{"error": 1}``. + Without the inline retry, a single conflict pushes the next attempt + 24 hours out (daily Celery beat), and under sustained write + pressure the prune can silently fail for many days in a row. + """ + if retention_days <= 0: + logger.info( + "version_history_retention: SUPERSET_VERSION_HISTORY_RETENTION_DAYS " + "<= 0; skipping", + ) + stats_logger_manager.instance.incr(f"{_METRIC_PREFIX}.skipped") + return {"skipped": 1} + + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + tables = _resolve_shadow_tables(versioning_manager.transaction_cls.__table__) + if not tables.parent: + logger.warning( + "version_history_retention: no versioned classes resolved; skipping", + ) + stats_logger_manager.instance.incr(f"{_METRIC_PREFIX}.skipped") + return {"skipped": 1} + + cutoff = datetime.utcnow() - timedelta(days=retention_days) + + for attempt in range(1, _MAX_RETRY_ATTEMPTS + 1): + try: + stats = _run_prune_pass(cutoff, tables) + except OperationalError as exc: + stats_logger_manager.instance.incr(f"{_METRIC_PREFIX}.retried") + if attempt == _MAX_RETRY_ATTEMPTS: + logger.warning( + "version_history_retention: gave up after %d attempts: %s", + _MAX_RETRY_ATTEMPTS, + exc, + ) + raise + backoff = _RETRY_BACKOFF_BASE_SECONDS * ( + _RETRY_BACKOFF_FACTOR ** (attempt - 1) + ) + logger.info( + "version_history_retention: attempt %d hit %s; retrying in %.2fs", + attempt, + type(exc).__name__, + backoff, + ) + time.sleep(backoff) + continue + if attempt > 1: + stats["retried"] = attempt - 1 + stats_logger_manager.instance.gauge( + f"{_METRIC_PREFIX}.pruned_transactions", + stats.get("pruned_transactions", 0), + ) + logger.info("version_history_retention: %s", stats) + return stats + + # The loop above always returns or re-raises; this is the type checker's + # placate-line. If it ever fires, the loop's exit condition has been + # changed incorrectly. + raise AssertionError("retention retry loop exited without result") + + +@celery_app.task(name="version_history.prune_old_versions") +def prune_old_versions() -> dict[str, Any]: + """Celery beat task entry point. Wraps the implementation with + config lookup + broad exception handling so a single failed run + doesn't poison the schedule (the next firing retries from a clean + slate). + """ + retention_days: int = current_app.config.get( + "SUPERSET_VERSION_HISTORY_RETENTION_DAYS", 30 + ) + try: + return _prune_old_versions_impl(retention_days) + except Exception: # pylint: disable=broad-except + logger.exception("version_history.prune_old_versions: task failed") + return {"error": 1} diff --git a/superset/translations/fr/LC_MESSAGES/messages.po b/superset/translations/fr/LC_MESSAGES/messages.po index 99d3cfe3741f..a9d291698362 100644 --- a/superset/translations/fr/LC_MESSAGES/messages.po +++ b/superset/translations/fr/LC_MESSAGES/messages.po @@ -18891,3 +18891,136 @@ msgstr "" msgid "© Layer attribution" msgstr "© Attribution de la couche" + +msgid "Show Value" +msgstr "Afficher la valeur" + +msgid "Truncate Axis" +msgstr "Tronquer l'axe" + +msgid "deck.gl charts" +msgstr "Graphiques deck.gl" + +# sc-103156 entity-versioning UI strings +msgid "Added" +msgstr "Ajouté" + +msgid "Removed" +msgstr "Supprimé" + +msgid "Changed" +msgstr "Modifié" + +msgid "Moved" +msgstr "Déplacé" + +msgid "Edited" +msgstr "Modifié" + +msgid "Baseline" +msgstr "Version initiale" + +msgid "Cleared %(field)s" +msgstr "%(field)s effacé" + +msgid "title" +msgstr "titre" + +msgid "chart name" +msgstr "nom du graphique" + +msgid "table name" +msgstr "nom de la table" + +msgid "chart" +msgstr "graphique" + +msgid "row" +msgstr "ligne" + +msgid "column" +msgstr "colonne" + +msgid "tab" +msgstr "onglet" + +msgid "tabs" +msgstr "onglets" + +msgid "header" +msgstr "en-tête" + +msgid "markdown" +msgstr "markdown" + +msgid "divider" +msgstr "séparateur" + +msgid "metric" +msgstr "mesure" + +msgid "Added %(kind)s" +msgstr "%(kind)s ajouté(e)" + +msgid "Added %(kind)s \"%(name)s\"" +msgstr "%(kind)s « %(name)s » ajouté(e)" + +msgid "Removed %(kind)s" +msgstr "%(kind)s supprimé(e)" + +msgid "Removed %(kind)s \"%(name)s\"" +msgstr "%(kind)s « %(name)s » supprimé(e)" + +msgid "Moved %(kind)s" +msgstr "%(kind)s déplacé(e)" + +msgid "Moved %(kind)s \"%(name)s\"" +msgstr "%(kind)s « %(name)s » déplacé(e)" + +msgid "Edited %(kind)s" +msgstr "%(kind)s modifié(e)" + +msgid "Edited %(kind)s \"%(name)s\"" +msgstr "%(kind)s « %(name)s » modifié(e)" + +msgid "Changed %(kind)s" +msgstr "%(kind)s modifié(e)" + +msgid "Changed %(kind)s \"%(name)s\"" +msgstr "%(kind)s « %(name)s » modifié(e)" + +msgid "Added %(kind)s %(detail)s" +msgstr "%(kind)s %(detail)s ajouté(e)" + +msgid "Removed %(kind)s %(detail)s" +msgstr "%(kind)s %(detail)s supprimé(e)" + +msgid "Changed %(kind)s %(detail)s" +msgstr "%(kind)s %(detail)s modifié(e)" + +msgid "Added chart %(id)s" +msgstr "Graphique %(id)s ajouté" + +msgid "Removed chart %(id)s" +msgstr "Graphique %(id)s supprimé" + +msgid "Changed chart %(id)s" +msgstr "Graphique %(id)s modifié" + +msgid "Added %(field)s" +msgstr "%(field)s ajouté" + +msgid "Removed %(field)s" +msgstr "%(field)s supprimé" + +msgid "Changed %(field)s" +msgstr "%(field)s modifié" + +msgid "Changed %(field)s to \"%(value)s\"" +msgstr "%(field)s changé en « %(value)s »" + +msgid "Set %(field)s to \"%(value)s\"" +msgstr "%(field)s défini à « %(value)s »" + +msgid "%(first)s (+%(more)s more)" +msgstr "%(first)s (+%(more)s autres)" diff --git a/superset/versioning/__init__.py b/superset/versioning/__init__.py new file mode 100644 index 000000000000..13a83393a912 --- /dev/null +++ b/superset/versioning/__init__.py @@ -0,0 +1,16 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. diff --git a/superset/versioning/activity/__init__.py b/superset/versioning/activity/__init__.py new file mode 100644 index 000000000000..9d17483ab00a --- /dev/null +++ b/superset/versioning/activity/__init__.py @@ -0,0 +1,102 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Read-side queries for the cross-entity activity-view API (sc-107283). + +Companion to :mod:`superset.versioning.queries`. Whereas ``queries.py`` +returns transaction-level history for a single entity, the helpers +here unify change-record history across an entity's transitive +dependency chain — a dashboard's activity stream includes edits to +charts that were attached to it AND edits to datasets those charts +pointed at, each time-bounded by when the relationship was active. + +One public entry point — :func:`get_activity` — dispatches on the +first argument to serve all three endpoint families: + +* ``get_activity(Dashboard, dashboard_uuid, ...)`` — own edits + + charts attached during their dashboard window + datasets those + charts used during their chart window. +* ``get_activity(Slice, chart_uuid, ...)`` — own edits + datasets the + chart pointed at during association. +* ``get_activity(SqlaTable, dataset_uuid, ...)`` — own edits only. + Datasets are not transitive recipients of activity in V2. + +Package layout (descends from public entry point to leaf helpers): + +* :mod:`.orchestrator` — :func:`get_activity` (public), the + ``activity_endpoint`` REST helper, the request param parser + (:func:`parse_activity_query_params`), and the observability + instrumentation that T037/T038 specify. +* :mod:`.scope` — scope resolution (DB-touching): + :func:`resolve_scope` / :func:`_resolve_dashboard_scope` / + :func:`_resolve_chart_scope` / :func:`_resolve_related_scope`. +* :mod:`.windows` — pure window arithmetic on half-open + ``[start_tx, end_tx)`` intervals: :func:`intersect_windows` / + :func:`union_windows` / :func:`merge_entity_windows` / + :func:`row_within_any_window`. Extracted from :mod:`.scope` so + :mod:`.queries` can import the pure helpers at module-top instead + of through a cycle-dodging lazy import. +* :mod:`.queries` — every DB-touching helper: Phase A relationship + walks, Phase B change-record fetch, name denormalization, + path-entity resolution, and tombstone-state lookup. +* :mod:`.impact` — per-record impact-count computation (the only + field that requires its own batched query). +* :mod:`.visibility` — the AV-008 silent visibility filter; uses + the same SQL-side access filters FAB applies on list endpoints. +* :mod:`.render` — record-decoration helpers that turn raw rows into + the ActivityRecord DTO (summary headlines, ``changed_by`` projection, + uuid lookup). +* :mod:`.kinds` — the kind-translation tables, the ``Window`` / + ``EntityWindows`` type aliases, and :func:`load_shadow_model`. + +The public surface (re-exported here) is the eight symbols below. +Sub-module privates are intentionally NOT re-exported — tests and +new internal callers should import them from their owning submodule +(e.g. ``from superset.versioning.activity.windows import +intersect_windows``) so the package's public API stays scannable. + +``PathEntityResponseError`` and ``resolve_endpoint_path_entity`` are +re-exported here from :mod:`superset.versioning.api_helpers` (where +they live alongside the ``/versions/`` endpoint handlers) so the +three ``/activity/`` endpoint callers can ``from +superset.versioning.activity import resolve_endpoint_path_entity`` +without crossing into the ``/versions/`` module name. +""" + +from __future__ import annotations + +from superset.versioning.activity.kinds import EntityWindows, Window +from superset.versioning.activity.orchestrator import ( + activity_endpoint, + ActivityParamsError, + get_activity, + parse_activity_query_params, +) +from superset.versioning.api_helpers import ( + PathEntityResponseError, + resolve_endpoint_path_entity, +) + +__all__ = [ + "ActivityParamsError", + "EntityWindows", + "PathEntityResponseError", + "Window", + "activity_endpoint", + "get_activity", + "parse_activity_query_params", + "resolve_endpoint_path_entity", +] diff --git a/superset/versioning/activity/impact.py b/superset/versioning/activity/impact.py new file mode 100644 index 000000000000..651f81f10604 --- /dev/null +++ b/superset/versioning/activity/impact.py @@ -0,0 +1,160 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Per-record impact computation for the activity DTO. + +Only dashboard-path activity records pointing at a ``SqlaTable`` +related entity carry an ``impact`` field — the number of charts on +the dashboard at that transaction that were pointing at the dataset. +This module computes that count in a single batched query per +request: + +* :func:`collect_impact_pairs` — pulls the distinct + ``(dataset_id, transaction_id)`` pairs that need counts. +* :func:`batch_chart_counts` — one SQL query joining + ``dashboard_slices_version`` and ``slices_version`` to count + the matching charts validity-strategy-style. +* :func:`impact_for_record` — pure projection from the pre-fetched + counts onto each record (returns ``None`` for non-Dashboard paths + or non-SqlaTable kinds, matching data-model.md §"``impact`` + computation"). + +Splitting the count batching from the pure projection keeps the SQL +inside one function (the batched read) and the per-record decoration +inside another (no DB). +""" + +from __future__ import annotations + +from typing import Any + +import sqlalchemy as sa + +from superset.extensions import db +from superset.versioning.activity.kinds import TABLE_KIND_TO_API + + +def collect_impact_pairs( + records: list[dict[str, Any]], path_kind: str +) -> set[tuple[int, int]]: + """Distinct ``(dataset_id, transaction_id)`` pairs from *records* + that require an impact computation per data-model.md. + + Only dashboard-path records whose related entity is a ``SqlaTable`` + produce a non-null ``impact`` field; for any other shape this set + is empty and no DB query needs to fire. + """ + if path_kind != "Dashboard": + return set() + return { + (record["entity_id"], record["transaction_id"]) + for record in records + if TABLE_KIND_TO_API.get(record["entity_kind"]) == "SqlaTable" + } + + +def batch_chart_counts( + dashboard_id: int, pairs: set[tuple[int, int]] +) -> dict[tuple[int, int], int]: + """For every ``(dataset_id, target_tx)`` in *pairs*, count the + distinct charts that were both on *dashboard_id* and pointing at + *dataset_id* at *target_tx*. + + One SELECT against ``dashboard_slices_version`` ⨝ ``slices_version``, + pulling the (slice, dataset, validity-window) state for every slice + ever on the dashboard whose dataset matches one of the requested + dataset_ids. The Python loop then applies the validity-strategy + predicate per pair. Replaces the previous N+1 shape that fired one + COUNT per related record. + + Returns ``{(dataset_id, target_tx): count}``; pairs whose count + would be zero are omitted so the caller's ``.get(key, 0)`` is + correct. + """ + if not pairs: + return {} + + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.slice import Slice + + metadata = version_class(Slice).__table__.metadata + m2m_tbl = metadata.tables.get("dashboard_slices_version") + slices_tbl = version_class(Slice).__table__ + if m2m_tbl is None: + return {} + + dataset_ids = {dataset_id for dataset_id, _ in pairs} + stmt = sa.select( + m2m_tbl.c.slice_id, + slices_tbl.c.datasource_id, + m2m_tbl.c.transaction_id.label("m2m_start"), + m2m_tbl.c.end_transaction_id.label("m2m_end"), + slices_tbl.c.transaction_id.label("slice_start"), + slices_tbl.c.end_transaction_id.label("slice_end"), + ).where( + m2m_tbl.c.dashboard_id == dashboard_id, + m2m_tbl.c.operation_type != 2, + slices_tbl.c.id == m2m_tbl.c.slice_id, + slices_tbl.c.datasource_id.in_(dataset_ids), + slices_tbl.c.datasource_type == "table", + slices_tbl.c.operation_type != 2, + ) + rows = db.session.connection().execute(stmt).mappings().all() + + # For each pair, collect the slice_ids whose two validity windows + # both straddle target_tx. ``set`` dedupes within a pair. + matches: dict[tuple[int, int], set[int]] = {} + pairs_by_dataset: dict[int, list[int]] = {} + for dataset_id, target_tx in pairs: + pairs_by_dataset.setdefault(dataset_id, []).append(target_tx) + + for row in rows: + ds_id = row["datasource_id"] + for target_tx in pairs_by_dataset.get(ds_id, ()): + in_m2m = row["m2m_start"] <= target_tx and ( + row["m2m_end"] is None or row["m2m_end"] > target_tx + ) + in_slice = row["slice_start"] <= target_tx and ( + row["slice_end"] is None or row["slice_end"] > target_tx + ) + if in_m2m and in_slice: + matches.setdefault((ds_id, target_tx), set()).add(row["slice_id"]) + + return {pair: len(slice_ids) for pair, slice_ids in matches.items()} + + +def impact_for_record( + record: dict[str, Any], + path_kind: str, + counts: dict[tuple[int, int], int], +) -> dict[str, int] | None: + """Synthesize the ``impact`` field for one record using the pre- + fetched *counts* mapping. Pure function — no DB. + + Per data-model.md §"``impact`` computation": only + ``path=Dashboard`` and ``related=SqlaTable`` shapes carry an + impact; everything else returns ``None``. + """ + api_kind = TABLE_KIND_TO_API.get(record["entity_kind"]) + if path_kind != "Dashboard" or api_kind != "SqlaTable": + return None + key = (record["entity_id"], record["transaction_id"]) + chart_count = counts.get(key, 0) + if chart_count == 0: + return None + return {"charts": chart_count} diff --git a/superset/versioning/activity/kinds.py b/superset/versioning/activity/kinds.py new file mode 100644 index 000000000000..4c11a05ad13c --- /dev/null +++ b/superset/versioning/activity/kinds.py @@ -0,0 +1,186 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Kind translation tables, shared types, and the shadow-model loader. + +The activity-view module operates in two "kind" namespaces — the +table-stored value (``"chart"`` / ``"dashboard"`` / ``"dataset"``) that +``version_changes.entity_kind`` carries, and the Python class-name +form (``"Slice"`` / ``"Dashboard"`` / ``"SqlaTable"``) used internally +for dispatch and returned in the DTO's ``entity_kind`` field. The four +mappings here translate between them. Adjacent kind-keyed dicts live +here too: the per-kind human-readable label, the user-facing +lowercase form, and the 404 exception class. + +The :func:`load_shadow_model` helper exists in the same module +because each lookup is keyed on the same set of class names — keeping +it adjacent to the mappings makes the kind-translation surface +discoverable at a glance. +""" + +from __future__ import annotations + +from dataclasses import dataclass + +from flask_appbuilder import Model + +from superset.commands.chart.exceptions import ChartNotFoundError +from superset.commands.dashboard.exceptions import DashboardNotFoundError +from superset.commands.dataset.exceptions import DatasetNotFoundError +from superset.versioning.changes import ENTITY_KIND_BY_CLASS_NAME + +# ---- Kind translation ----------------------------------------------------- + +# ``version_changes.entity_kind`` stores the friendly downstream-tooling +# value (``"chart"``, ``"dashboard"``, ``"dataset"``) per sc-103156's +# ``ENTITY_KIND_BY_CLASS_NAME``. The activity-view DTO returns the +# Python class name instead (``"Slice"``, ``"Dashboard"``, +# ``"SqlaTable"``) so the contract aligns with ``__class__.__name__`` +# (data-model.md §"``ActivityRecord`` DTO"). Translate at the boundary. +TABLE_KIND_TO_API: dict[str, str] = { + table_kind: class_name + for class_name, table_kind in ENTITY_KIND_BY_CLASS_NAME.items() +} +API_KIND_TO_TABLE: dict[str, str] = dict(ENTITY_KIND_BY_CLASS_NAME) + +# Human-readable label for AV-012 summary headlines +# ("Dataset updated: Sales Transactions"). Keyed by the internal API kind +# (Python class name; matches ``model_cls.__name__``). +API_KIND_LABEL: dict[str, str] = { + "Dashboard": "Dashboard", + "Slice": "Chart", + "SqlaTable": "Dataset", +} + +# User-facing lowercase rendering of the kind. This is what appears in +# the JSON response's ``entity_kind`` field and the +# ``ActivityRecordSchema.entity_kind`` enum. Internal code keeps the +# Python class-name form because it matches ``model_cls.__name__`` and is +# convenient for dispatch — translation happens at serialization time +# only, in :func:`render.apply_record_decoration`. +USER_FACING_KIND: dict[str, str] = { + "Dashboard": "dashboard", + "Slice": "chart", + "SqlaTable": "dataset", +} + +# 404 exception class per API kind. Each accepts a string positional arg +# (the path-entity UUID) that gets formatted into the exception message. +NOT_FOUND_EXC: dict[str, type[Exception]] = { + "Dashboard": DashboardNotFoundError, + "Slice": ChartNotFoundError, + "SqlaTable": DatasetNotFoundError, +} + +# Per-API-kind (model class name, display column) used by +# ``_resolve_names_for_kind`` to read the user-facing entity name from +# the shadow table valid at a given transaction. +NAME_COLUMN: dict[str, tuple[str, str]] = { + "Dashboard": ("Dashboard", "dashboard_title"), + "Slice": ("Slice", "slice_name"), + "SqlaTable": ("SqlaTable", "table_name"), +} + + +# ---- Types ---------------------------------------------------------------- + + +@dataclass(frozen=True) +class Window: + """A validity window in Continuum transaction-id space, half-open as + ``[start_tx, end_tx)``. + + Immutable and equal-by-attributes — two windows with the same + ``start_tx`` / ``end_tx`` are interchangeable. Constructor rejects + ``end_tx <= start_tx``. ``end_tx = None`` means "open ended + (current)" and acts as positive infinity throughout the helpers. + + Helper methods (``contains`` / ``intersect`` / ``merges_with``) + live on the type so callers don't re-implement the half-open + predicate inline. Previously a ``tuple[int, int | None]`` alias; + promoted to a dataclass so a function accepting a ``Window`` can't + silently accept any other 2-tuple and so the constructor enforces + the half-open invariant. + """ + + start_tx: int + end_tx: int | None + + def __post_init__(self) -> None: + if self.end_tx is not None and self.end_tx <= self.start_tx: + raise ValueError( + f"Window end_tx must be > start_tx; " + f"got [{self.start_tx}, {self.end_tx})" + ) + + def contains(self, tx_id: int) -> bool: + """``True`` iff *tx_id* falls inside this half-open interval.""" + return self.start_tx <= tx_id and (self.end_tx is None or tx_id < self.end_tx) + + def intersect(self, other: Window) -> Window | None: + """Return the clipped overlap of this window with *other*, or + ``None`` when they are disjoint. ``end_tx = None`` acts as + positive infinity on either side.""" + start = max(self.start_tx, other.start_tx) + end: int | None + if self.end_tx is None: + end = other.end_tx + elif other.end_tx is None: + end = self.end_tx + else: + end = min(self.end_tx, other.end_tx) + if end is not None and end <= start: + return None + return Window(start, end) + + def merges_with(self, other: Window) -> bool: + """``True`` iff *self* and *other* overlap or touch (so their + union is one contiguous window). Assumes the caller has placed + them in start-ascending order.""" + if self.end_tx is None: + # self extends to +∞; everything past it merges in. + return True + return other.start_tx <= self.end_tx + + +#: A related-entity scope row: ``(api_kind, entity_id, [windows])``. +#: ``api_kind`` is the DTO-facing kind (``"Slice"``, etc.), not the +#: table-stored kind. Left as a tuple alias for now — the +#: ``(api_kind, entity_id)`` pair is logically a key with the window +#: list as its value, so a dict shape may fit better than a flat +#: dataclass when this is revisited. +EntityWindows = tuple[str, int, list[Window]] + + +def load_shadow_model(model_name: str) -> type[Model]: + """Inline-import a shadow model class by name. Deferred until call + time because the versioning package is initialised before all model + mappers are configured (same idiom used throughout + :mod:`superset.versioning.changes`).""" + # pylint: disable=import-outside-toplevel + if model_name == "Dashboard": + from superset.models.dashboard import Dashboard + + return Dashboard + if model_name == "Slice": + from superset.models.slice import Slice + + return Slice + if model_name == "SqlaTable": + from superset.connectors.sqla.models import SqlaTable + + return SqlaTable + raise LookupError(f"No shadow class registered for {model_name!r}") diff --git a/superset/versioning/activity/orchestrator.py b/superset/versioning/activity/orchestrator.py new file mode 100644 index 000000000000..9f6d81180ec1 --- /dev/null +++ b/superset/versioning/activity/orchestrator.py @@ -0,0 +1,352 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Top-level orchestrator + query-param parsing + observability. + +This is the public entry point for the activity-view read path. One +function — :func:`get_activity` — dispatches on the path entity's +model class to assemble the cross-entity activity stream: + +1. ``resolve_path_entity`` (queries.py) — resolve UUID → live entity. +2. ``resolve_scope`` (scope.py) — build the related-entity window list. +3. ``fetch_change_records`` (queries.py) — pull rows from + ``version_changes`` joined with ``version_transaction`` and ``ab_user``. +4. ``filter_records_by_visibility`` (visibility.py) — silent AV-008 + drop of records the requester can't read. +5. ``apply_entity_name_denormalization`` (queries.py) — resolve entity names + from the shadow row valid at each record's transaction_id. +6. ``apply_record_decoration`` (render.py) — synthesize the ActivityRecord + DTO fields and strip internal-only columns. +7. Paginate in Python over the post-filter list. + +Parameter parsing for the REST endpoints lives here too — +:func:`parse_activity_query_params` is called by the three +``/activity/`` endpoint handlers before they call ``get_activity``. +Same for the observability instrumentation: ``_phase_timer`` and +``_emit_request_shape_attributes`` emit the metrics that T037/T038 +specify, on the same prefix the cross-coupling test pins. +""" + +from __future__ import annotations + +import contextlib +from collections.abc import Iterator +from datetime import datetime, timezone +from typing import Any +from uuid import UUID + +from flask import Response +from flask_appbuilder import Model + +from superset.versioning.activity.kinds import EntityWindows +from superset.versioning.activity.queries import ( + apply_entity_name_denormalization, + fetch_change_records, + resolve_path_entity, +) +from superset.versioning.activity.render import apply_record_decoration +from superset.versioning.activity.scope import resolve_scope +from superset.versioning.activity.visibility import filter_records_by_visibility +from superset.versioning.api_helpers import ( + PathEntityResponseError, + resolve_endpoint_path_entity, +) + +_DEFAULT_PAGE_SIZE = 25 +_MAX_PAGE_SIZE = 200 +_VALID_INCLUDE_VALUES: frozenset[str] = frozenset({"self", "related", "all"}) + + +class ActivityParamsError(ValueError): + """Raised by :func:`parse_activity_query_params` when a query param is + malformed. The endpoint catches this and maps to ``response_400``; + no other callers should depend on the exception type.""" + + +def parse_activity_query_params(args: Any) -> dict[str, Any]: + """Parse the ``since`` / ``until`` / ``include`` / ``page`` / ``page_size`` + query parameters into the kwargs ``get_activity`` accepts. + + Raises :class:`ActivityParamsError` (subclass of ``ValueError``) when + a parameter is malformed. Shared across the three endpoint families + (dashboards, charts, datasets) so the parsing and 400-messaging stay + consistent. + """ + params: dict[str, Any] = { + "include": _parse_include(args.get("include", "all")), + "page": _parse_page(args.get("page", "0")), + "page_size": _parse_page_size(args.get("page_size")), + } + if (since := _parse_optional_iso(args.get("since"), name="since")) is not None: + params["since"] = since + if (until := _parse_optional_iso(args.get("until"), name="until")) is not None: + params["until"] = until + return params + + +def _parse_optional_iso(raw: str | None, *, name: str) -> datetime | None: + """Parse a missing-or-ISO-datetime field; ``None`` for missing, + ``ActivityParamsError`` for malformed.""" + if not raw: + return None + parsed = _parse_iso_datetime(raw) + if parsed is None: + raise ActivityParamsError(f"Invalid {name!r} datetime: {raw!r}") + return parsed + + +def _parse_include(value: str) -> str: + if value not in _VALID_INCLUDE_VALUES: + raise ActivityParamsError( + f"Invalid 'include' value: {value!r}; " + f"must be one of {sorted(_VALID_INCLUDE_VALUES)}" + ) + return value + + +def _parse_page(raw: str) -> int: + try: + value = int(raw) + except (TypeError, ValueError) as exc: + raise ActivityParamsError(f"Invalid 'page' value: {raw!r}") from exc + if value < 0: + raise ActivityParamsError("Invalid 'page' value: must be >= 0") + return value + + +def _parse_page_size(raw: str | None) -> int: + """``page_size`` honours the default when missing, raises when invalid, + and silently clamps to ``_MAX_PAGE_SIZE`` (so ``?page_size=500`` + returns 200 records instead of a 400).""" + if raw is None: + return _DEFAULT_PAGE_SIZE + try: + value = int(raw) + except (TypeError, ValueError) as exc: + raise ActivityParamsError(f"Invalid 'page_size' value: {raw!r}") from exc + if value < 1: + raise ActivityParamsError("Invalid 'page_size' value: must be >= 1") + return min(value, _MAX_PAGE_SIZE) + + +def _parse_iso_datetime(value: str) -> datetime | None: + """Parse an ISO-8601 datetime string. Tolerates the trailing ``Z`` + suffix that Python <3.11 ``fromisoformat`` rejects, and normalises any + timezone-aware result to naive UTC. + + The ``since`` / ``until`` filters bind directly against + ``version_transaction.issued_at``, which is ``sa.DateTime()`` — a + timezone-*naive* column (UTC by convention). Binding a tz-aware value + against it shifts the comparison by the session offset on PostgreSQL + (and raises on some drivers), so collapse aware inputs to naive UTC + here. Naive inputs pass through unchanged (already treated as UTC). + """ + candidate = value[:-1] + "+00:00" if value.endswith("Z") else value + try: + parsed = datetime.fromisoformat(candidate) + except ValueError: + return None + if parsed.tzinfo is not None: + parsed = parsed.astimezone(timezone.utc).replace(tzinfo=None) + return parsed + + +def get_activity( + model_cls: type[Model], + entity_uuid: UUID, + *, + since: datetime | None = None, + until: datetime | None = None, + include: str = "all", + page: int = 0, + page_size: int = _DEFAULT_PAGE_SIZE, +) -> tuple[list[dict[str, Any]], int]: + """Cross-entity activity stream for one path entity. + + Single polymorphic entry point. Dispatches on *model_cls* to + assemble the path entity's self records plus the transitive related- + entity records (charts attached to a dashboard, datasets a chart + pointed at, etc.) per data-model.md §"Query phases". + + Returns ``(records, total_count)``. The count is post-visibility + (AV-008) and post-include-filter, not just the size of the returned + slice — clients paginate by passing ``page`` forward until + ``page * page_size >= count``. + + Raises ``DashboardNotFoundError`` / ``ChartNotFoundError`` / + ``DatasetNotFoundError`` when the path entity doesn't exist (AV-009). + """ + _path_entity, path_id = resolve_path_entity(model_cls, entity_uuid) + path_kind = model_cls.__name__ + kind_key = path_kind.lower() # "dashboard" / "slice" / "sqlatable" + + with _phase_timer(kind_key, "relationship_resolution_ms"): + entity_windows = resolve_scope(path_kind, path_id, include) + if not entity_windows: + _emit_request_shape_attributes( + kind_key, + include=include, + has_since_filter=since is not None, + page_size=page_size, + record_count=0, + entity_windows=[], + path_kind=path_kind, + path_id=path_id, + ) + return [], 0 + + # Visibility filter runs before decoration: it needs the raw + # ``entity_id`` column (which decoration strips), and dropping + # invisible records early means we don't pay for name lookup + + # tombstone probes + impact counts on records the requester + # can't see (AV-008's silent-filter contract). + with _phase_timer(kind_key, "fetch_ms"): + records = fetch_change_records(entity_windows, since, until) + with _phase_timer(kind_key, "visibility_filter_ms"): + records = filter_records_by_visibility(records) + with _phase_timer(kind_key, "denormalize_ms"): + apply_entity_name_denormalization(records) + with _phase_timer(kind_key, "decorate_ms"): + apply_record_decoration(records, path_kind, path_id) + + total = len(records) + bounded_size = max(1, min(page_size, _MAX_PAGE_SIZE)) + offset = max(0, page) * bounded_size + + _emit_request_shape_attributes( + kind_key, + include=include, + has_since_filter=since is not None, + page_size=bounded_size, + record_count=total, + entity_windows=entity_windows, + path_kind=path_kind, + path_id=path_id, + ) + + return records[offset : offset + bounded_size], total + + +def activity_endpoint( + api: Any, model_cls: type[Model], uuid_str: str, request_args: Any +) -> Response: + """Body of ``GET /api/v1/{resource}//activity/``. + + Same shape as :func:`superset.versioning.api_helpers.list_versions_endpoint` + for the ``/versions/`` endpoint family. Resolves the path entity, + parses the request query params, runs :func:`get_activity`, and + wraps the result through ``ActivityResponseSchema``. + + *api* is the FAB ``ModelRestApi`` instance (pass ``self`` from the + endpoint method). *request_args* is ``request.args`` from + ``flask.request`` — passed explicitly so the helper is testable + without a live Flask context. + """ + # pylint: disable=import-outside-toplevel + from superset.versioning.schemas import ActivityResponseSchema + + try: + entity, _ = resolve_endpoint_path_entity(api, model_cls, uuid_str) + except PathEntityResponseError as exc: + return exc.response + + try: + params = parse_activity_query_params(request_args) + except ActivityParamsError as exc: + return api.response_400(message=str(exc)) + + records, count = get_activity(model_cls, entity.uuid, **params) + payload = ActivityResponseSchema().dump({"result": records, "count": count}) + return api.response(200, **payload) + + +# ---- Observability (T037 / T038) ------------------------------------------ + +#: Common prefix for every metric this module emits. Per plan §D-17. +_METRIC_PREFIX = "superset.activity_view" + + +@contextlib.contextmanager +def _phase_timer(kind_key: str, phase: str) -> Iterator[None]: + """Time the wrapped block and emit + ``superset.activity_view..`` to ``stats_logger_manager``. + Wrapper around :func:`superset.utils.decorators.stats_timing` that + centralises the key construction. + """ + # pylint: disable=import-outside-toplevel + from superset.extensions import stats_logger_manager + from superset.utils.decorators import stats_timing + + with stats_timing( + f"{_METRIC_PREFIX}.{kind_key}.{phase}", + stats_logger_manager.instance, + ): + yield + + +def _emit_request_shape_attributes( + kind_key: str, + *, + include: str, + has_since_filter: bool, + page_size: int, + record_count: int, + entity_windows: list[EntityWindows], + path_kind: str, + path_id: int, +) -> None: + """Emit non-PII shape counters about the request and its result set. + + Per T038: include_mode / has_since_filter / page_size / record_count + + per-related-kind entity counts. **No PII**: entity names, diff + content, user identifiers — none of those reach the metric layer. + The counters use ``incr`` (counters) since they're tags, not + latencies; the timing keys above carry the latency dimension. + """ + # pylint: disable=import-outside-toplevel + from superset.extensions import stats_logger_manager + + sl = stats_logger_manager.instance + + # Tag-style metrics: one counter per attribute value. The statsd + # bridge accepts arbitrary strings; downstream dashboards filter by + # the value segment. + sl.incr(f"{_METRIC_PREFIX}.{kind_key}.requests.include_{include}") + sl.incr( + f"{_METRIC_PREFIX}.{kind_key}.requests." + f"has_since_filter_{'true' if has_since_filter else 'false'}" + ) + sl.gauge(f"{_METRIC_PREFIX}.{kind_key}.page_size", float(page_size)) + sl.gauge(f"{_METRIC_PREFIX}.{kind_key}.record_count", float(record_count)) + + # Per-related-kind entity counts (T038 explicit fields). The scope + # list includes the path entity itself (the "self" window); exclude + # it so the gauge reflects only the *related* entities the request + # fanned out to, not "this request touched itself". + by_kind: dict[str, int] = {"Slice": 0, "SqlaTable": 0, "Dashboard": 0} + for api_kind, entity_id, _windows in entity_windows: + if (api_kind, entity_id) == (path_kind, path_id): + continue + if api_kind in by_kind: + by_kind[api_kind] += 1 + sl.gauge( + f"{_METRIC_PREFIX}.{kind_key}.related_entity_count.charts", + float(by_kind["Slice"]), + ) + sl.gauge( + f"{_METRIC_PREFIX}.{kind_key}.related_entity_count.datasets", + float(by_kind["SqlaTable"]), + ) diff --git a/superset/versioning/activity/queries.py b/superset/versioning/activity/queries.py new file mode 100644 index 000000000000..5e4cd74021a9 --- /dev/null +++ b/superset/versioning/activity/queries.py @@ -0,0 +1,518 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""DB-touching helpers for the activity-view read path. + +All Phase A relationship walks (``charts_attached_to_dashboard``, +``datasets_used_by_chart``, ``batch_datasets_used_by_charts``), +the Phase B change-record fetch (``fetch_change_records`` / +``_select_change_rows_for_kinds``), the name-denormalization helpers +(``_resolve_names_for_kind`` / ``apply_entity_name_denormalization``), the +path-entity resolution helper (``resolve_path_entity``), and the +tombstone-state lookup (``check_entity_tombstones``) live here. + +Each helper is a thin SELECT-and-shape function — no orchestration, +no business logic. Callers in :mod:`scope`, :mod:`render`, and +:mod:`orchestrator` compose them into the end-to-end request. + +**Inline imports.** Continuum's ``version_class`` / ``versioning_manager`` +and the Superset model classes are imported inside each helper because +this package is loaded from ``init_versioning()`` before all SQLAlchemy +mappers are configured. +""" + +from __future__ import annotations + +from collections.abc import Iterator +from datetime import datetime +from typing import Any +from uuid import UUID + +import sqlalchemy as sa +from flask_appbuilder import Model + +from superset.extensions import db +from superset.versioning.activity.kinds import ( + API_KIND_TO_TABLE, + EntityWindows, + load_shadow_model, + NAME_COLUMN, + NOT_FOUND_EXC, + TABLE_KIND_TO_API, + Window, +) +from superset.versioning.activity.windows import row_within_any_window +from superset.versioning.changes import version_changes_table + +# ---- Path-entity resolution ----------------------------------------------- + + +def resolve_path_entity(model_cls: type[Model], entity_uuid: UUID) -> tuple[Any, int]: + """Resolve *entity_uuid* to ``(live_entity, entity_id)`` or raise a + typed 404 per AV-009. + + Soft-delete handling (sc-103157) is inherited transparently from + :func:`superset.versioning.queries.find_active_by_uuid` once it + learns to filter out ``deleted_at IS NOT NULL`` rows; at that point + soft-deleted paths will also raise here. + """ + # pylint: disable=import-outside-toplevel + from superset.versioning.queries import find_active_by_uuid + + entity = find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + api_kind = model_cls.__name__ + exc_cls = NOT_FOUND_EXC.get(api_kind) + if exc_cls is None: + raise LookupError( + f"Activity view does not support model class {api_kind!r}" + ) + raise exc_cls(str(entity_uuid)) + return entity, entity.id + + +# ---- Phase A: relationship-traversal queries ------------------------------ + + +def charts_attached_to_dashboard(dashboard_id: int) -> list[tuple[int, Window]]: + """Return ``(slice_id, window)`` for every chart that has ever been on + *dashboard_id*, with each association's validity window in + transaction-id space. + + Reads from ``dashboard_slices_version`` (Continuum's auto-generated + M2M shadow). Rows with ``operation_type = 2`` (DELETE) are excluded + so we don't synthesize a phantom window from a detachment row. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.dashboard import Dashboard + + metadata = version_class(Dashboard).__table__.metadata + m2m_tbl = metadata.tables.get("dashboard_slices_version") + if m2m_tbl is None: + return [] + + rows = ( + db.session.connection() + .execute( + sa.select( + m2m_tbl.c.slice_id, + m2m_tbl.c.transaction_id, + m2m_tbl.c.end_transaction_id, + ).where( + m2m_tbl.c.dashboard_id == dashboard_id, + m2m_tbl.c.operation_type != 2, + m2m_tbl.c.slice_id.is_not(None), + ) + ) + .all() + ) + return [(row[0], Window(row[1], row[2])) for row in rows] + + +def datasets_used_by_chart(slice_id: int) -> list[tuple[int, Window]]: + """Return ``(datasource_id, window)`` for every dataset that *slice_id* + has ever pointed at, with each association's validity window. + + Single-slice form, used by ``_resolve_chart_scope`` where there + is only one chart to walk. The dashboard-scope path calls + :func:`batch_datasets_used_by_charts` instead so the query fires + once for all slices on the dashboard, not once per slice. + + Reads from ``slices_version`` (the chart parent shadow). Filters to + ``datasource_type = 'table'`` because the activity view only follows + the chart → ``SqlaTable`` dependency edge (not legacy/other + datasources). Rows with ``operation_type = 2`` are excluded. + """ + return batch_datasets_used_by_charts({slice_id}).get(slice_id, []) + + +def batch_datasets_used_by_charts( + slice_ids: set[int], +) -> dict[int, list[tuple[int, Window]]]: + """Batch form of :func:`datasets_used_by_chart`. Returns + ``{slice_id: [(dataset_id, window), ...]}`` in a single query so the + dashboard-scope walker doesn't fire one query per chart on the + dashboard. The previous per-slice shape became O(n_charts) round- + trips, which dominated ``get_activity`` latency on dashboards with + rich history (profile run 2026-05-26 showed `resolve_scope` + accounting for ~1.9s out of 4s p95). + """ + if not slice_ids: + return {} + + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.slice import Slice + + slices_tbl = version_class(Slice).__table__ + rows = ( + db.session.connection() + .execute( + sa.select( + slices_tbl.c.id, + slices_tbl.c.datasource_id, + slices_tbl.c.transaction_id, + slices_tbl.c.end_transaction_id, + ).where( + slices_tbl.c.id.in_(slice_ids), + slices_tbl.c.datasource_type == "table", + slices_tbl.c.operation_type != 2, + slices_tbl.c.datasource_id.is_not(None), + ) + ) + .mappings() + .all() + ) + grouped: dict[int, list[tuple[int, Window]]] = {} + for row in rows: + grouped.setdefault(row["id"], []).append( + ( + row["datasource_id"], + Window(row["transaction_id"], row["end_transaction_id"]), + ) + ) + return grouped + + +# ---- Phase B: change-record fetch ----------------------------------------- + + +def fetch_change_records( + entity_window_tuples: list[EntityWindows], + since: datetime | None, + until: datetime | None, +) -> list[dict[str, Any]]: + """Fetch all ``version_changes`` rows matching any of the supplied + entity-window tuples, joined with ``version_transaction`` for + ``issued_at`` and ``user_id``. + + Each tuple is ``(api_kind, entity_id, [(start_tx, end_tx), ...])``; + a record matches when ``entity_kind`` equals the table-stored form + of *api_kind*, ``entity_id`` matches, and ``transaction_id`` falls + inside at least one of the entity's windows. ``since``/``until`` + further restrict by ``issued_at``. + + Implementation: one SELECT per kind with ``entity_id IN (...)`` and + a wide ``transaction_id`` bound (the union of all windows for that + kind). Per-window precision is applied in Python afterward. This + keeps the SQL shape proportional to the number of *kinds* (≤3) and + the bound proportional to the union of windows, not the cross- + product of (entity, window) — which previously generated one OR + clause per (entity, window) pair and hit SQLite's + ``SQLITE_MAX_EXPR_DEPTH`` limit on dashboards with many slices + or many historical attachment windows. + + Per AV-008 the visibility filter runs after this function (records + the requester can't read are silently dropped and must not + contribute to ``count``), so the orchestrator paginates in Python + over the filtered list — no DB-level ``LIMIT``/``OFFSET`` here. + + Returned rows are ordered by ``(issued_at DESC, transaction_id DESC, + sequence DESC)`` — the secondary keys break ties for AV-006's + stable-ordering contract. + """ + if not entity_window_tuples: + return [] + + # Group windows by (table_kind, entity_id) and by table_kind for SQL + # narrowing. The fetch is per-kind; the post-filter is per-entity. + windows_by_entity: dict[tuple[str, int], list[Window]] = {} + ids_by_kind: dict[str, set[int]] = {} + for api_kind, entity_id, windows in entity_window_tuples: + table_kind = API_KIND_TO_TABLE.get(api_kind) + if table_kind is None or not windows: + continue + ids_by_kind.setdefault(table_kind, set()).add(entity_id) + windows_by_entity.setdefault((table_kind, entity_id), []).extend(windows) + + if not ids_by_kind: + return [] + + rows = _select_change_rows_for_kinds(ids_by_kind, since, until) + filtered = [ + row + for row in rows + if row_within_any_window( + row, windows_by_entity.get((row["entity_kind"], row["entity_id"]), []) + ) + ] + filtered.sort( + key=lambda r: (r["issued_at"], r["transaction_id"], r["sequence"]), + reverse=True, + ) + return filtered + + +def _select_change_rows_for_kinds( + ids_by_kind: dict[str, set[int]], + since: datetime | None, + until: datetime | None, +) -> list[dict[str, Any]]: + """Fire one SELECT per entity_kind with ``entity_id IN (...)``; + concatenate the results. Each SELECT joins ``version_transaction`` + + ``ab_user`` so the orchestrator has the columns it needs for + decoration. + + Per-kind, not one query: SQLAlchemy's ``tuple_(entity_kind, + entity_id).in_(...)`` would collapse the three queries into one, + but its SQL emission is not portable across Postgres, MySQL, and + SQLite. The per-kind shape is the correct trade-off given + Superset's multi-dialect requirement (at most 3 round-trips per + request, bounded by the kind taxonomy). Do not "optimise" into a + composite-tuple IN clause without verifying the SQL on all three + dialects. + + **Init-order dependency.** ``tx_tbl.c.action_kind`` resolves only + after ``init_versioning()`` has run — the column is appended onto + Continuum's transaction Table by + ``superset.versioning.factory.VersionTransactionFactory`` at app + start via ``append_column`` + ``add_property``. This helper is + safe to call from request-path code because the app is fully + initialised by then; calling it from a script that imports the + versioning package without going through ``init_versioning()`` + will raise ``AttributeError`` on the ``action_kind`` attribute + access below.""" + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + from superset import security_manager + + tx_tbl = versioning_manager.transaction_cls.__table__ + user_tbl = security_manager.user_model.__table__ + vc = version_changes_table + join_tree = vc.join(tx_tbl, vc.c.transaction_id == tx_tbl.c.id).outerjoin( + user_tbl, tx_tbl.c.user_id == user_tbl.c.id + ) + select_cols = ( + vc.c.transaction_id, + vc.c.entity_kind, + vc.c.entity_id, + vc.c.sequence, + vc.c.kind, + vc.c.operation, + vc.c.path, + vc.c.from_value, + vc.c.to_value, + tx_tbl.c.issued_at, + tx_tbl.c.user_id, + # ``action_kind`` is the high-level avenue (restore / import / + # clone / NULL=ordinary save) stamped by the originating + # command via the change-record listener. All records sharing a + # ``transaction_id`` share the same value. The column is + # declared on the Continuum Table by ``VersionTransactionFactory``, + # so ``tx_tbl.c.action_kind`` resolves cleanly here. See + # sc-103156 data-model.md §"Three dimensions". + tx_tbl.c.action_kind, + user_tbl.c.id.label("changed_by_id"), + user_tbl.c.first_name, + user_tbl.c.last_name, + ) + + out: list[dict[str, Any]] = [] + for table_kind, entity_ids in ids_by_kind.items(): + # Chunk ``entity_ids`` to stay inside SQLite's + # ``SQLITE_MAX_VARIABLE_NUMBER`` floor (default 999, raised to + # 32766 in 3.32+ but the older limit ships in many builds). The + # bind count grows linearly with chart-on-dashboard count; a + # dashboard built from a huge chart library can reach the floor. + # Postgres + MySQL accept the full list, but the chunk is + # dialect-agnostic for simplicity. + for chunk in _chunked_ids(entity_ids, _ENTITY_ID_CHUNK_SIZE): + stmt = ( + sa.select(*select_cols) + .select_from(join_tree) + .where( + vc.c.entity_kind == table_kind, + vc.c.entity_id.in_(chunk), + ) + ) + if since is not None: + stmt = stmt.where(tx_tbl.c.issued_at >= since) + if until is not None: + stmt = stmt.where(tx_tbl.c.issued_at < until) + out.extend( + dict(row) + for row in db.session.connection().execute(stmt).mappings().all() + ) + return out + + +# Bind-parameter floor: see ``_select_change_rows_for_kinds`` docstring. +# 500 leaves room for the two literal-string filters and the optional +# since/until datetime params. +_ENTITY_ID_CHUNK_SIZE = 500 + + +def _chunked_ids(ids: set[int], size: int) -> Iterator[list[int]]: + """Yield *ids* in fixed-size lists. Final chunk may be smaller.""" + items = list(ids) + for i in range(0, len(items), size): + yield items[i : i + size] + + +# ---- Name denormalization ------------------------------------------------- + + +def _resolve_names_for_kind( + api_kind: str, pairs: set[tuple[int, int]] +) -> dict[tuple[int, int], str]: + """For one entity kind, return ``{(entity_id, target_tx): name}`` from + the shadow row valid at *target_tx* (validity-strategy predicate). + Empty mapping when the kind has no name column registered. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + if api_kind not in NAME_COLUMN: + return {} + + model_name, name_col = NAME_COLUMN[api_kind] + model_cls = load_shadow_model(model_name) + shadow_tbl = version_class(model_cls).__table__ + ids = sorted({eid for eid, _ in pairs}) + rows = ( + db.session.connection() + .execute( + sa.select( + shadow_tbl.c.id, + shadow_tbl.c.transaction_id, + shadow_tbl.c.end_transaction_id, + shadow_tbl.c[name_col], + ).where(shadow_tbl.c.id.in_(ids)) + ) + .all() + ) + per_entity: dict[int, list[tuple[int, int | None, Any]]] = {} + for row in rows: + per_entity.setdefault(row[0], []).append((row[1], row[2], row[3])) + + resolved: dict[tuple[int, int], str] = {} + for entity_id, target_tx in pairs: + for start_tx, end_tx, name in per_entity.get(entity_id, []): + if start_tx <= target_tx and (end_tx is None or end_tx > target_tx): + resolved[(entity_id, target_tx)] = name + break + return resolved + + +def apply_entity_name_denormalization(records: list[dict[str, Any]]) -> None: + """Resolve each record's ``entity_name`` from the shadow row valid at + its ``transaction_id``. Adds an ``entity_name`` key to every record + in place; returns ``None``. + + The lookup is per (table-stored ``entity_kind``, ``entity_id``, + ``transaction_id``) triple. One ``IN``-clause query per kind keeps + round-trips bounded by the number of distinct kinds (≤3) regardless + of result-set size. The in-place mutation avoids re-allocating + thousands of dicts on hot dashboards; the name + return signature + make the side effect explicit instead of pretending to be a pure + projection. + """ + if not records: + return + + needed_by_kind: dict[str, set[tuple[int, int]]] = {} + for record in records: + api_kind = TABLE_KIND_TO_API.get(record["entity_kind"]) + if api_kind is None or api_kind not in NAME_COLUMN: + continue + needed_by_kind.setdefault(api_kind, set()).add( + (record["entity_id"], record["transaction_id"]) + ) + + resolved: dict[tuple[str, int, int], str] = {} + for api_kind, pairs in needed_by_kind.items(): + for (entity_id, target_tx), name in _resolve_names_for_kind( + api_kind, pairs + ).items(): + resolved[(api_kind, entity_id, target_tx)] = name + + for record in records: + api_kind_for_record = TABLE_KIND_TO_API.get(record["entity_kind"], "") + key = (api_kind_for_record, record["entity_id"], record["transaction_id"]) + record["entity_name"] = resolved.get(key, "") + + +# ---- Live-row existence + soft-delete state ------------------------------- + + +def check_entity_tombstones( + distinct_entities: set[tuple[str, int]], +) -> dict[tuple[str, int], dict[str, Any]]: + """For each ``(api_kind, entity_id)``, report ``deleted`` (no live + row) and ``deletion_state`` (``"soft_deleted"`` iff the live row has + a non-null ``deleted_at`` per sc-103157, else ``None``). + + Pre-sc-103157 the model classes don't have a ``deleted_at`` column; + we probe with ``hasattr`` and report ``deletion_state=None`` + universally in that case. Once sc-103157 lands, this helper picks up + the new column automatically. + """ + result: dict[tuple[str, int], dict[str, Any]] = {} + if not distinct_entities: + return result + + by_kind: dict[str, list[int]] = {} + for api_kind, entity_id in distinct_entities: + by_kind.setdefault(api_kind, []).append(entity_id) + + # ``no_autoflush`` mirrors the defensive posture of the listener- + # side reads. Today's callers run from request-path code with no + # pending writes; a future caller that probes tombstones before a + # flush would otherwise trigger autoflush mid-read. + with db.session.no_autoflush: + for api_kind, entity_ids in by_kind.items(): + if api_kind not in NAME_COLUMN: + for entity_id in entity_ids: + result[(api_kind, entity_id)] = { + "deleted": True, + "deletion_state": None, + } + continue + + model_name, _ = NAME_COLUMN[api_kind] + model_cls = load_shadow_model(model_name) + live_tbl = model_cls.__table__ # type: ignore[attr-defined] + has_deleted_at = "deleted_at" in live_tbl.c + + cols = [live_tbl.c.id] + if has_deleted_at: + cols.append(live_tbl.c.deleted_at) + rows = ( + db.session.connection() + .execute(sa.select(*cols).where(live_tbl.c.id.in_(entity_ids))) + .all() + ) + live: dict[int, Any] = {} + for row in rows: + live[row[0]] = row[1] if has_deleted_at else None + + for entity_id in entity_ids: + if entity_id not in live: + result[(api_kind, entity_id)] = { + "deleted": True, + "deletion_state": None, + } + else: + deleted_at = live[entity_id] + result[(api_kind, entity_id)] = { + "deleted": False, + "deletion_state": "soft_deleted" if deleted_at else None, + } + return result diff --git a/superset/versioning/activity/render.py b/superset/versioning/activity/render.py new file mode 100644 index 000000000000..3e7993e6eaf3 --- /dev/null +++ b/superset/versioning/activity/render.py @@ -0,0 +1,218 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Decoration: turn raw change-records into the ActivityRecord DTO. + +After fetching + filtering, each record needs the synthesized fields +the API contract documents — ``entity_kind`` translated to the user- +facing form, ``entity_uuid``, ``entity_deleted`` / +``entity_deletion_state``, ``source`` (self vs. related), +``summary`` (the AV-012 headline), ``impact`` (chart-count for +dashboard→dataset records), ``version_uuid``, ``changed_by``. + +This module collects all those decorations: + +* :func:`apply_record_decoration` — orchestrates the per-page additions in + one pass: pulls tombstones + uuids + impact counts in batches, then + walks records adding the synthesized fields and stripping the + internal-only columns the API contract doesn't expose. +* :func:`_lookup_entity_uuids` — one IN-clause query per kind to + resolve live ``uuid`` for non-tombstoned entities. +* :func:`_build_summary` — pure projection of (api_kind, record kind, + entity_name) onto the AV-012 headline string. +* :func:`_changed_by_dict` — projects the user columns onto the + ``changed_by`` DTO shape. +""" + +from __future__ import annotations + +from typing import Any +from uuid import UUID + +import sqlalchemy as sa + +from superset.extensions import db +from superset.versioning.activity.impact import ( + batch_chart_counts, + collect_impact_pairs, + impact_for_record, +) +from superset.versioning.activity.kinds import ( + API_KIND_LABEL, + load_shadow_model, + NAME_COLUMN, + TABLE_KIND_TO_API, + USER_FACING_KIND, +) +from superset.versioning.activity.queries import check_entity_tombstones +from superset.versioning.queries import derive_version_uuid + +_SUMMARY_VERBS: dict[str, str] = { + # The kind taxonomy from FR-016 mapped to past-tense verbs for the + # AV-012 " : " headline. "field" is + # the fallback for scalar changes that don't map to a named verb. + "filter": "filter changed", + "metric": "metric changed", + "dimension": "dimension changed", + "column": "column changed", + "chart": "chart changed", + "time_range": "time range changed", + "color_palette": "palette changed", + "restore": "restored", + "field": "updated", +} + + +def apply_record_decoration( + records: list[dict[str, Any]], + path_kind: str, + path_id: int, +) -> None: + """Add the synthesized ActivityRecord fields to each record in place: + ``entity_kind`` (translated to API form), ``entity_uuid``, + ``entity_deleted``, ``entity_deletion_state``, ``source``, + ``summary``, ``impact``, ``version_uuid``, ``changed_by``. + + Mutates *records* in place; returns ``None``. Records are expected + to already carry ``entity_name`` from + :func:`apply_entity_name_denormalization`. The in-place mutation + avoids re-allocating thousands of dicts on hot dashboards; the + name + return signature make the side effect explicit instead of + pretending to be a pure projection. + """ + if not records: + return + + distinct: set[tuple[str, int]] = { + ( + TABLE_KIND_TO_API.get(r["entity_kind"], ""), + r["entity_id"], + ) + for r in records + if TABLE_KIND_TO_API.get(r["entity_kind"]) + } + tombstones = check_entity_tombstones(distinct) + uuids = _lookup_entity_uuids(distinct, tombstones) + # Pre-compute impact counts for the whole page in one batch query + # instead of one COUNT per related record (was N+1). + impact_counts = batch_chart_counts( + path_id, collect_impact_pairs(records, path_kind) + ) + + for record in records: + api_kind = TABLE_KIND_TO_API.get(record["entity_kind"], "") + entity_id = record["entity_id"] + tombstone = tombstones.get( + (api_kind, entity_id), {"deleted": True, "deletion_state": None} + ) + entity_uuid = uuids.get((api_kind, entity_id)) + is_self = api_kind == path_kind and entity_id == path_id + + # Emit the user-facing form ("dashboard"/"chart"/"dataset") on the + # wire; the internal class-name (api_kind) is kept above for the + # remaining decoration steps that key off model_cls.__name__. + record["entity_kind"] = USER_FACING_KIND.get(api_kind, api_kind) + record["entity_uuid"] = str(entity_uuid) if entity_uuid else None + record["entity_deleted"] = tombstone["deleted"] + record["entity_deletion_state"] = tombstone["deletion_state"] + record["source"] = "self" if is_self else "related" + record["version_uuid"] = ( + str(derive_version_uuid(entity_uuid, record["transaction_id"])) + if entity_uuid + else None + ) + record["changed_by"] = _changed_by_dict(record) + + if is_self: + record["summary"] = "" + record["impact"] = None + else: + record["summary"] = _build_summary(api_kind, record) + record["impact"] = impact_for_record(record, path_kind, impact_counts) + + # Strip the internal-only columns the API contract doesn't expose. + for key in ( + "entity_id", + "sequence", + "user_id", + "changed_by_id", + "first_name", + "last_name", + ): + record.pop(key, None) + + +def _lookup_entity_uuids( + distinct: set[tuple[str, int]], + tombstones: dict[tuple[str, int], dict[str, Any]], +) -> dict[tuple[str, int], UUID | None]: + """Batch-fetch live ``uuid`` per ``(api_kind, entity_id)``. Tombstoned + entities are skipped (their ``entity_uuid`` is null per data-model.md). + """ + result: dict[tuple[str, int], UUID | None] = {} + by_kind: dict[str, list[int]] = {} + for api_kind, entity_id in distinct: + if tombstones.get((api_kind, entity_id), {}).get("deleted"): + continue + by_kind.setdefault(api_kind, []).append(entity_id) + + # ``no_autoflush`` mirrors the defensive posture of the baseline + + # change-record listeners: this helper reads from live tables to + # resolve uuids, and a future caller that resolves an entity before + # the parent flush would otherwise trigger autoflush mid-read. + # Today's call sites run from request-path code with no pending + # session state, so the cost of the guard is zero. + with db.session.no_autoflush: + for api_kind, entity_ids in by_kind.items(): + if api_kind not in NAME_COLUMN: + continue + model_cls = load_shadow_model(NAME_COLUMN[api_kind][0]) + live_tbl = model_cls.__table__ # type: ignore[attr-defined] + rows = ( + db.session.connection() + .execute( + sa.select(live_tbl.c.id, live_tbl.c.uuid).where( + live_tbl.c.id.in_(entity_ids) + ) + ) + .all() + ) + for row in rows: + result[(api_kind, row[0])] = row[1] + return result + + +def _build_summary(api_kind: str, record: dict[str, Any]) -> str: + """Build the AV-012 headline for a related record: + ``" : "``.""" + label = API_KIND_LABEL.get(api_kind, api_kind) + verb = _SUMMARY_VERBS.get(record.get("kind", ""), "updated") + name = record.get("entity_name") or "" + return f"{label} {verb}: {name}" if name else f"{label} {verb}" + + +def _changed_by_dict(record: dict[str, Any]) -> dict[str, Any] | None: + """Project the user columns onto the ``changed_by`` shape, or + ``None`` when no Flask user was attached to the save (CLI / Celery) + or when the user has since been deleted from ``ab_user``. + """ + if record.get("changed_by_id") is None: + return None + return { + "id": record["changed_by_id"], + "first_name": record.get("first_name"), + "last_name": record.get("last_name"), + } diff --git a/superset/versioning/activity/scope.py b/superset/versioning/activity/scope.py new file mode 100644 index 000000000000..f24140c68823 --- /dev/null +++ b/superset/versioning/activity/scope.py @@ -0,0 +1,103 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Scope resolution — turn a path entity into the related-entity walk. + +Composes :mod:`~superset.versioning.activity.queries` (Phase A +relationship walks) and :mod:`~superset.versioning.activity.windows` +(pure interval arithmetic) into the +``list[EntityWindows]`` scope that +:func:`~superset.versioning.activity.queries.fetch_change_records` +consumes. + +The functions here read the DB (via the Phase A helpers in +:mod:`~superset.versioning.activity.queries`); the pure window- +arithmetic functions previously colocated here now live in +:mod:`~superset.versioning.activity.windows` so the package no longer +needs a lazy import to dodge a ``scope ↔ queries`` cycle. +""" + +from __future__ import annotations + +from superset.versioning.activity.kinds import EntityWindows, Window +from superset.versioning.activity.queries import ( + batch_datasets_used_by_charts, + charts_attached_to_dashboard, + datasets_used_by_chart, +) +from superset.versioning.activity.windows import ( + intersect_windows, + merge_entity_windows, +) + + +def resolve_scope(path_kind: str, path_id: int, include: str) -> list[EntityWindows]: + """Build the ``[(api_kind, entity_id, [windows])]`` list that + :func:`~superset.versioning.activity.queries.fetch_change_records` + consumes, branching by *path_kind* and *include* mode.""" + want_self = include in ("all", "self") + want_related = include in ("all", "related") + + scope: list[EntityWindows] = [] + if want_self: + scope.append((path_kind, path_id, [Window(0, None)])) + if want_related: + scope.extend(_resolve_related_scope(path_kind, path_id)) + return scope + + +def _resolve_related_scope(path_kind: str, path_id: int) -> list[EntityWindows]: + """Walk the dependency edges from the path entity to its related + entities. Per AV-004, datasets have no transitive layer in V2.""" + if path_kind == "Dashboard": + return _resolve_dashboard_scope(path_id) + if path_kind == "Slice": + return _resolve_chart_scope(path_id) + return [] + + +def _resolve_dashboard_scope(dashboard_id: int) -> list[EntityWindows]: + """Charts on the dashboard during their attachment window, plus + datasets each chart pointed at during the intersection of (chart- + attachment, chart-on-dataset).""" + scope: list[EntityWindows] = [] + chart_windows: dict[int, list[Window]] = {} + for slice_id, window in charts_attached_to_dashboard(dashboard_id): + chart_windows.setdefault(slice_id, []).append(window) + + # One query for the dataset-history of every chart on the dashboard, + # not one query per chart. The per-slice form was O(n_charts) round- + # trips which dominated p95 on rich dashboards. + dataset_windows_by_slice = batch_datasets_used_by_charts(set(chart_windows)) + + for slice_id, attachment_windows in chart_windows.items(): + scope.append(("Slice", slice_id, list(attachment_windows))) + dataset_windows = dataset_windows_by_slice.get(slice_id, []) + for attachment in attachment_windows: + for dataset_id, chart_dataset_window in dataset_windows: + if ( + intersect := intersect_windows(attachment, chart_dataset_window) + ) is not None: + scope.append(("SqlaTable", dataset_id, [intersect])) + return merge_entity_windows(scope) + + +def _resolve_chart_scope(slice_id: int) -> list[EntityWindows]: + """Datasets the chart pointed at over its full history.""" + scope: list[EntityWindows] = [] + for dataset_id, window in datasets_used_by_chart(slice_id): + scope.append(("SqlaTable", dataset_id, [window])) + return merge_entity_windows(scope) diff --git a/superset/versioning/activity/visibility.py b/superset/versioning/activity/visibility.py new file mode 100644 index 000000000000..5e4c376ed91e --- /dev/null +++ b/superset/versioning/activity/visibility.py @@ -0,0 +1,203 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Per-AV-008 silent visibility filter for activity-view records. + +Drops records whose source entity the requester can't read. Silent in +the sense that dropped records contribute no count and no placeholder +to the response — the user sees only what they're entitled to see, and +the response shape can't be used to infer the existence of entities +they're gated out of. + +Visibility is resolved SQL-side via each resource's existing FAB +access filter (``DashboardAccessFilter`` / ``ChartFilter`` / +``DatasourceFilter``). Two SQL queries per kind (one for live ids, one +for the access-filtered subset) replace the N-call +``security_manager.can_access_(entity)`` loop that dominated +latency on dashboard-scope responses with many related entities +(sqlalchemy-review W-NEW-1). + +**Inline imports.** ``_resolve_visibility`` defers the FAB-filter +imports (``DashboardAccessFilter`` / ``ChartFilter`` / +``DatasourceFilter`` and ``SQLAInterface``) until call time. Same +init-order rationale as :mod:`superset.versioning.baseline` — +``versioning`` packages are imported from ``init_versioning()`` +before all model mappers are configured, and the filter classes pull +in their resource's model graph (Chart → Dataset → Database for +``ChartFilter``, etc.); a module-top import would trip mapper +resolution before Continuum's ``make_versioned()`` has finished. + +**Integration shape.** The activity-view consumes FAB's access-filter +classes (``DashboardAccessFilter`` / ``ChartFilter`` / +``DatasourceFilter``) directly rather than translating them or +re-implementing the predicate. That keeps the activity stream's +visibility posture identical to the list endpoints' — operator- +controlled and stable — at the cost of coupling to FAB's exact +filter shape. Future entities added to the activity surface must +extend the dispatch table in ``_resolve_visibility`` to include +their access-filter class. +""" + +from __future__ import annotations + +import logging +from typing import Any + +from superset.extensions import db +from superset.versioning.activity.kinds import ( + load_shadow_model, + NAME_COLUMN, + TABLE_KIND_TO_API, +) + +logger = logging.getLogger(__name__) + + +def filter_records_by_visibility( + records: list[dict[str, Any]], +) -> list[dict[str, Any]]: + """Drop records whose source entity the requester can't read. + + Per AV-008 the filter is silent: dropped records contribute no count + and no placeholder. Tombstoned entities (no live row) pass through + — the decorator step marks them ``entity_deleted: true`` and the + payload exposes no navigable ``entity_uuid``, so there's nothing + sensitive left to gate. + + Visibility is resolved SQL-side via each resource's existing access + filter, which reads the requesting user from Flask-Login internally + (no explicit user parameter threads through here). If a CLI/Celery + bypass becomes necessary in the future, add it then with a real call + site. + """ + if not records: + return records + + distinct: set[tuple[str, int]] = { + ( + TABLE_KIND_TO_API.get(r["entity_kind"], r["entity_kind"]), + r["entity_id"], + ) + for r in records + } + visible = _resolve_visibility(distinct) + return [ + r + for r in records + if visible.get( + ( + TABLE_KIND_TO_API.get(r["entity_kind"], r["entity_kind"]), + r["entity_id"], + ), + True, # tombstone / unknown kind → pass through + ) + ] + + +def _resolve_visibility( + distinct_entities: set[tuple[str, int]], +) -> dict[tuple[str, int], bool]: + """Return ``{(api_kind, entity_id): can_read}`` for the live row of + each entity. Missing live rows (tombstoned) map to ``True`` — the + decorator handles the deleted-state messaging separately. + + Visibility is computed SQL-side via each resource's existing access + filter (``DashboardAccessFilter`` / ``ChartFilter`` / + ``DatasourceFilter``). These are the same filters FAB's + ``ModelRestApi`` applies to ``base_filters`` on list endpoints, so + the activity-view visibility check matches the rest of the read + surface byte-for-byte. Two queries per kind (one for live ids, one + for the access-filtered subset) replace the N-call + ``security_manager.can_access_(entity)`` loop that dominated + latency on dashboard-scope activity responses with many related + entities (sqlalchemy-review W-NEW-1). + """ + # pylint: disable=import-outside-toplevel + from flask_appbuilder.models.sqla.interface import SQLAInterface + + from superset.charts.filters import ChartFilter + from superset.dashboards.filters import DashboardAccessFilter + from superset.views.base import DatasourceFilter + + access_filter_classes: dict[str, type] = { + "Dashboard": DashboardAccessFilter, + "Slice": ChartFilter, + "SqlaTable": DatasourceFilter, + } + + by_kind: dict[str, list[int]] = {} + for api_kind, entity_id in distinct_entities: + by_kind.setdefault(api_kind, []).append(entity_id) + + visible: dict[tuple[str, int], bool] = {} + for api_kind, entity_ids in by_kind.items(): + if api_kind in NAME_COLUMN and api_kind not in access_filter_classes: + # The kind is in the change-records taxonomy but is missing + # an access-filter wiring — almost certainly a future-entity + # addition that updated ``TABLE_KIND_TO_API`` but forgot + # the visibility dispatch. Fail closed: the activity stream + # must not silently disclose change records for an entity + # whose access predicate is unimplemented. Warn so the gap + # surfaces in CI / staging logs before production. + logger.warning( + "activity visibility: no access filter wired for kind %r; " + "denying %d records", + api_kind, + len(entity_ids), + ) + for entity_id in entity_ids: + visible[(api_kind, entity_id)] = False + continue + if api_kind not in NAME_COLUMN: + # Kind isn't in the change-records taxonomy at all — not + # something the activity-view emits today. Pass through so + # the decorator can mark it as a tombstone if appropriate. + for entity_id in entity_ids: + visible[(api_kind, entity_id)] = True + continue + model_cls = load_shadow_model(NAME_COLUMN[api_kind][0]) + + # Live ids — what exists at all. Used to decide tombstone vs + # not-visible: an id missing from this set is tombstoned and + # passes through (True); an id in this set but absent from the + # access-filtered set is denied (False). + live_ids = { + row[0] + for row in db.session.query(model_cls.id) # type: ignore[attr-defined] + .filter(model_cls.id.in_(entity_ids)) # type: ignore[attr-defined] + .all() + } + + # Apply the SQL-side access filter to a query restricted to the + # candidate ids. Same predicate FAB uses for list endpoints, so + # results are consistent with the rest of the read surface. + access_filter = access_filter_classes[api_kind]("id", SQLAInterface(model_cls)) + visible_ids = { + row[0] + for row in access_filter.apply( + db.session.query(model_cls.id).filter( # type: ignore[attr-defined] + model_cls.id.in_(entity_ids) # type: ignore[attr-defined] + ), + value=None, + ).all() + } + + for entity_id in entity_ids: + if entity_id not in live_ids: + visible[(api_kind, entity_id)] = True + else: + visible[(api_kind, entity_id)] = entity_id in visible_ids + return visible diff --git a/superset/versioning/activity/windows.py b/superset/versioning/activity/windows.py new file mode 100644 index 000000000000..b76ae43c5dc5 --- /dev/null +++ b/superset/versioning/activity/windows.py @@ -0,0 +1,108 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Pure window arithmetic on half-open ``[start_tx, end_tx)`` intervals. + +Extracted from the DB-touching scope resolution so that: + +* :mod:`scope` (DB-touching) can import this module at module-top. +* :mod:`queries.fetch_change_records` can import + :func:`row_within_any_window` at module-top instead of through a + lazy import that previously dodged a ``scope ↔ queries`` cycle. + +Everything here is pure Python — no DB, no Flask. ``end_tx = None`` +means "open-ended (current)" and behaves like positive infinity. +""" + +from __future__ import annotations + +from typing import Any + +from superset.versioning.activity.kinds import EntityWindows, Window + + +def intersect_windows(outer: Window, inner: Window) -> Window | None: + """Intersect two half-open ``[start_tx, end_tx)`` windows. + + Returns the clipped overlap, or ``None`` when they are disjoint. + ``end_tx = None`` means "open ended (current)" and acts like + positive infinity. Thin wrapper over :meth:`Window.intersect` — + kept as a free function so callers and tests don't have to migrate + to method form in lockstep with the dataclass promotion. + """ + return outer.intersect(inner) + + +def row_within_any_window(row: dict[str, Any], windows: list[Window]) -> bool: + """``True`` iff ``row['transaction_id']`` falls inside at least one + of *windows*. Half-open interval semantics match + :func:`intersect_windows`.""" + if not windows: + return False + tx_id = row["transaction_id"] + return any(w.contains(tx_id) for w in windows) + + +def merge_entity_windows(scope: list[EntityWindows]) -> list[EntityWindows]: + """Collapse repeated ``(api_kind, entity_id)`` entries by unioning + their window lists, and collapse overlapping/touching windows + within each entity into one. + + The OR-clause in + :func:`~superset.versioning.activity.queries.fetch_change_records` + generates one branch per (kind, id, window) tuple. Without the + within-entity union, a chart that's been attached-and-detached + many times (or that repeated fixture loads have populated the M2M + shadow for) yields a separate clause per redundant window — at + ~10 entities × ~50 windows the SQL hits SQLite's + ``SQLITE_MAX_EXPR_DEPTH`` (1000). Merging here keeps the clause + count proportional to the number of *distinct* validity intervals, + not the number of shadow rows. + """ + merged: dict[tuple[str, int], list[Window]] = {} + for api_kind, entity_id, windows in scope: + merged.setdefault((api_kind, entity_id), []).extend(windows) + return [ + (api_kind, entity_id, union_windows(windows)) + for (api_kind, entity_id), windows in merged.items() + ] + + +def union_windows(windows: list[Window]) -> list[Window]: + """Sort + merge overlapping/touching half-open intervals. + + Pure function — no DB. Touching ``[a, b)`` and ``[b, c)`` merge into + ``[a, c)``. ``end_tx = None`` (open-ended) absorbs everything to its + right. Returns a minimal disjoint cover of the input set. + """ + if not windows: + return [] + sorted_windows = sorted(windows, key=lambda w: w.start_tx) + out: list[Window] = [sorted_windows[0]] + for current in sorted_windows[1:]: + prev = out[-1] + if prev.end_tx is None: + # Prior window is open-ended; it absorbs everything past. + continue + if current.start_tx <= prev.end_tx: + # Overlapping or touching — extend the prior window. + new_end: int | None = ( + None if current.end_tx is None else max(prev.end_tx, current.end_tx) + ) + out[-1] = Window(prev.start_tx, new_end) + else: + out.append(current) + return out diff --git a/superset/versioning/api_helpers.py b/superset/versioning/api_helpers.py new file mode 100644 index 000000000000..cbe888b9e3c0 --- /dev/null +++ b/superset/versioning/api_helpers.py @@ -0,0 +1,237 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Shared handlers for the ``/versions/`` and ``/activity/`` REST endpoints. + +Each ``ChartRestApi`` / ``DashboardRestApi`` / ``DatasetRestApi`` carries +the same three ``/versions/`` endpoint methods — ``list_versions``, +``get_version``, ``restore_version`` — plus the ``activity`` endpoint +on each resource. The bodies were byte-for-byte identical apart from +the model class, the ``security_manager.raise_for_access`` kwarg, and +the resource-specific exception triplet on the restore path. + +Extracting the bodies here lets each per-resource method collapse to +a single delegation call, while the OpenAPI docstring + FAB decorators +stay at the method site where they belong. +""" + +from __future__ import annotations + +import logging +from dataclasses import dataclass +from typing import Any +from uuid import UUID + +from flask import Response +from flask_appbuilder import Model + +from superset.daos.version import VersionDAO +from superset.exceptions import SupersetSecurityException +from superset.extensions import security_manager +from superset.versioning.etag import set_version_etag_by_uuid + +logger = logging.getLogger(__name__) + + +@dataclass(frozen=True) +class RestoreEndpointSpec: + """Per-resource configuration for :func:`restore_version_endpoint`. + + Bundles the five fields that differ across the three /versions/restore + endpoint families (chart / dashboard / dataset) so the endpoint + function signature stays at four call-time parameters instead of + nine. Each per-resource RestApi declares a module-level instance + (e.g. ``_CHART_RESTORE_SPEC``) and passes it through. + + All fields are required; the dataclass is frozen so the spec can be + safely declared as a module-level constant. + """ + + command_cls: type + not_found_exc: type[Exception] + forbidden_exc: type[Exception] + update_failed_exc: type[Exception] + resource_label: str + + +# Maps the versioned model class name to the keyword argument +# ``security_manager.raise_for_access`` expects for the per-resource +# gate. Slice → ``chart=``, Dashboard → ``dashboard=``, SqlaTable → +# ``datasource=``. Centralised here so /versions/ and /activity/ +# endpoints share one source of truth for the dispatch. +_RAISE_FOR_ACCESS_KWARG: dict[str, str] = { + "Slice": "chart", + "Dashboard": "dashboard", + "SqlaTable": "datasource", +} + + +class PathEntityResponseError(Exception): + """Carries a pre-built error ``Response`` from + :func:`resolve_endpoint_path_entity`. Endpoints catch it and return + the carried response directly. The shape exists so the + UUID-parse + find-by-uuid + read-access check can live in one + place across the ``/versions/`` and ``/activity/`` endpoint + families.""" + + def __init__(self, response: Any) -> None: + super().__init__("PathEntityResponseError") + self.response = response + + +def resolve_endpoint_path_entity( + api: Any, model_cls: type[Model], uuid_str: str +) -> tuple[Any, UUID]: + """Run the standard path-entity preflight for a /versions/ or + /activity/ endpoint: + + 1. Parse *uuid_str* into a UUID (or raise → 400). + 2. Look up the live entity via ``VersionDAO.find_active_by_uuid`` + (or raise → 404). + 3. Run ``security_manager.raise_for_access`` with the resource-typed + kwarg (or raise → 403). + + Returns ``(entity, entity_uuid)`` on success — the parsed UUID is + threaded out so callers don't re-parse the path-string. Raises + :class:`PathEntityResponseError` carrying the appropriate error + Response on any failure; the endpoint method should:: + + try: + entity, entity_uuid = resolve_endpoint_path_entity( + self, Dashboard, uuid_str + ) + except PathEntityResponseError as exc: + return exc.response + + *api* is the FAB ``ModelRestApi`` instance — we call + ``api.response_400`` / ``api.response_403`` / ``api.response_404`` + on it. Pass ``self`` from the endpoint method. + """ + try: + entity_uuid = UUID(uuid_str) + except ValueError as exc: + raise PathEntityResponseError(api.response_400(message="Invalid UUID")) from exc + + entity = VersionDAO.find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + raise PathEntityResponseError(api.response_404()) + + # Direct ``[…]`` would leak the unknown model name into a generic 500 + # via the unhandled ``KeyError`` exception text. The three resource + # families wired today cover every key; a future entity added to the + # versioning surface without updating this dispatch table should fail + # closed (the test suite picks it up) rather than silently disclose. + kwarg = _RAISE_FOR_ACCESS_KWARG.get(model_cls.__name__) + if kwarg is None: + raise LookupError( + f"No raise_for_access kwarg registered for {model_cls.__name__!r}" + ) + try: + security_manager.raise_for_access(**{kwarg: entity}) + except SupersetSecurityException as exc: + raise PathEntityResponseError(api.response_403()) from exc + + return entity, entity_uuid + + +def list_versions_endpoint( + api: Any, + model_cls: type[Model], + uuid_str: str, +) -> Response: + """Body of ``GET /api/v1/{resource}//versions/``.""" + try: + entity, entity_uuid = resolve_endpoint_path_entity(api, model_cls, uuid_str) + except PathEntityResponseError as exc: + return exc.response + + versions = VersionDAO.list_versions(model_cls, entity_uuid, entity=entity) + if versions is None: + return api.response_404() + return set_version_etag_by_uuid( + api.response(200, result=versions, count=len(versions)), + model_cls, + entity_uuid, + entity_id=entity.id, + ) + + +def get_version_endpoint( + api: Any, + model_cls: type[Model], + uuid_str: str, + version_uuid_str: str, +) -> Response: + """Body of ``GET /api/v1/{resource}//versions//``.""" + try: + entity, entity_uuid = resolve_endpoint_path_entity(api, model_cls, uuid_str) + except PathEntityResponseError as exc: + return exc.response + + try: + version_uuid = UUID(version_uuid_str) + except ValueError: + return api.response_400(message="Invalid version UUID") + + snapshot = VersionDAO.get_version( + model_cls, entity_uuid, version_uuid, entity=entity + ) + if snapshot is None: + return api.response_404() + return set_version_etag_by_uuid( + api.response(200, result=snapshot), + model_cls, + entity_uuid, + entity_id=entity.id, + ) + + +def restore_version_endpoint( + api: Any, + model_cls: type[Model], + uuid_str: str, + version_uuid_str: str, + spec: RestoreEndpointSpec, +) -> Response: + """Body of ``POST /api/v1/{resource}//versions//restore``. + + Does not use :func:`resolve_endpoint_path_entity` — the restore + command runs its own ownership / existence checks via + ``raise_for_ownership`` in ``BaseRestoreVersionCommand.validate`` + and turns failures into the resource-specific exception triplet + packed in *spec*. + """ + try: + entity_uuid = UUID(uuid_str) + except ValueError: + return api.response_400(message="Invalid UUID") + try: + version_uuid = UUID(version_uuid_str) + except ValueError: + return api.response_400(message="Invalid version UUID") + + try: + spec.command_cls(entity_uuid, version_uuid).run() + except spec.not_found_exc: + return api.response_404() + except spec.forbidden_exc: + return api.response_403() + except spec.update_failed_exc as ex: + logger.error("Error restoring %s version: %s", spec.resource_label, ex) + return api.response_422(message=str(ex)) + return set_version_etag_by_uuid( + api.response(200, message="OK"), model_cls, entity_uuid + ) diff --git a/superset/versioning/baseline/__init__.py b/superset/versioning/baseline/__init__.py new file mode 100644 index 000000000000..664af88c4cb3 --- /dev/null +++ b/superset/versioning/baseline/__init__.py @@ -0,0 +1,69 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""``before_flush`` listener that captures a baseline version (version 0) +for entities being updated for the first time after the versioning +migration. + +Package layout (descends from public entry point to leaf builders): + +* :mod:`.listener` — public :func:`register_baseline_listener` that + wires the before-flush event on ``db.session``. +* :mod:`.dirty` — :func:`force_parent_dirty_on_child_change` and + :func:`pin_audit_columns`: promote a parent into ``session.dirty`` + when only its versioned children changed, and pin its audit columns + so the synthetic flush doesn't bump them. +* :mod:`.collection` — discovery: which parents need a baseline row? + Holds ``VERSIONED_MODELS`` (populated at app start), + :func:`collect_parents_to_baseline`, the + :func:`child_to_parent_registry` mapping, and the per-parent + Continuum-shadow-table lookups. +* :mod:`.insertion` — parent baseline insertion + child-handler + dispatch. +* :mod:`.children` — per-entity child baseline handlers + (``_baseline_dataset_children`` / ``_baseline_dashboard_children``) + plus the leaf helpers that synthesize child / slice shadow rows. +* :mod:`.shadow` — low-level :func:`insert_baseline_shadow_row` + helper used by every module that writes a shadow row, and the + :data:`CONTINUUM_BOOKKEEPING_COLUMNS` constant re-used outside this + package (the change-record listener and ``queries.py`` filter on it). + +The re-exports below preserve the prior ``from +superset.versioning.baseline import …`` call shape; no caller outside +this package needs to change. +""" + +from __future__ import annotations + +from superset.versioning.baseline.collection import ( + child_to_parent_registry, + VERSIONED_MODELS, +) +from superset.versioning.baseline.dirty import pin_audit_columns +from superset.versioning.baseline.listener import register_baseline_listener +from superset.versioning.baseline.shadow import ( + CONTINUUM_BOOKKEEPING_COLUMNS, + insert_baseline_shadow_row, +) + +__all__ = [ + "CONTINUUM_BOOKKEEPING_COLUMNS", + "VERSIONED_MODELS", + "child_to_parent_registry", + "insert_baseline_shadow_row", + "pin_audit_columns", + "register_baseline_listener", +] diff --git a/superset/versioning/baseline/children.py b/superset/versioning/baseline/children.py new file mode 100644 index 000000000000..e55a5cf4fb57 --- /dev/null +++ b/superset/versioning/baseline/children.py @@ -0,0 +1,212 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Per-entity child-baseline handlers. + +After a parent baseline row lands in :mod:`.insertion`, this module's +handlers write the parent's child baselines under the same transaction +id. The dispatch table :data:`CHILD_BASELINE_HANDLERS` is keyed on +the parent class name (avoids an import-cycle with the entity modules, +which can't be loaded at app-init time). + +The dataset handler baselines :class:`TableColumn` and +:class:`SqlMetric` children. The dashboard handler baselines the +``dashboard_slices`` M2M membership *and* synthesizes +``operation_type=0`` rows in ``slices_version`` for attached slices +that have no prior shadow — without those slice-side baselines, +Continuum's M2M revert query returns empty. + +Leaf-level helpers (:func:`_insert_child_baseline_rows`, +:func:`_baseline_attached_slices`, +:func:`_insert_synthetic_slice_baseline`) live here too — they're +shared between the two parent-specific handlers. +""" + +from __future__ import annotations + +from collections.abc import Callable +from typing import Any + +import sqlalchemy as sa +from sqlalchemy.orm import Session + +from superset.versioning.baseline.shadow import insert_baseline_shadow_row + + +def _baseline_dataset_children(session: Session, dataset: Any, tx_id: int) -> None: + """Baseline a dataset's ``TableColumn`` and ``SqlMetric`` children + under the dataset's baseline tx. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.connectors.sqla.models import SqlMetric, TableColumn + + for child_cls in (TableColumn, SqlMetric): + _insert_child_baseline_rows( + session, + dataset, + child_cls.__table__, + version_class(child_cls).__table__, + "table_id", + tx_id, + ) + + +def _baseline_dashboard_children(session: Session, dashboard: Any, tx_id: int) -> None: + """Baseline a dashboard's ``dashboard_slices`` M2M plus synthesize + ``operation_type=0`` rows in ``slices_version`` for attached slices + with no prior shadow. + + Continuum's M2M version-side relationship for ``Dashboard.slices`` + joins through both ``dashboard_slices_version`` AND + ``slices_version``: the second exists clause filters slices by + "latest slices_version row with tx <= dashboard.tx". If a slice + has no slices_version rows at all, that join produces no match + and ``version_obj.slices`` returns empty — leaving the dashboard + restore with no slices to append. The synthetic slice baseline at + this dashboard's tx gives the M2M query a slice version it can match. + + Doesn't try to be clever about slices shared across dashboards: a + slice is baselined at this dashboard's tx_id only when it has no + shadow rows at all. If a later dashboard baseline references the + same slice, this baseline (now at lower tx) is still found by + that dashboard's restore. The reverse — a dashboard baselined + AFTER the slice was first baselined under another dashboard at + a higher tx — is a residual gap deferred to a future fix. + """ + metadata = type(dashboard).__table__.metadata + live_tbl = metadata.tables.get("dashboard_slices") + shadow_tbl = metadata.tables.get("dashboard_slices_version") + if live_tbl is None or shadow_tbl is None: + return + + _insert_child_baseline_rows( + session, dashboard, live_tbl, shadow_tbl, "dashboard_id", tx_id + ) + _baseline_attached_slices(session, dashboard, live_tbl, tx_id) + + +# Dispatch table keyed by parent CLASS NAME rather than class, to avoid +# the import-cycle between baseline.py (loaded at app init) and the +# entity modules. The class-name string is set once at app start by +# the model definitions — typo-prone if extended. Declared after the +# handlers it references because module-level dict literals evaluate +# at import time and need the names already bound. +_ChildBaselineHandler = Callable[[Session, Any, int], None] +CHILD_BASELINE_HANDLERS: dict[str, _ChildBaselineHandler] = { + "SqlaTable": _baseline_dataset_children, + "Dashboard": _baseline_dashboard_children, +} + + +def _insert_child_baseline_rows( + session: Session, + parent_obj: Any, + child_table: sa.Table, + child_version_table: sa.Table, + fk_column_name: str, + tx_id: int, +) -> None: + """Synthesize ``operation_type=0`` shadow rows for every live child of + *parent_obj* under transaction id *tx_id*. + + Parallels :func:`~superset.versioning.baseline.insertion._insert_baseline_row` + but iterates over child rows. Used to give Continuum's ``Reverter`` + baseline data for children of pre-existing parents (children that + predate this commit have no shadow rows otherwise, so Reverter + would treat them as "deleted at the target tx" and try to remove + them on revert — the ADR-004 Failure 1 reproduction scenario). + + :param child_table: the live child SQLAlchemy ``Table`` (e.g. + ``TableColumn.__table__`` or the bare ``dashboard_slices`` association) + :param child_version_table: the corresponding Continuum shadow ``Table`` + :param fk_column_name: column on *child_table* that points to the parent + (e.g. ``"table_id"`` for ``TableColumn``, ``"dashboard_id"`` for + ``dashboard_slices``) + """ + conn = session.connection() + fk_col = getattr(child_table.c, fk_column_name) + + rows = ( + conn.execute(sa.select(child_table).where(fk_col == parent_obj.id)) + .mappings() + .all() + ) + if not rows: + return + + for row in rows: + insert_baseline_shadow_row(conn, child_version_table, row, tx_id) + + +def _baseline_attached_slices( + session: Session, dashboard: Any, live_tbl: sa.Table, tx_id: int +) -> None: + """Insert ``operation_type=0`` rows in ``slices_version`` for each + slice attached to *dashboard* that has no shadow row yet. + + Batched: one membership SELECT, one existing-shadow SELECT, one live + SELECT for the missing slices. Per-slice work happens only on + ``_insert_synthetic_slice_baseline``. The previous per-slice + ``COUNT(*)`` + ``SELECT`` pattern was O(N) round-trips and surfaced + as a measurable first-save hotspot on dashboards with many charts. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.slice import Slice + + slice_ver_table = version_class(Slice).__table__ + slice_table = Slice.__table__ + conn = session.connection() + + attached_slice_ids = [ + r.slice_id + for r in conn.execute( + sa.select(live_tbl.c.slice_id).where( + live_tbl.c.dashboard_id == dashboard.id + ) + ).all() + ] + if not attached_slice_ids: + return + + existing_shadow_ids = { + row[0] + for row in conn.execute( + sa.select(slice_ver_table.c.id.distinct()).where( + slice_ver_table.c.id.in_(attached_slice_ids) + ) + ).all() + } + missing_ids = [sid for sid in attached_slice_ids if sid not in existing_shadow_ids] + if not missing_ids: + return + + slice_rows = ( + conn.execute(sa.select(slice_table).where(slice_table.c.id.in_(missing_ids))) + .mappings() + .all() + ) + for slice_row in slice_rows: + _insert_synthetic_slice_baseline(conn, slice_ver_table, slice_row, tx_id) + + +def _insert_synthetic_slice_baseline( + conn: Any, slice_ver_table: sa.Table, slice_row: Any, tx_id: int +) -> None: + insert_baseline_shadow_row(conn, slice_ver_table, slice_row, tx_id) diff --git a/superset/versioning/baseline/collection.py b/superset/versioning/baseline/collection.py new file mode 100644 index 000000000000..0e5ef73643b2 --- /dev/null +++ b/superset/versioning/baseline/collection.py @@ -0,0 +1,147 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Discovery: figure out which parents need a baseline row. + +Three helpers cooperate on the listener's "should I baseline" decision: + +* :func:`collect_parents_to_baseline` — walks ``session.dirty`` / + ``new`` / ``deleted`` and returns the unique parent entities to + consider (directly-dirty versioned parents + parents reachable from + dirty children via :func:`child_to_parent_registry`). +* :func:`version_table_for` — resolves a Continuum shadow Table for + one parent object. +* :func:`shadow_row_count` — counts existing shadow rows for the + parent's id; ``0`` is the signal to insert a baseline. + +:func:`child_to_parent_registry` is also exposed because +:mod:`superset.versioning.factory` consumes it via inline import. + +**Inline imports.** ``versioning.baseline`` is imported during +``init_versioning()`` before all SQLAlchemy mappers are configured; +the lazy imports defer Continuum + model resolution until call time. +""" + +from __future__ import annotations + +import functools +import logging +from typing import Any + +import sqlalchemy as sa +from sqlalchemy.exc import OperationalError +from sqlalchemy.orm import Session + +# Populated at app startup (superset/initialization/__init__.py) before +# register_baseline_listener() is called. +VERSIONED_MODELS: list[type] = [] + +logger = logging.getLogger(__name__) + + +def collect_parents_to_baseline(session: Session) -> dict[int, Any]: + """Return parents-to-baseline as ``{id(obj): obj}`` keyed by Python + object identity to dedupe across ``session.dirty + new + deleted``. + + Includes both directly-dirty versioned parents and parents reachable + from dirty/new/deleted children via the child→parent registry. + """ + parents: dict[int, Any] = {} + child_map = child_to_parent_registry() + for obj in list(session.dirty) + list(session.new) + list(session.deleted): + if type(obj) in VERSIONED_MODELS: + parents[id(obj)] = obj + continue + entry = child_map.get(type(obj)) + if entry is None: + continue + parent_attr, parent_cls = entry + parent = getattr(obj, parent_attr, None) + if parent is not None and type(parent) is parent_cls: # noqa: E721 + parents[id(parent)] = parent + return parents + + +@functools.cache +def child_to_parent_registry() -> dict[type, tuple[str, type]]: + """Map child entity class → (parent-relationship-attr, parent class). + + When a dirty child of a known type appears in session.dirty/new/deleted, + we walk to its parent and baseline the parent (+ siblings) under the + SAME flush so pre-edit child values land in the baseline shadow rows. + Without this, edits that only touch child rows produce a "silent" flush + A (just ``TableColumn``) followed by flush B (``SqlaTable.changed_on``); + flush B reads children from DB AFTER flush A already pushed UPDATEs, + capturing post-edit state. + + Cached because this is called from ``force_parent_dirty_on_child_change`` + and ``collect_parents_to_baseline`` on every save flush. The returned + mapping depends only on the (fixed at import time) child model classes, + so an unbounded ``functools.cache`` is the right shape — no invalidation + needed. + """ + # Lazy import: ``baseline`` is imported during ``init_versioning``, which + # runs before all model mappers are configured. Importing model classes + # at module load would either cycle or hit unresolved mappers. + # pylint: disable=import-outside-toplevel + from superset.connectors.sqla.models import SqlaTable, SqlMetric, TableColumn + + return { + TableColumn: ("table", SqlaTable), + SqlMetric: ("table", SqlaTable), + } + + +def version_table_for(obj: Any) -> Any: + """Return Continuum's shadow ``Table`` for *obj*'s class, or ``None`` + when the class isn't registered (forks / plugins that subclass without + ``__versioned__``). + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + from sqlalchemy_continuum.exc import ClassNotVersioned + + try: + return version_class(type(obj)).__table__ + except ClassNotVersioned: + return None + + +def shadow_row_count(session: Session, obj: Any, version_table: Any) -> int | None: + """Return number of shadow rows for *obj.id* in *version_table*, or + ``None`` when the version table is missing (migration not yet applied) + or the count query raised unexpectedly. + """ + try: + with session.no_autoflush: + return ( + session.connection() + .execute( + sa.select(sa.func.count()) + .select_from(version_table) + .where(version_table.c.id == obj.id) + ) + .scalar() + ) + except OperationalError: + return None + except Exception: # pylint: disable=broad-except + logger.exception( + "baseline_listener: count query failed for %s id=%s", + type(obj).__name__, + getattr(obj, "id", None), + ) + return None diff --git a/superset/versioning/baseline/dirty.py b/superset/versioning/baseline/dirty.py new file mode 100644 index 000000000000..f351cbb38bbe --- /dev/null +++ b/superset/versioning/baseline/dirty.py @@ -0,0 +1,243 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Parent-dirty force machinery for child-only saves. + +When a versioned child (``TableColumn`` / ``SqlMetric``) is in +``session.dirty`` / ``new`` / ``deleted`` but its parent's scalar +columns haven't been touched, the parent is *missing* from the dirty +set — so Continuum's UnitOfWork never creates a parent UPDATE +operation, no parent shadow row is written, and the version-history +dropdown comes back empty for column/metric-only saves. + +:func:`force_parent_dirty_on_child_change` walks dirty/new/deleted +children, looks them up in the child→parent registry (in +:mod:`.collection`), and ``attributes.flag_modified``s a deterministic +non-excluded column on the parent. SQLAlchemy adds the parent to +``session.dirty``; Continuum then writes a parent shadow row whose +scalars mirror the previous version (only the children actually +changed). + +:func:`pin_audit_columns` is a companion: when the parent is force- +flagged, we pin ``changed_by_fk`` / ``changed_on`` to their current +in-memory values so the parent UPDATE doesn't invoke the audit +columns' ``onupdate=get_user_id`` / ``onupdate=datetime.now`` hooks +(which would attribute the synthetic flush to whoever ``g.user`` is +at the time, possibly a deleted test user under autoflush teardown). + +**Inline imports.** Same init-order rationale as +:mod:`superset.versioning.baseline.collection`. +""" + +from __future__ import annotations + +import logging +from collections.abc import Iterator +from typing import Any + +from sqlalchemy.exc import InvalidRequestError +from sqlalchemy.orm import attributes, Session + +from superset.versioning.baseline.collection import child_to_parent_registry + +logger = logging.getLogger(__name__) + + +def force_parent_dirty_on_child_change(session: Session) -> None: + """Mark a versioned parent as dirty whenever one of its versioned + children appears in ``session.dirty``/``new``/``deleted`` but the + parent's own scalars haven't been edited. + + Without this hook, edits that only touch ``TableColumn`` or + ``SqlMetric`` rows leave the parent ``SqlaTable`` out of + ``session.dirty`` — so Continuum's UnitOfWork never creates a + parent UPDATE operation and ``list_versions`` (which queries the + parent shadow ``tables_version``) returns just the baseline. The + user-visible symptom is "I edited a column description but the + dataset's version history dropdown is empty". + + For each child that represents a real edit, we resolve its parent + and ``attributes.flag_modified`` a deterministic non-excluded + column so SQLAlchemy adds the parent to ``session.dirty`` without + altering any column values. Continuum then writes a parent shadow + row at this transaction; its scalar columns mirror the previous + version (only the children changed). + ``SkipUnmodifiedPlugin._is_no_op_update`` is taught to recognize + the "scalars match but children dirty" case and keep the row. + """ + child_map = child_to_parent_registry() + new_set = session.new + for child in _real_dirty_versioned_children(session, child_map): + parent = _resolve_parent(child, child_map) + if parent is None: + continue + if parent in new_set: + # Already-new short-circuit. If the parent itself is in + # ``session.new`` (typical during an import that adds a + # ``SqlaTable`` plus 50 fresh ``TableColumn`` children), it + # will INSERT in this flush regardless — the + # ``flag_modified`` call is redundant (and the attribute- + # default-not-yet-fired case in ``_flag_parent`` would just + # swallow an ``InvalidRequestError``). Skip the work. + continue + if _flag_parent(parent): + pin_audit_columns(parent) + + +def _real_dirty_versioned_children( + session: Session, child_map: dict[type, Any] +) -> Iterator[Any]: + """Yield child instances that are (a) of a versioned-child class + registered in *child_map*, and (b) represent a real content edit — + not a phantom-dirty entry from lazy-load side effects or audit- + column auto-bumps. + + Phantom-dirty filter rationale: a child can appear in + ``session.dirty`` for reasons that don't represent real content + edits — lazy-load side effects, ``AuditMixin`` auto-bumps from + prior code paths, M2M relationship-cascade artifacts (e.g., + ``rls_entry.tables.extend([dataset])`` in setUp), Reverter side + passes. Force-touching the parent in those cases produces an + incidental ``UPDATE tables SET description=…, changed_on=…, + changed_by_fk=…`` that can violate FK integrity on some dialects + (observed in ``test_rls_filter_alters_no_role_user_birth_names_query``). + + The filter applies ONLY to persistent rows in ``session.dirty``: + ``session.new`` (creation) and ``session.deleted`` (removal) are + always real content changes — deletion in particular is a state + transition with no attribute history, so ``is_modified`` returns + False there even when the change is real (column-removed records + must still emit). + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import is_modified + + # ``session.dirty`` / ``session.new`` are IdentitySets — ``__contains__`` + # uses identity comparison, which is what we need for the phantom- + # dirty filter below. + dirty_set = session.dirty + for obj in list(session.dirty) + list(session.new) + list(session.deleted): + if type(obj) not in child_map: + continue + if obj in dirty_set and not is_modified(obj): + continue + yield obj + + +def _resolve_parent(child: Any, child_map: dict[type, Any]) -> Any | None: + """Resolve the versioned parent for *child* via the child→parent + registry; return ``None`` when the registered parent attribute + isn't loaded or has been swapped for an unexpected type.""" + parent_attr, parent_cls = child_map[type(child)] + parent = getattr(child, parent_attr, None) + if parent is None or type(parent) is not parent_cls: # noqa: E721 + return None + return parent + + +def _flag_parent(parent: Any) -> bool: + """``flag_modified`` a stable non-excluded column on *parent* so + SQLAlchemy adds it to ``session.dirty`` without altering values. + Returns ``True`` on success. + + Column choice: ``description`` is a plain ``Text`` column on all + three versioned parent classes (Dashboard, Slice, SqlaTable) and is + in none of their ``__versioned__`` excludes — pick it + deterministically so the flagged attribute is stable across + SQLAlchemy versions / mapper-configuration orders. We deliberately + avoid ``uuid``: when a versioned-parent UPDATE goes through with + ``uuid`` flagged, the column's ``UUIDType``/BLOB round-trip + produces a memoryview that fails an FK integrity check on some + dialects (observed in + ``test_rls_filter_alters_no_role_user_birth_names_query`` and + ``test_restore_applies_scalar_field``). ``description`` is a plain + text column with no marshaling layer, so flagging it safely + round-trips its current value. Falls back to ``uuid`` then + ``col_keys[0]`` for forks that excluded ``description``. + + Returns ``False`` for the freshly-constructed ``session.new`` + instance whose attribute defaults haven't fired yet — the + attribute is unloaded in instance state, so ``flag_modified`` + rejects it with ``InvalidRequestError``. The parent will INSERT in + this flush regardless, so the flag was redundant; safely skip. + Hit by ``test_create_dataset_item`` (POST /api/v1/dataset/). + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum.utils import versioned_column_properties + + col_keys = [prop.key for prop in versioned_column_properties(parent)] + if not col_keys: + return False + if "description" in col_keys: + flag_col = "description" + elif "uuid" in col_keys: + flag_col = "uuid" + else: + flag_col = col_keys[0] + try: + attributes.flag_modified(parent, flag_col) + except InvalidRequestError: + return False + return True + + +def pin_audit_columns(parent: Any) -> None: + """Pin ``changed_by_fk`` and ``changed_on`` to their current in-memory + values on a flag-flushed parent. + + ``changed_by_fk`` carries ``onupdate=get_user_id`` from ``AuditMixin``: + any UPDATE statement that doesn't explicitly set this column lets + SQLAlchemy invoke ``get_user_id()`` and write whoever ``g.user`` is + at flush time. When the flush is autoflush-triggered during an + earlier test's teardown (after the test user has been deleted from + ``ab_user``), the bumped value points at a non-existent row and the + parent UPDATE fails the FK to ``ab_user``. The same applies to + ``changed_on``'s ``onupdate=datetime.now`` (cosmetic only, but it's + cheap to pin together). + + ``flag_modified`` on both columns marks them as having dirty + attribute history, which tells SQLAlchemy to use the in-memory + (previously-committed) values instead of invoking ``onupdate`` — + the parent UPDATE then carries the existing audit values rather + than whatever ``g.user`` resolves to during the synthetic flag + flush. Hits ``test_rls_filter_alters_no_role_user_birth_names_query`` + and ``TestDatasetRestoreApi::test_restore_applies_scalar_field`` + in CI's full-suite ordering (autoflush during teardown). + """ + pinned_any = False + for audit_col in ("changed_by_fk", "changed_on"): + if hasattr(parent, audit_col): + try: + attributes.flag_modified(parent, audit_col) + pinned_any = True + except InvalidRequestError: + continue + if not pinned_any and hasattr(parent, "changed_by_fk"): + # Both audit columns are present on the parent but neither + # ``flag_modified`` succeeded — typically because the parent is + # a freshly-constructed ``session.new`` instance whose attribute + # defaults haven't fired yet. Without the pin, the synthetic + # parent UPDATE in this flush invokes ``onupdate=get_user_id`` + # and writes whoever ``g.user`` is at flush time, which under + # autoflush-during-teardown can point at a deleted test user + # and fail the FK to ``ab_user``. Surface this so the failure + # mode is debuggable from the log without inspection. + logger.info( + "baseline: skipped audit-column pin on %s id=%s " + "(attribute defaults not loaded)", + type(parent).__name__, + getattr(parent, "id", None), + ) diff --git a/superset/versioning/baseline/insertion.py b/superset/versioning/baseline/insertion.py new file mode 100644 index 000000000000..303a1e4f202c --- /dev/null +++ b/superset/versioning/baseline/insertion.py @@ -0,0 +1,149 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Parent baseline insertion + child-handler dispatch. + +Two complementary helpers: + +* :func:`insert_baseline_and_children` — top-level glue called by + the listener. Wraps the work in ``session.no_autoflush`` (so + ``session.connection()`` doesn't trigger a flush of Continuum's + pending Transaction object before our direct-SQL insert claims its + tx_id) and logs any failures as listener-boundary errors. +* :func:`_insert_baseline_row` — actually writes the + ``version_transaction`` row and the parent shadow row. Returns the + allocated ``transaction_id``. +* :func:`_baseline_children_for_parent` — dispatches to the per- + entity handler in :mod:`.children` under the same tx_id. +""" + +from __future__ import annotations + +import logging +from typing import Any + +import sqlalchemy as sa +from sqlalchemy.orm import Session + +from superset.versioning.baseline.children import CHILD_BASELINE_HANDLERS +from superset.versioning.baseline.shadow import insert_baseline_shadow_row +from superset.versioning.utils import read_row_outside_flush + +logger = logging.getLogger(__name__) + + +def insert_baseline_and_children( + session: Session, obj: Any, version_table: Any +) -> None: + """Insert the parent baseline row, then baseline the parent's child + collections under the same transaction id. + + Wrapped in ``no_autoflush`` so ``session.connection()`` inside + ``_insert_baseline_row`` does not trigger a flush of Continuum's + pending Transaction object before our direct-SQL insert claims its + tx_id. + """ + try: + with session.no_autoflush: + tx_id = _insert_baseline_row(session, obj, version_table) + if tx_id is None: + return + _baseline_children_for_parent(session, obj, tx_id) + logger.debug( + "baseline_listener: inserted baseline tx_id=%s for %s id=%s", + tx_id, + type(obj).__name__, + getattr(obj, "id", None), + ) + except Exception: # pylint: disable=broad-except + logger.exception( + "baseline_listener: failed to insert baseline for %s id=%s", + type(obj).__name__, + getattr(obj, "id", None), + ) + + +def _insert_baseline_row( + session: Session, obj: Any, version_table: sa.Table +) -> int | None: + """Insert a synthetic baseline row capturing the pre-edit DB state of *obj*. + + Creates a version_transaction entry and an operation_type=0 version row. + All writes use the session's existing connection so they share the same + database transaction as the triggering flush. + + Returns the allocated ``transaction_id`` so the caller can baseline child + collections under the same tx (see + :func:`~superset.versioning.baseline.children._insert_child_baseline_rows`), + or ``None`` when the entity has no live row. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + main_table = type(obj).__table__ + row = read_row_outside_flush(session, main_table, obj.id) + if row is None: + return None + + conn = session.connection() + + # Insert a version_transaction row for the baseline. + # + # ``issued_at`` and ``user_id`` are sourced from the entity's audit fields + # (``changed_on`` / ``changed_by_fk``, falling back to ``created_on`` / + # ``created_by_fk`` if the row was never edited), so the baseline reads + # in the version-history UI as "this is the state at the time of the + # last pre-versioning edit, by that user." Using ``now()`` and the + # current user would have made the baseline look chronologically newer + # than subsequent edits and attributed historical content to the user + # who happened to trigger the first save under versioning. + baseline_issued_at = row.get("changed_on") or row.get("created_on") or sa.func.now() + baseline_user_id = row.get("changed_by_fk") or row.get("created_by_fk") + tx_table = versioning_manager.transaction_cls.__table__ + result = conn.execute( + tx_table.insert().values( + issued_at=baseline_issued_at, + user_id=baseline_user_id, + remote_addr=None, + ) + ) + tx_id = result.inserted_primary_key[0] + insert_baseline_shadow_row(conn, version_table, row, tx_id) + return tx_id + + +def _baseline_children_for_parent( + session: Session, parent_obj: Any, tx_id: int +) -> None: + """Baseline a parent's child collections under the parent's baseline tx. + + Dispatches via the + :data:`~superset.versioning.baseline.children.CHILD_BASELINE_HANDLERS` + table to per-entity handlers. A handler failure is logged but does + not block the parent baseline. + """ + parent_name = type(parent_obj).__name__ + handler = CHILD_BASELINE_HANDLERS.get(parent_name) + if handler is None: + return + try: + handler(session, parent_obj, tx_id) + except Exception: # pylint: disable=broad-except + logger.exception( + "baseline_listener: failed to baseline children of %s id=%s", + parent_name, + getattr(parent_obj, "id", None), + ) diff --git a/superset/versioning/baseline/listener.py b/superset/versioning/baseline/listener.py new file mode 100644 index 000000000000..0b16f7365c6a --- /dev/null +++ b/superset/versioning/baseline/listener.py @@ -0,0 +1,95 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Public entry point: attach the ``before_flush`` baseline listener. + +:func:`register_baseline_listener` is called from +:class:`superset.initialization.SupersetAppInitializer.init_versioning` +after ``make_versioned()`` has run and all versioned model classes +have been imported. It registers one ``before_flush`` listener on +``db.session`` that: + +1. force-dirties versioned parents whose only changes are + child-collection edits (:mod:`.dirty`); +2. collects the parents that need a baseline row + (:mod:`.collection`); +3. for each parent with no prior shadow row, inserts the synthetic + baseline row + its child baseline rows (:mod:`.insertion` + + :mod:`.children`). +""" + +from __future__ import annotations + +from typing import Any + +from sqlalchemy import event +from sqlalchemy.orm import Session + +from superset.versioning.baseline.collection import ( + collect_parents_to_baseline, + shadow_row_count, + version_table_for, + VERSIONED_MODELS, +) +from superset.versioning.baseline.dirty import force_parent_dirty_on_child_change +from superset.versioning.baseline.insertion import insert_baseline_and_children + +# Sentinel attribute set on the session target after first successful +# registration — same pattern as +# :mod:`superset.versioning.changes.listener`. Subsequent calls become +# no-ops so test fixtures that instantiate multiple Superset apps per +# process don't attach a second copy of the listener to the shared +# ``db.session`` (every flush would otherwise run the baseline pass +# twice). +_REGISTERED_SENTINEL = "_versioning_baseline_listener_registered" + + +def register_baseline_listener() -> None: + """Attach the before_flush listener that captures baseline versions. + + Call this after ``VERSIONED_MODELS`` has been populated and + ``make_versioned()`` has run. Idempotent — repeat calls are no-ops. + """ + # pylint: disable=import-outside-toplevel + from superset.extensions import db + + if getattr(db.session, _REGISTERED_SENTINEL, False): + return + + # insert=True prepends us in the listener chain so we run BEFORE + # Continuum's before_flush. Continuum's pending Transaction object + # (added in its own before_flush) would otherwise get a lower + # auto-increment tx_id than our direct-SQL baseline insert, placing the + # baseline row after the update in version_number order. Prepending + # ensures our baseline's tx_id comes first. + @event.listens_for(db.session, "before_flush", insert=True) + def capture_baseline(session: Session, flush_context: Any, instances: Any) -> None: + if not VERSIONED_MODELS: + return + # Make sure a child-only edit promotes the parent to ``session.dirty`` + # before Continuum's before_flush reads the dirty set. + force_parent_dirty_on_child_change(session) + for obj in collect_parents_to_baseline(session).values(): + if type(obj) not in VERSIONED_MODELS: + continue + version_table = version_table_for(obj) + if version_table is None: + continue + count = shadow_row_count(session, obj, version_table) + if count == 0: + insert_baseline_and_children(session, obj, version_table) + + setattr(db.session, _REGISTERED_SENTINEL, True) diff --git a/superset/versioning/baseline/shadow.py b/superset/versioning/baseline/shadow.py new file mode 100644 index 000000000000..c31a5835bcf7 --- /dev/null +++ b/superset/versioning/baseline/shadow.py @@ -0,0 +1,72 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Continuum-shaped shadow-row writer. + +Two pieces: + +* :data:`CONTINUUM_BOOKKEEPING_COLUMNS` — the set of column names + Continuum uses for per-row bookkeeping (``transaction_id`` / + ``end_transaction_id`` / ``operation_type``). Re-used outside this + package as a filter (the change-record listener strips these from + JSON record values). +* :func:`insert_baseline_shadow_row` — copies a live row into a + shadow ``Table`` as a synthetic ``operation_type=0`` baseline at + the given transaction id. The other modules in this package use it + for every parent and child baseline insert. +""" + +from __future__ import annotations + +from typing import Any + +import sqlalchemy as sa + +# Continuum's per-shadow-row bookkeeping columns. Skipped when copying +# content from a live row into a synthetic baseline shadow row; set +# explicitly by the baseline writer so the row reads as a freshly-created +# live row at the baseline transaction. +CONTINUUM_BOOKKEEPING_COLUMNS: frozenset[str] = frozenset( + {"transaction_id", "end_transaction_id", "operation_type"} +) + + +def insert_baseline_shadow_row( + conn: Any, + version_table: sa.Table, + source_row: Any, + tx_id: int, +) -> None: + """Copy *source_row* into *version_table* as a synthetic baseline + (``operation_type=0``) shadow row at *tx_id*. + + Content columns are copied through; the three Continuum bookkeeping + columns are set explicitly so the row reads as a freshly-created + live row at *tx_id*. Column objects (not names) are used as + ``values()`` keys to avoid the "Unconsumed column names" error that + a name-based dict hits when a Column's ``.key`` differs from its + ``.name`` — a thing Continuum-generated tables occasionally produce. + """ + col_values: dict[Any, Any] = {} + for col in version_table.columns: + if col.name in CONTINUUM_BOOKKEEPING_COLUMNS: + continue + if col.name in source_row: + col_values[col] = source_row[col.name] + col_values[version_table.c.transaction_id] = tx_id + col_values[version_table.c.end_transaction_id] = None + col_values[version_table.c.operation_type] = 0 + conn.execute(version_table.insert().values(col_values)) diff --git a/superset/versioning/changes/__init__.py b/superset/versioning/changes/__init__.py new file mode 100644 index 000000000000..736e7e458e58 --- /dev/null +++ b/superset/versioning/changes/__init__.py @@ -0,0 +1,70 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Write-side change-record capture for ``version_changes``. + +The package is split into four submodules that descend from public +entry point to leaf helpers: + +* :mod:`.listener` — public ``register_change_record_listener`` plus + the session-event machinery (``before_flush`` / ``after_flush`` / + ``after_commit`` / ``after_rollback``) that drives the capture. + Holds ``ACTION_KIND_KEY``, the buffer-key constants, and the per-tx + ``action_kind`` stamper. +* :mod:`.state` — per-entity diff dispatch: pre-state read, + post-state serialisation, JSON-safety coercion (``jsonable``), + cached scalar-field discovery, and bulk-insert into the + ``version_changes`` table. +* :mod:`.shadow_queries` — shadow-table reads that drive child- + collection diffs (dataset columns/metrics, dashboard slice + membership). Includes the validity-strategy ``shadow_rows_valid_at`` + helper consumed externally by :mod:`superset.versioning.queries`. +* :mod:`.table` — the SQLAlchemy ``Table`` definition for + ``version_changes`` plus the ``ENTITY_KIND_BY_CLASS_NAME`` mapping + consumed by the API + activity-view modules. + +The re-exports below preserve the prior ``from +superset.versioning.changes import …`` call shape; no caller outside +this package needs to change. +""" + +from __future__ import annotations + +from superset.versioning.changes.listener import ( + ACTION_KIND_CLONE, + ACTION_KIND_IMPORT, + ACTION_KIND_KEY, + ACTION_KIND_RESTORE, + ACTION_KINDS, + register_change_record_listener, +) +from superset.versioning.changes.shadow_queries import shadow_rows_valid_at +from superset.versioning.changes.table import ( + ENTITY_KIND_BY_CLASS_NAME, + version_changes_table, +) + +__all__ = [ + "ACTION_KIND_CLONE", + "ACTION_KIND_IMPORT", + "ACTION_KIND_KEY", + "ACTION_KIND_RESTORE", + "ACTION_KINDS", + "ENTITY_KIND_BY_CLASS_NAME", + "register_change_record_listener", + "shadow_rows_valid_at", + "version_changes_table", +] diff --git a/superset/versioning/changes/listener.py b/superset/versioning/changes/listener.py new file mode 100644 index 000000000000..48179ad80859 --- /dev/null +++ b/superset/versioning/changes/listener.py @@ -0,0 +1,421 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Session-level listeners that drive ``version_changes`` writes. + +Two flush events cooperate, plus two post-commit / post-rollback +cleanups: + +- ``before_flush``: for each versioned entity in ``session.dirty``, + reads the pre-save scalar state from the DB via raw SQL inside + ``session.no_autoflush`` (same idiom as the baseline listener, not + Continuum's internal ``units_of_work`` which is a private API), reads + the post-save state from the in-memory ORM object, calls the diff + engine, and buffers the resulting :class:`ChangeRecord` list on + ``session.info``. This must run before the flush because after the + flush the DB already reflects the post-state; we can't recover the + pre-state from it. + +- ``after_flush``: drains the buffer, resolves the current Continuum + transaction id via ``versioning_manager.units_of_work``, and bulk- + inserts one ``version_changes`` row per record with a monotonic + ``sequence`` number. Records accumulated across multiple before_flush + calls within one transaction share the same ``transaction_id`` and + contiguous sequence numbers. + +- ``after_commit`` / ``after_rollback``: clean up session-scoped + state (processed-tx set, ``action_kind`` key) so a long-lived session + doesn't accumulate stale buffer entries. + +Scope: + - Slice, Dashboard, SqlaTable **scalar fields** (via the cached + field set from :mod:`superset.versioning.changes.state` — new + columns are picked up automatically). + - ``Slice.params`` kind-classification (filter / metric / time_range + / color_palette / dimension, plus generic ``field`` fallback). + +Child-collection diffs (dataset ``TableColumn`` / ``SqlMetric``, +dashboard ``dashboard_slices``) read the pre- and post-state from +Continuum shadow tables via the helpers in +:mod:`superset.versioning.changes.shadow_queries`, executed in +``after_flush`` once Continuum has written its tx-N rows. + +``session.new`` entities are not processed in this listener: +operation_type=0 transactions (baseline capture and first-save INSERTs) +produce zero change records per spec §Clarifications 2026-04-24. +""" + +from __future__ import annotations + +import logging +from typing import Any + +import sqlalchemy as sa +from sqlalchemy import event +from sqlalchemy.exc import OperationalError +from sqlalchemy.orm import Session + +from superset.versioning.changes.shadow_queries import ( + _dashboard_child_records_for_tx_from_shadows, + _dataset_child_records_for_tx_from_shadows, +) +from superset.versioning.changes.state import ( + bulk_insert_records, + compute_records_for_entity, +) +from superset.versioning.changes.table import ENTITY_KIND_BY_CLASS_NAME +from superset.versioning.diff import ( + ChangeRecord, + fold_dashboard_layout_with_chart_changes, +) + +logger = logging.getLogger(__name__) + + +# Key under which the pending-records buffer is stored on ``session.info``. +# Using ``session.info`` (SQLAlchemy's user-data dict) avoids the need +# for a module-level WeakKeyDictionary and keeps buffers naturally scoped +# to the session's lifetime. +_BUFFER_KEY = "_version_changes_pending" + +# Key for the set of Continuum transaction ids whose change records +# have already been written in this session. ``after_flush`` can fire +# more than once for a single transaction (e.g. autoflush triggered by +# a mid-commit query), and our child-diff path reads snapshot tables +# that don't care about the buffer state — without this marker we'd +# re-insert the same child records on the second flush and hit the +# UNIQUE(transaction_id, entity_kind, entity_id, sequence) constraint. +_PROCESSED_TXS_KEY = "_version_changes_processed_txs" + +# Key on ``session.info`` that commands set to declare the high-level +# action that produced the current transaction. Read once per flush by +# the change-record listener and stamped onto the +# ``version_transaction.action_kind`` column via ``sa.update()``. +# ``None`` (the default) means "ordinary save". +# +# Commands set this immediately before ``db.session.commit()``: +# +# db.session.info[ACTION_KIND_KEY] = ACTION_KIND_RESTORE +# db.session.commit() +# +# The listener pops the key after stamping, and ``after_commit`` / +# ``after_rollback`` cleanup pop it again as a safety net, so a +# long-lived session can't accidentally carry the value into the next +# transaction. +ACTION_KIND_KEY = "_versioning_action_kind" + +# Recognised ``action_kind`` values — the single source of truth shared +# by the four command-side stampers (restore / import / clone) and the +# listener that writes them to ``version_transaction.action_kind``. +# Schemas / response decorators that need an allowlist read from +# ``ACTION_KINDS`` so a future addition (e.g. ``"thumbnail_warm"``) only +# has to update this one constant. ``None`` is *not* a member — it +# represents the default "ordinary save" path that never sets the key. +ACTION_KIND_RESTORE = "restore" +ACTION_KIND_IMPORT = "import" +ACTION_KIND_CLONE = "clone" +ACTION_KINDS: frozenset[str] = frozenset( + {ACTION_KIND_RESTORE, ACTION_KIND_IMPORT, ACTION_KIND_CLONE} +) + +# Sentinel attribute set on the session target after first successful +# registration. Subsequent calls become no-ops. Storing the flag on the +# target itself (rather than module-level state) keeps the guard +# naturally scoped — a fresh session proxy gets a fresh registration — +# and avoids the TOCTOU race between ``event.contains`` and +# ``event.listen`` that a module-level ref would have under concurrent +# init. In test fixtures that instantiate multiple Superset apps per +# process, the shared ``db.session`` carries the sentinel and re-entry +# is correctly deduped. +_REGISTERED_SENTINEL = "_versioning_change_listener_registered" + +#: Metric namespace for swallowed capture-path failures. The capture +#: listeners fail open (a versioning bug must never break a user's save), +#: so the read path (``activity/orchestrator``) is richly instrumented but +#: the write path historically logged-and-swallowed with no counter. Each +#: ``_incr_capture_error(stage)`` emits ``..error`` so a +#: systematic capture regression is alertable rather than log-grep-only. +_CAPTURE_METRIC_PREFIX = "superset.versioning.capture" + + +def _incr_capture_error(stage: str) -> None: + """Emit a counter for a swallowed capture-path failure at *stage*. + + Best-effort: metrics emission must never itself break a user's save, + so it is wrapped in the same fail-open posture as the call site. + """ + # pylint: disable=import-outside-toplevel + try: + from superset.extensions import stats_logger_manager + + stats_logger_manager.instance.incr(f"{_CAPTURE_METRIC_PREFIX}.{stage}.error") + except Exception: # pylint: disable=broad-except + logger.exception("version_changes: failed to emit capture-error metric") + + +def _process_dirty_entity_into_buffer( + session: Session, + obj: Any, + buffer: dict[tuple[str, int], list[ChangeRecord]], +) -> None: + """Compute scalar change records for one dirty entity + append to buffer.""" + entity_kind = ENTITY_KIND_BY_CLASS_NAME.get(type(obj).__name__) + if entity_kind is None: + return + entity_id = getattr(obj, "id", None) + if entity_id is None: + return + try: + records = compute_records_for_entity(session, obj) + except Exception: # pylint: disable=broad-except + logger.exception( + "version_changes: diff failed for %s id=%s", + type(obj).__name__, + entity_id, + ) + _incr_capture_error("scalar_diff") + return + if records: + buffer.setdefault((entity_kind, entity_id), []).extend(records) + + +def _append_child_records_to_buffer( + session: Session, + tx_id: int, + buffer: dict[tuple[str, int], list[ChangeRecord]], +) -> None: + """Compute dataset + dashboard child-collection records + append to buffer. + + Runs in ``after_flush`` so the shadow tables already have the + current-tx rows. Reads from Continuum shadow tables + (``table_columns_version`` / ``sql_metrics_version`` / + ``dashboard_slices_version`` / ``slices_version``). + """ + try: + for dataset_id, records in _dataset_child_records_for_tx_from_shadows( + session, tx_id + ).items(): + buffer.setdefault(("dataset", dataset_id), []).extend(records) + for dashboard_id, records in ( + _dashboard_child_records_for_tx_from_shadows(session, tx_id) + ).items(): + buffer.setdefault(("dashboard", dashboard_id), []).extend(records) + + # Post-merge fold: when a dashboard save adds/removes charts, + # drop the redundant ``position_json.*`` records that mirror + # the membership change. See + # ``diff.fold_dashboard_layout_with_chart_changes``. + for key in list(buffer.keys()): + if key[0] == "dashboard": + buffer[key] = fold_dashboard_layout_with_chart_changes(buffer[key]) + if not buffer[key]: + del buffer[key] + except Exception: # pylint: disable=broad-except + logger.exception("version_changes: child-diff failed for tx %s", tx_id) + _incr_capture_error("child_diff") + + +def _current_transaction_id(session: Session) -> int | None: + """Return the Continuum transaction id for *session*'s current unit of + work, or ``None`` when Continuum has no active transaction (e.g. raw + SQL execution outside the ORM's flush flow). + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + uow = versioning_manager.units_of_work.get(session.connection()) + if uow is None or uow.current_transaction is None: + return None + return uow.current_transaction.id + + +def _stamp_action_kind_on_transaction(session: Session, tx_id: int) -> None: + """Pop the per-tx action_kind from ``session.info`` and stamp it + onto the ``version_transaction`` row identified by *tx_id*. + + No-op when no command set the action_kind (the default for + ordinary saves). Emits via ``sa.update()`` against Continuum's + transaction Table so the identifier is auto-quoted per dialect + (MySQL would otherwise reject the unquoted column name if it ever + collided with a reserved word) and the dialect-portable column + binding is reused instead of hand-written SQL. + + The action_kind is popped (not just read) so a long-lived session + can't accidentally carry the value into the next transaction. A + failed stamp is logged and swallowed — action_kind is a + descriptive enrichment, not a correctness invariant; refusing to + write change records because an UPDATE on a single column failed + would punish the user save for an audit-log nicety. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + action_kind = session.info.pop(ACTION_KIND_KEY, None) + if action_kind is None: + return + tx_tbl = versioning_manager.transaction_cls.__table__ + try: + session.connection().execute( + sa.update(tx_tbl) + .where(tx_tbl.c.id == tx_id) + .values(action_kind=action_kind) + ) + except Exception: # pylint: disable=broad-except + logger.exception( + "version_changes: failed to stamp action_kind=%s on tx %s", + action_kind, + tx_id, + ) + _incr_capture_error("action_kind_stamp") + + +def _persist_buffered_records( + session: Session, + tx_id: int, + buffer: dict[tuple[str, int], list[ChangeRecord]], +) -> None: + """Bulk-insert *buffer*'s records under *tx_id* and reset the buffer. + + Catches ``OperationalError`` to handle the pre-migration startup race + (version_changes table missing), and ``Exception`` as the listener- + boundary safety net so a malformed record can't crash the user's save. + """ + try: + bulk_insert_records(session, tx_id, buffer) + except OperationalError: + # version_changes table missing (migration not yet applied). + pass + except Exception: # pylint: disable=broad-except + logger.exception( + "version_changes: bulk insert failed for tx %s (%d entities)", + tx_id, + len(buffer), + ) + _incr_capture_error("bulk_insert") + + +def register_change_record_listener() -> None: # noqa: C901 + """Attach the before_flush + after_flush listeners. + + Registered from :class:`superset.initialization.SupersetAppInitializer` + (``init_versioning``) alongside the baseline, dataset-snapshot, + and dashboard-snapshot listeners. Must run after Continuum's + ``make_versioned()`` so the ``versioning_manager`` is available + and has installed its own before_flush hook. + """ + # pylint: disable=import-outside-toplevel + from superset.connectors.sqla.models import SqlaTable + from superset.extensions import db + from superset.models.dashboard import Dashboard + from superset.models.slice import Slice + + if getattr(db.session, _REGISTERED_SENTINEL, False): + return + + versioned_classes: tuple[type, ...] = (Dashboard, Slice, SqlaTable) + + def compute_change_records( + session: Session, _flush_context: Any, _instances: Any + ) -> None: + # session.info persists across before_flush/after_flush within + # a single transaction. The buffer is keyed on + # ``(entity_kind, entity_id)`` so scalar records captured here + # and child records captured in after_flush (T048b) merge + # under the same entity without duplication. + buffer: dict[tuple[str, int], list[ChangeRecord]] = session.info.setdefault( + _BUFFER_KEY, {} + ) + for obj in list(session.dirty): + if isinstance(obj, versioned_classes): + _process_dirty_entity_into_buffer(session, obj, buffer) + + def flush_change_records(session: Session, _flush_context: Any) -> None: + buffer: dict[tuple[str, int], list[ChangeRecord]] = session.info.setdefault( + _BUFFER_KEY, {} + ) + + tx_id = _current_transaction_id(session) + if tx_id is None: + session.info[_BUFFER_KEY] = {} + return + + # Skip if we've already written records for this tx (after_flush + # can fire more than once per commit — e.g. autoflush from a + # mid-commit query). Without this guard the child-diff path would + # re-read the same shadow rows and re-emit the same records, + # tripping the UNIQUE(transaction_id, entity_kind, entity_id, + # sequence) constraint on insert. + processed: set[int] = session.info.setdefault(_PROCESSED_TXS_KEY, set()) + if tx_id in processed: + return + + # Stamp action_kind eagerly, before the buffer-empty short- + # circuit. Restores / imports / clones may flush across multiple + # cycles; the FIRST firing for this tx is the one with the + # value still on ``session.info``. The helper pops on success + # so subsequent firings see ``None`` and short-circuit cleanly. + _stamp_action_kind_on_transaction(session, tx_id) + + _append_child_records_to_buffer(session, tx_id, buffer) + + if not buffer: + # Don't mark tx as processed when nothing was inserted. A + # later after_flush firing for the same tx may carry the + # records — e.g. when an entity's edit lands across two + # flushes (a child-only flush followed by a parent-dirty + # flush): the parent shadow only lands in the parent-dirty + # flush, so the child-diff path can't find a prior tx to + # compare against until then. + session.info[_BUFFER_KEY] = {} + return + + try: + _persist_buffered_records(session, tx_id, buffer) + finally: + session.info[_BUFFER_KEY] = {} + processed.add(tx_id) + + def reset_processed_after_commit(session: Session) -> None: + # ``_PROCESSED_TXS_KEY`` accumulates Continuum tx ids whose change + # records have already been written, to dedup against multiple + # ``after_flush`` firings within one transaction. After commit + # the tx is closed and its id will never recur on this session + # — drop the set so a long-lived session (Celery worker, CLI) + # doesn't grow it without bound. + session.info.pop(_PROCESSED_TXS_KEY, None) + # If a command set the action_kind but no flush fired (e.g. a + # save that touched nothing versioned), the value would + # otherwise leak into the next transaction. Drop it here as a + # belt-and-suspenders cleanup; the + # ``_stamp_action_kind_on_transaction`` helper already pops on + # the normal path. + session.info.pop(ACTION_KIND_KEY, None) + + def reset_action_kind_after_rollback(session: Session) -> None: + # When a command sets ``ACTION_KIND_KEY`` and then an exception + # fires before flush (e.g. validation error after the key is + # set), the transaction rolls back without the listener ever + # popping the key. The next save on the same session would + # then inherit the stale value and label an unrelated commit + # as "restore" / "import" / "clone". Pop here so a rolled-back + # action's intent doesn't leak forward. + session.info.pop(ACTION_KIND_KEY, None) + + event.listen(db.session, "before_flush", compute_change_records) + event.listen(db.session, "after_flush", flush_change_records) + event.listen(db.session, "after_commit", reset_processed_after_commit) + event.listen(db.session, "after_rollback", reset_action_kind_after_rollback) + setattr(db.session, _REGISTERED_SENTINEL, True) diff --git a/superset/versioning/changes/shadow_queries.py b/superset/versioning/changes/shadow_queries.py new file mode 100644 index 000000000000..1f6a290007f1 --- /dev/null +++ b/superset/versioning/changes/shadow_queries.py @@ -0,0 +1,323 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Shadow-table queries that drive child-collection diffs. + +Reads Continuum shadow tables (``table_columns_version`` / +``sql_metrics_version`` / ``dashboard_slices_version`` / +``slices_version``) under the validity-strategy semantics to compute +the pre/post state of child collections at a given transaction. Used +by the change-record listener's ``after_flush`` path once Continuum +has written the current transaction's shadow rows. + +**Inline imports.** Continuum's ``version_class`` and the Superset +model classes are imported inside each helper because this package is +loaded from ``init_versioning()`` before all SQLAlchemy mappers are +configured. The deferred imports keep the module-load graph free of +mapper-resolution side effects. +""" + +from __future__ import annotations + +from typing import Any + +import sqlalchemy as sa +from sqlalchemy.orm import Session + +from superset.versioning.baseline import CONTINUUM_BOOKKEEPING_COLUMNS +from superset.versioning.changes.state import jsonable +from superset.versioning.diff import ( + ChangeRecord, + diff_dashboard_slices, + diff_dataset_columns, + diff_dataset_metrics, +) + + +def shadow_rows_valid_at( + session: Session, + shadow_table: sa.Table, + fk_col_name: str, + fk_value: int, + tx: int, +) -> list[dict[str, Any]]: + """Return the live state of *shadow_table* rows whose FK column + (``fk_col_name``) equals *fk_value*, as of transaction *tx*. + + Uses Continuum's validity-strategy semantics: a row is "valid at tx" + when ``transaction_id <= tx`` AND (``end_transaction_id`` IS NULL OR + ``end_transaction_id`` > tx) AND it isn't a DELETE shadow. + + The returned dicts mirror the live row's column set (no Continuum + bookkeeping columns), so they can be passed straight to the + natural-key diff helpers (``diff_dataset_columns`` etc.). + """ + fk_col = getattr(shadow_table.c, fk_col_name) + rows = ( + session.connection() + .execute( + sa.select(shadow_table).where( + fk_col == fk_value, + shadow_table.c.transaction_id <= tx, + sa.or_( + shadow_table.c.end_transaction_id.is_(None), + shadow_table.c.end_transaction_id > tx, + ), + shadow_table.c.operation_type != 2, + ) + ) + .mappings() + .all() + ) + # Coerce values to JSON-safe forms — raw shadow rows can carry + # ``UUID``, ``datetime``, ``bytes`` etc. that don't survive the + # ``version_changes.from_value/to_value`` JSON column write. + return [ + { + k: jsonable(v) + for k, v in dict(row).items() + if k not in CONTINUUM_BOOKKEEPING_COLUMNS + } + for row in rows + ] + + +def _affected_dataset_ids_at_tx(session: Session, tx: int) -> set[int]: + """Datasets touched at *tx* — directly (parent shadow at tx) or + indirectly (column / metric shadow at tx).""" + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.connectors.sqla.models import SqlaTable, SqlMetric, TableColumn + + dataset_ids: set[int] = set() + parent_tbl = version_class(SqlaTable).__table__ + for row in session.connection().execute( + sa.select(parent_tbl.c.id).where(parent_tbl.c.transaction_id == tx) + ): + dataset_ids.add(row[0]) + for child_cls in (TableColumn, SqlMetric): + child_tbl = version_class(child_cls).__table__ + for row in session.connection().execute( + sa.select(child_tbl.c.table_id).where(child_tbl.c.transaction_id == tx) + ): + if row[0] is not None: + dataset_ids.add(row[0]) + return dataset_ids + + +def _dataset_child_records_for_tx_from_shadows( + session: Session, transaction_id: int +) -> dict[int, list[ChangeRecord]]: + """Compute column + metric diff records for each dataset touched at + *transaction_id*, reading from Continuum shadow tables. + + For each dataset: + * Post-state = rows valid at ``transaction_id`` in + ``table_columns_version`` / ``sql_metrics_version``. + * Pre-state = rows valid at ``transaction_id - 1`` in the same + shadow tables. + + With Continuum's validity-strategy semantics, "valid at tx N - 1" + is the state immediately before this transaction's effects (the + row that gets superseded at tx=N has ``end_transaction_id=N``, so + it satisfies ``end > N - 1``). Unrelated transactions between this + dataset's edits are transparent — they don't change validity for + this dataset's children. + + First-edit case: when there is no prior tx (the dataset's earliest + shadow IS at *transaction_id*), pre-state is empty. We skip rather + than emit "Added X" for every column — same "baseline = zero + records" semantics as the snapshot path. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.connectors.sqla.models import SqlMetric, TableColumn + + cols_tbl = version_class(TableColumn).__table__ + metrics_tbl = version_class(SqlMetric).__table__ + + result: dict[int, list[ChangeRecord]] = {} + for dataset_id in _affected_dataset_ids_at_tx(session, transaction_id): + # Skip the very first transaction for this dataset (no pre-state). + prior_tx = ( + session.connection() + .execute( + sa.select(sa.func.max(cols_tbl.c.transaction_id)).where( + cols_tbl.c.table_id == dataset_id, + cols_tbl.c.transaction_id < transaction_id, + ) + ) + .scalar() + ) + if prior_tx is None: + # No prior column shadow — could still be a metric-only edit; + # check metrics shadow too. + prior_tx = ( + session.connection() + .execute( + sa.select(sa.func.max(metrics_tbl.c.transaction_id)).where( + metrics_tbl.c.table_id == dataset_id, + metrics_tbl.c.transaction_id < transaction_id, + ) + ) + .scalar() + ) + if prior_tx is None: + continue + + post_cols = shadow_rows_valid_at( + session, cols_tbl, "table_id", dataset_id, transaction_id + ) + pre_cols = shadow_rows_valid_at( + session, cols_tbl, "table_id", dataset_id, prior_tx + ) + post_metrics = shadow_rows_valid_at( + session, metrics_tbl, "table_id", dataset_id, transaction_id + ) + pre_metrics = shadow_rows_valid_at( + session, metrics_tbl, "table_id", dataset_id, prior_tx + ) + + records: list[ChangeRecord] = [] + records.extend(diff_dataset_columns(pre_cols, post_cols)) + records.extend(diff_dataset_metrics(pre_metrics, post_metrics)) + if records: + result[dataset_id] = records + return result + + +def _affected_dashboard_ids_at_tx(session: Session, tx: int) -> set[int]: + """Dashboards touched at *tx* — directly (parent shadow at tx) or + indirectly (slice-membership shadow at tx).""" + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.dashboard import Dashboard + + dashboard_ids: set[int] = set() + parent_tbl = version_class(Dashboard).__table__ + for row in session.connection().execute( + sa.select(parent_tbl.c.id).where(parent_tbl.c.transaction_id == tx) + ): + dashboard_ids.add(row[0]) + + # M2M shadow: ``dashboard_slices_version`` is auto-generated by + # Continuum and lives in metadata — not a model class. Look it up + # from the metadata bag rather than via ``version_class``. + metadata = parent_tbl.metadata + if (m2m_tbl := metadata.tables.get("dashboard_slices_version")) is not None: + for row in session.connection().execute( + sa.select(m2m_tbl.c.dashboard_id).where(m2m_tbl.c.transaction_id == tx) + ): + if row[0] is not None: + dashboard_ids.add(row[0]) + return dashboard_ids + + +def _dashboard_slice_uuids_at_tx( + session: Session, dashboard_id: int, tx: int +) -> list[str]: + """Slice UUIDs attached to *dashboard_id* as of *tx*, read by joining + ``dashboard_slices_version`` (M2M membership) against + ``slices_version`` (slice content). + + Joining through both is necessary — and matches the same query + Continuum's M2M ``Reverter`` uses — because a slice that's + referenced by the M2M but has no slice-version row at this tx is + treated as "not yet versioned" and excluded. + + Returns UUIDs (strings) so the result can be diffed by the existing + :func:`diff_dashboard_slices` helper, which keys on uuid. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.slice import Slice + + metadata = version_class(Slice).__table__.metadata + m2m_tbl = metadata.tables.get("dashboard_slices_version") + slices_tbl = version_class(Slice).__table__ + if m2m_tbl is None: + return [] + + rows = ( + session.connection() + .execute( + sa.select(slices_tbl.c.uuid).where( + slices_tbl.c.id == m2m_tbl.c.slice_id, + m2m_tbl.c.dashboard_id == dashboard_id, + m2m_tbl.c.transaction_id <= tx, + sa.or_( + m2m_tbl.c.end_transaction_id.is_(None), + m2m_tbl.c.end_transaction_id > tx, + ), + m2m_tbl.c.operation_type != 2, + slices_tbl.c.transaction_id <= tx, + sa.or_( + slices_tbl.c.end_transaction_id.is_(None), + slices_tbl.c.end_transaction_id > tx, + ), + slices_tbl.c.operation_type != 2, + ) + ) + .all() + ) + return [str(r[0]) for r in rows if r[0] is not None] + + +def _dashboard_child_records_for_tx_from_shadows( + session: Session, transaction_id: int +) -> dict[int, list[ChangeRecord]]: + """Compute slice-membership diff records for each dashboard touched + at *transaction_id*, reading from Continuum shadow tables. + + Same pre/post logic as + :func:`_dataset_child_records_for_tx_from_shadows`. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.models.dashboard import Dashboard + + metadata = version_class(Dashboard).__table__.metadata + m2m_tbl = metadata.tables.get("dashboard_slices_version") + + result: dict[int, list[ChangeRecord]] = {} + for dashboard_id in _affected_dashboard_ids_at_tx(session, transaction_id): + prior_tx = None + if m2m_tbl is not None: + prior_tx = ( + session.connection() + .execute( + sa.select(sa.func.max(m2m_tbl.c.transaction_id)).where( + m2m_tbl.c.dashboard_id == dashboard_id, + m2m_tbl.c.transaction_id < transaction_id, + ) + ) + .scalar() + ) + if prior_tx is None: + continue + + post_uuids = _dashboard_slice_uuids_at_tx(session, dashboard_id, transaction_id) + pre_uuids = _dashboard_slice_uuids_at_tx(session, dashboard_id, prior_tx) + + records = diff_dashboard_slices(pre_uuids, post_uuids) + if records: + result[dashboard_id] = records + return result diff --git a/superset/versioning/changes/state.py b/superset/versioning/changes/state.py new file mode 100644 index 000000000000..541a87629ae9 --- /dev/null +++ b/superset/versioning/changes/state.py @@ -0,0 +1,237 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Pre/post-state reading and the per-entity diff dispatch. + +Three concerns live here: + +1. **JSON-safety coercion** — raw column values (``datetime``, ``UUID``, + ``bytes``, ``Decimal``) get converted to strings before they land in + the ``version_changes.from_value`` / ``to_value`` JSON columns. +2. **State capture** — :func:`_orm_to_post_state` serialises the + in-memory ORM object; :func:`_read_pre_state` reads the corresponding + pre-flush row directly from the DB inside ``session.no_autoflush``. +3. **Diff dispatch** — :func:`compute_records_for_entity` routes to the + right :mod:`superset.versioning.diff` helper based on the model + class name (string dispatch keeps this module free of hard imports + on the three entity classes, which avoids import-order coupling at + app-init time). + +Bulk insert of the computed records into the ``version_changes`` table +lives here too — it's the tail of the per-entity compute pipeline. +""" + +from __future__ import annotations + +import logging +from datetime import date, datetime +from decimal import Decimal +from typing import Any +from uuid import UUID + +import sqlalchemy as sa +from flask_appbuilder import Model +from sqlalchemy.orm import Session + +from superset.versioning.changes.table import version_changes_table +from superset.versioning.diff import ( + ChangeRecord, + diff_dashboard, + diff_dataset, + diff_slice, + scalar_fields_for, +) +from superset.versioning.utils import read_row_outside_flush + +logger = logging.getLogger(__name__) + + +# Per-model-class cache of the scalar-field set. Populated lazily on +# first save of a model. Reading from ``__table__.columns`` is cheap +# but not free; memoising keeps the save-path overhead budget (FR-021) +# from slowly growing with the set of distinct model classes seen. +_SCALAR_FIELDS_CACHE: dict[type, frozenset[str]] = {} + + +def _cached_scalar_fields(model_cls: type[Model]) -> frozenset[str]: + """Cached wrapper around :func:`scalar_fields_for`.""" + if model_cls not in _SCALAR_FIELDS_CACHE: + # ``Slice.params`` is walked by ``diff_slice_params`` for kind + # promotion; emitting it as one opaque ``field`` change would + # defeat that and flood the log with meaningless records. + # ``last_saved_at`` / ``last_saved_by_fk`` are stamped by + # ``UpdateChartCommand`` on every chart save; they're audit + # noise (same shape as ``changed_on`` / ``changed_by_fk``) and + # don't carry user-authored signal. + # ``Dashboard.json_metadata`` and ``position_json`` are JSON + # blobs walked structurally by ``diff_json_field`` (one record + # per changed top-level key); the raw scalar diff would emit + # one giant multi-KB record per save and swamp the response. + special: frozenset[str] = frozenset() + audit: frozenset[str] = frozenset() + if model_cls.__name__ == "Slice": + special = frozenset({"params"}) + audit = frozenset({"last_saved_at", "last_saved_by_fk"}) + elif model_cls.__name__ == "Dashboard": + special = frozenset({"json_metadata", "position_json"}) + _SCALAR_FIELDS_CACHE[model_cls] = scalar_fields_for( + model_cls, special=special, audit=audit + ) + return _SCALAR_FIELDS_CACHE[model_cls] + + +def jsonable(value: Any) -> Any: + """Convert a column value into a JSON-serialisable form. + + Slice has ``last_saved_at`` (datetime), datasets have datetime + columns, and any of these fields can land in ``from_value`` / + ``to_value`` of a ``version_changes`` row, which is a JSON column. + Python's default JSON encoder rejects ``datetime`` / ``UUID`` / + ``bytes`` / ``Decimal``, so the whole bulk insert fails if a single + record carries one. Convert to ISO / hex / str at record-construction + time. + """ + if isinstance(value, (datetime, date)): + return value.isoformat() + if isinstance(value, UUID): + return str(value) + if isinstance(value, bytes): + return value.hex() + if isinstance(value, Decimal): + # Stringify rather than ``float()`` to preserve precision; the + # diff engine compares string equality on ``from_value`` / + # ``to_value``, so coercing both sides to the same form is what + # matters. + return str(value) + return value + + +def _orm_to_post_state(obj: Any) -> dict[str, Any]: + """Serialise an ORM object's column attributes to a plain dict. + + We only read declared column attributes — not relationships or + hybrid properties — because the diff engine operates on scalar + values per its documented API. Values are passed through + :func:`jsonable` so the dict is JSON-safe end-to-end. + """ + state = sa.inspect(obj) + return { + col.key: jsonable(getattr(obj, col.key)) for col in state.mapper.column_attrs + } + + +def _read_pre_state( + session: Session, model_cls: type[Model], entity_id: int +) -> dict[str, Any] | None: + """Read the entity's pre-flush row directly from the DB and convert + non-JSON-safe types to strings so both sides of the diff compare on + the same form. Delegates the autoflush-suppressed read itself to + :func:`superset.versioning.utils.read_row_outside_flush`. + + Returns ``None`` if the row is missing (shouldn't happen for a dirty + existing object, but defensive against race conditions). + """ + table = model_cls.__table__ # type: ignore[attr-defined] + result = read_row_outside_flush(session, table, entity_id) + if result is None: + return None + # Convert non-JSON-safe types (datetime, UUID, bytes, Decimal) to + # strings so both sides of the diff compare on the same form and + # any value that ends up in ``from_value`` / ``to_value`` is + # acceptable to the JSON column on insert. + return {key: jsonable(value) for key, value in result.items()} + + +def compute_records_for_entity(session: Session, obj: Any) -> list[ChangeRecord]: + """Diff the pre-state (from DB) against the post-state (in memory). + + Dispatches to :func:`diff_slice` / :func:`diff_dashboard` / + :func:`diff_dataset` based on the model class name — string-based + dispatch is used to keep this module free of hard imports on the + three entity classes, which in turn avoids import-order coupling + at app-init time. + """ + model_cls = type(obj) + entity_id = getattr(obj, "id", None) + if entity_id is None: + return [] + + try: + pre_state = _read_pre_state(session, model_cls, entity_id) + except Exception: # pylint: disable=broad-except + logger.exception( + "version_changes: pre-state read failed for %s id=%s", + model_cls.__name__, + entity_id, + ) + return [] + + if pre_state is None: + return [] + + post_state = _orm_to_post_state(obj) + fields = _cached_scalar_fields(model_cls) + + name = model_cls.__name__ + if name == "Slice": + return diff_slice(pre_state, post_state, fields=fields) + if name == "Dashboard": + return diff_dashboard(pre_state, post_state, fields=fields) + if name == "SqlaTable": + return diff_dataset(pre_state, post_state, fields=fields) + return [] + + +def bulk_insert_records( + session: Session, + transaction_id: int, + buffered: dict[tuple[str, int], list[ChangeRecord]], +) -> None: + """Insert ``version_changes`` rows for one transaction via raw SQL. + + Uses the module-level :data:`version_changes_table` Table object + (which carries JSON column types, unlike ``sa.table(...)``) so the + connection marshals ``path`` / ``from_value`` / ``to_value`` Python + structures into JSON on insert. Skips the ORM flush round that + ``session.bulk_insert_mappings`` would cost inside an already- + active flush. + + ``buffered`` is a dict keyed on ``(entity_kind, entity_id)`` so + records for one entity — scalars from ``before_flush`` plus + children collected in ``after_flush`` — merge naturally under the + same key. ``sequence`` resets per entity so each entity's records + form a self-contained replay sequence. + """ + if not buffered: + return + rows = [] + for (entity_kind, entity_id), records in buffered.items(): + for seq, r in enumerate(records): + rows.append( + { + "transaction_id": transaction_id, + "entity_kind": entity_kind, + "entity_id": entity_id, + "sequence": seq, + "kind": r.kind, + "operation": r.operation, + "path": r.path, + "from_value": r.from_value, + "to_value": r.to_value, + } + ) + if rows: + session.connection().execute(version_changes_table.insert(), rows) diff --git a/superset/versioning/changes/table.py b/superset/versioning/changes/table.py new file mode 100644 index 000000000000..5b9df23680d1 --- /dev/null +++ b/superset/versioning/changes/table.py @@ -0,0 +1,83 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Schema definition for ``version_changes``. + +Declared against the shared ``Model.metadata`` so integration tests +that build schema via ``metadata.create_all()`` pick it up without the +Alembic migration running. Mirrors the shape of the +``56cd24c07170_add_versioning_tables`` migration byte-for-byte. Typed +columns (``sa.JSON`` for path / values) are required so the +connection's bulk-insert path marshals Python lists/dicts into JSON — +a lightweight ``sa.table(...)`` would not carry the type info and +SQLite's driver would reject the ``list`` as an unsupported bind. + +The schema lives in its own module to keep the listener and the +shadow-table-query helpers free of schema-construction boilerplate at +import time. +""" + +from __future__ import annotations + +import sqlalchemy as sa +from flask_appbuilder import Model + +_metadata = Model.metadata # pylint: disable=no-member + +version_changes_table = sa.Table( + "version_changes", + _metadata, + sa.Column("id", sa.BigInteger, primary_key=True, autoincrement=True), + # ``transaction_id`` references ``version_transaction.id`` at the DB + # level only — the FK + ON DELETE CASCADE live in the Alembic + # migration. Declaring the FK here would fail to resolve at Table + # creation time because ``version_transaction`` is built + # dynamically by SQLAlchemy-Continuum at mapper-configuration time; + # integration tests that materialise schema via ``metadata.create_all`` + # before Continuum runs would hit ``NoReferencedTableError``. Same + # pattern as the other versioning tables. + sa.Column("transaction_id", sa.BigInteger, nullable=False), + sa.Column("entity_kind", sa.String(32), nullable=False), + sa.Column("entity_id", sa.Integer, nullable=False), + sa.Column("sequence", sa.SmallInteger, nullable=False), + sa.Column("kind", sa.String(32), nullable=False), + sa.Column("operation", sa.String(16), nullable=False), + sa.Column("path", sa.JSON, nullable=False), + sa.Column("from_value", sa.JSON, nullable=True), + sa.Column("to_value", sa.JSON, nullable=True), + sa.UniqueConstraint( + "transaction_id", + "entity_kind", + "entity_id", + "sequence", + name="uq_version_changes_tx_entity_sequence", + ), + sa.Index("ix_version_changes_kind", "kind"), + sa.Index("ix_version_changes_transaction_id", "transaction_id"), + sa.Index("ix_version_changes_entity", "entity_kind", "entity_id"), + extend_existing=True, +) + +# Mapping from Python class name to the ``entity_kind`` value written +# to ``version_changes.entity_kind``. The API filters change records +# by this value (``WHERE entity_kind = 'chart'`` for the chart history +# endpoint, etc.) — kept short and user-facing-ish so downstream tools +# consuming the raw table read sensibly. +ENTITY_KIND_BY_CLASS_NAME: dict[str, str] = { + "Slice": "chart", + "Dashboard": "dashboard", + "SqlaTable": "dataset", +} diff --git a/superset/versioning/diff.py b/superset/versioning/diff.py new file mode 100644 index 000000000000..52700e57d123 --- /dev/null +++ b/superset/versioning/diff.py @@ -0,0 +1,1026 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Diff engine for the ``version_changes`` table (FR-016..FR-019). + +Hand-rolled because: + +- The on-disk ``path`` shape (array of segments) is a direct + representation of our chosen format; external diff libraries + return string paths or JSON-Pointer forms that would need + translation. +- Kind classification (``filter`` vs ``metric`` vs ``field`` etc.) + is co-located with diff walking, avoiding a second classification + pass over the generic diff output. +- Child-collection identity uses natural keys (``column_name``, + ``metric_name``, slice ``uuid``) — the same identity model + ``DatasetDAO.update_columns`` settled on (ADR-004). External + libraries default to list-index matching, which is wrong for our + data. + +See ADR (plan.md §"Key Design Decision: Hand-rolled diff engine") for +the full rationale. + +All functions in this module are pure: they take dicts (or lists of +dicts) and return a list of :class:`ChangeRecord`. The ORM->dict +conversion and Continuum transaction lookup happen in the capture +listener (T048), not here. This keeps the engine unit-testable without +an app context or DB. +""" + +from __future__ import annotations + +import logging +from collections.abc import Callable, Iterable +from dataclasses import dataclass +from typing import Any + +from superset.utils import json as _json + +logger = logging.getLogger(__name__) + +# Per-field recursion depth caps for the leaf-level diff walker. +# A cap is a usefulness bound, not a safety bound: it controls how deep +# into a nested JSON value the engine emits per-leaf records before +# stopping and treating the sub-tree as an opaque value. Values are +# tuned to the field's semantic shape — layout meta is shallow +# (text/sizes/colors), json_metadata and chart params can carry deep +# structures (native filters, adhoc filter sub-queries). +_LAYOUT_META_DIFF_DEPTH = 3 +_JSON_METADATA_DIFF_DEPTH = 6 +_SLICE_PARAMS_DIFF_DEPTH = 6 + +# Columns that are always excluded from change records, regardless of +# what ``__versioned__`` says. ``id`` / ``uuid`` are stable identifiers +# (not edited in normal flows). The four audit fields change on every +# save — emitting records for them would double every history entry +# with meaningless "timestamp changed, user stamped" rows that the UI +# would have to filter out anyway. +_AUDIT_FIELDS: frozenset[str] = frozenset( + { + "id", + "uuid", + "created_on", + "changed_on", + "created_by_fk", + "changed_by_fk", + } +) + +# Fields stripped from child-collection dict items (TableColumn, +# SqlMetric) before comparison and emission. ``changed_on`` / +# ``created_on`` / ``*_by_fk`` are audit fields that update on every +# save of the parent — without this filter, saving a dataset to add +# one column produces a record per existing column too (because their +# ``changed_on`` timestamps all refreshed). ``id`` and ``table_id`` +# are implementation details — ``id`` can change under the +# ``override_columns`` delete-and-reinsert pattern (ADR-004) even +# when the column is semantically unchanged; ``table_id`` is the +# parent FK and never meaningfully differs within one dataset's +# history. ``uuid`` stays stable across normal saves and is kept so +# the renderer can use it for identity if it needs to. +_CHILD_ITEM_OPAQUE_FIELDS: frozenset[str] = frozenset( + { + "id", + "table_id", + "changed_on", + "created_on", + "changed_by_fk", + "created_by_fk", + } +) + + +def _strip_opaque_fields(item: Any) -> Any: + """Return *item* with child-item audit/implementation fields removed. + + Pass-through for non-dict values (scalars, strings) — the strip + only applies where it matters (dataset column / metric dicts). + """ + if not isinstance(item, dict): + return item + return {k: v for k, v in item.items() if k not in _CHILD_ITEM_OPAQUE_FIELDS} + + +# Chart ``params`` sub-keys that are promoted to first-class kinds. +# Every other params sub-key falls through to ``kind="field"``. +_CHART_PARAMS_KIND_BY_KEY: dict[str, str] = { + "adhoc_filters": "filter", + "time_range": "time_range", + "color_scheme": "color_palette", + "metrics": "metric", + "groupby": "dimension", + "columns": "dimension", +} + +# Chart ``params`` sub-keys that are machine-stamped on save and don't +# carry user-authored signal — same category as ``last_saved_at`` on +# the scalar side. ``slice_id`` is a self-reference to the chart's +# own primary id; Superset's save paths add or refresh it on every +# save, producing a spurious "field" record on the first save after +# a chart's params were stored without it. +_CHART_PARAMS_AUDIT_KEYS: frozenset[str] = frozenset({"slice_id"}) + + +def scalar_fields_for( + model_cls: Any, + *, + special: frozenset[str] = frozenset(), + audit: frozenset[str] = frozenset(), +) -> frozenset[str]: + """Scalar columns on ``model_cls`` that should produce change records. + + Derived from the model itself at call time so contributors (and + downstream derivatives) don't have to maintain a parallel whitelist + in this module. Adding a new column to ``Dashboard``, ``Slice``, or + ``SqlaTable`` — whether upstream or in a fork — automatically flows + through to ``version_changes`` on the next save. + + Excludes, in order: + + 1. The model's own ``__versioned__.exclude`` list, so change records + stay consistent with Continuum's shadow tables. If Continuum + isn't tracking a column, the change log shouldn't either. + 2. :data:`_AUDIT_FIELDS` — ``id``, ``uuid``, and the audit + timestamps / user-id columns shared across the three entity types. + 3. The caller's ``audit`` set — model-specific save-side-effect + columns that aren't user-authored content. ``Slice.last_saved_at`` + / ``last_saved_by_fk`` are stamped on every chart save by + ``UpdateChartCommand``, similar to how ``changed_on`` is stamped + by the ORM event listener; emitting "field" records for them + would noise up the change log with one entry per save that + carries no user-meaningful signal. + 4. The caller's ``special`` set — columns handled by a dedicated + differ elsewhere. ``Slice.params``, for example, is walked by + :func:`diff_slice_params` to produce first-class ``filter`` / + ``time_range`` / ``metric`` / ``dimension`` records; emitting + it as a single opaque ``field`` would defeat that. + """ + try: + table = model_cls.__table__ + except AttributeError: + return frozenset() + columns = frozenset(c.name for c in table.columns) + continuum_exclude = frozenset( + getattr(model_cls, "__versioned__", {}).get("exclude", []) or [] + ) + return columns - continuum_exclude - _AUDIT_FIELDS - audit - special + + +@dataclass(frozen=True) +class ChangeRecord: + """One atomic change, as stored in ``version_changes``. + + Fields match the ``version_changes`` columns one-to-one so the + capture listener can serialise a list of these to + ``session.bulk_insert_mappings`` without translation. + + Three orthogonal dimensions: + * ``kind`` — what type of thing changed (``filter`` / ``column`` / + ``header`` / ``field`` / etc.). Content category. + * ``operation`` — what happened to it (``add`` / ``remove`` / + ``move`` / ``edit``). ``move`` only fires for layout records. + * ``path`` — pure navigation address; no verb encoded. + + The transaction-level fourth dimension (``trigger``: ``restore`` / + ``import`` / ``clone``) lives on ``version_transaction``, not here. + """ + + kind: str + operation: str + path: list[Any] + from_value: Any + to_value: Any + + +Key = str | int + + +def _operation_from_values(from_value: Any, to_value: Any) -> str: + """Derive the per-record ``operation`` verb from ``from_value`` / + ``to_value`` nullability. + + * ``add`` — ``from_value`` is ``None`` and ``to_value`` is not. + * ``remove`` — ``to_value`` is ``None`` and ``from_value`` is not. + * ``edit`` — both populated (or both null, which shouldn't reach here). + + Used by every emit site except ``_diff_layout_node``, which emits + ``move`` records (parent reparenting) that cannot be derived from + value nullability alone. + """ + if from_value is None and to_value is not None: + return "add" + if to_value is None and from_value is not None: + return "remove" + return "edit" + + +def _values_equivalent(from_value: Any, to_value: Any) -> bool: + """True if a transition from ``from_value`` to ``to_value`` should + NOT produce a record. + + Beyond plain ``==`` equality, treats ``None`` and ``""`` as equivalent: + Superset's save paths normalize nullable strings to ``""`` on first + write (e.g. ``Dashboard.css``, ``certified_by``, + ``certification_details``), so a first-save transition between + null and empty string carries no user-authored signal. + """ + if from_value == to_value: + return True + if from_value in (None, "") and to_value in (None, ""): + return True + return False + + +def _diff_scalar( + field_name: str, + from_value: Any, + to_value: Any, +) -> ChangeRecord | None: + """Emit a generic ``kind="field"`` record when a scalar differs.""" + if _values_equivalent(from_value, to_value): + return None + return ChangeRecord( + kind="field", + operation=_operation_from_values(from_value, to_value), + path=[field_name], + from_value=from_value, + to_value=to_value, + ) + + +def _recursive_leaf_diff( + kind: str, + path_prefix: list[Any], + pre: Any, + post: Any, + *, + max_depth: int, +) -> list[ChangeRecord]: + """Walk matched dict structures and emit one ``ChangeRecord`` per + changed leaf. + + Recursion rules: + + * Both sides equal (per :func:`_values_equivalent`) → no record. + * Both sides ``dict`` AND recursion depth below ``max_depth`` → + recurse into each key, extending the path by the key. + * All other cases (scalar mismatch, list on either side, mismatched + types, both dicts but depth-capped) → emit one leaf record with + ``from_value`` / ``to_value`` carrying the raw pre/post values. + + Lists are treated as opaque on purpose — positional paths break on + reorder and most lists in Superset's JSON blobs (adhoc filters, + metrics, dataset columns) already have a dedicated natural-key + walker upstream that emits per-element records with the right + identity. + + A depth-cap hit on dict-vs-dict emits a debug log so production + tuning can see when a field's cap is too tight to capture all + meaningful change. + """ + + def _walk(pre: Any, post: Any, path: list[Any], depth: int) -> list[ChangeRecord]: + if _values_equivalent(pre, post): + return [] + if depth < max_depth and isinstance(pre, dict) and isinstance(post, dict): + records: list[ChangeRecord] = [] + for key in sorted(set(pre) | set(post)): + records.extend( + _walk(pre.get(key), post.get(key), [*path, key], depth + 1) + ) + return records + if isinstance(pre, dict) and isinstance(post, dict): + logger.debug( + "version_changes: depth cap %d hit at path=%s — sub-tree " + "emitted as opaque leaf", + max_depth, + path, + ) + return [ + ChangeRecord( + kind=kind, + operation=_operation_from_values(pre, post), + path=list(path), + from_value=pre, + to_value=post, + ) + ] + + return _walk(pre, post, path_prefix, 0) + + +def _diff_list_by_natural_key( + kind: str, + path_prefix: list[Any], + from_list: list[Any] | None, + to_list: list[Any] | None, + key_fn: Callable[[Any], Key | None], +) -> list[ChangeRecord]: + """Diff two lists, matching elements by natural key. + + Emits one record per add / remove / modify. When ``key_fn`` returns + ``None`` for an item (natural key missing or empty), the item falls + back to its position as a synthetic key — so insertions in the + middle of a keyless list still produce sensible records, at the + cost of position-dependent identity. + """ + from_list = from_list or [] + to_list = to_list or [] + + def _effective_key(raw: Key | None, idx: int) -> Key: + if raw is None or raw == "": + return idx + return raw + + from_by_key: dict[Key, Any] = {} + for idx, item in enumerate(from_list): + from_by_key[_effective_key(key_fn(item), idx)] = item + to_by_key: dict[Key, Any] = {} + for idx, item in enumerate(to_list): + to_by_key[_effective_key(key_fn(item), idx)] = item + + records: list[ChangeRecord] = [] + # Preserve `from` order then append `to`-only keys, so sequence is + # deterministic across runs. For dict items (dataset columns / + # metrics) we strip audit/implementation fields before comparing + # AND before emitting — otherwise a save that only adds a new + # column would also emit "changed" records for every existing + # column, because their ``changed_on`` timestamps all refreshed. + # The stripped from/to are what the renderer sees; the per-column + # audit trail is already aggregated at the transaction level in + # ``version_transaction`` (``user_id`` + ``issued_at``). + for k, from_item in from_by_key.items(): + to_item = to_by_key.get(k) + stripped_from = _strip_opaque_fields(from_item) + if to_item is None: + records.append( + ChangeRecord( + kind=kind, + operation="remove", + path=[*path_prefix, k], + from_value=stripped_from, + to_value=None, + ) + ) + continue + stripped_to = _strip_opaque_fields(to_item) + if stripped_from != stripped_to: + records.append( + ChangeRecord( + kind=kind, + operation="edit", + path=[*path_prefix, k], + from_value=stripped_from, + to_value=stripped_to, + ) + ) + for k, to_item in to_by_key.items(): + if k not in from_by_key: + records.append( + ChangeRecord( + kind=kind, + operation="add", + path=[*path_prefix, k], + from_value=None, + to_value=_strip_opaque_fields(to_item), + ) + ) + return records + + +def _filter_key(f: Any) -> Key | None: + """Natural key for an adhoc filter — its subject (column name). + + Users rarely have two filters on the same column; when they do the + secondary dimensions (operator, comparator) appear in the record's + from/to values so the renderer can disambiguate. + """ + return f.get("subject") if isinstance(f, dict) else None + + +def _metric_key(m: Any) -> Key | None: + """Natural key for a metric: prefer ``label``, fall back to column+aggregate.""" + if not isinstance(m, dict): + return None + if label := m.get("label"): + return label + column = m.get("column") + col_name = column.get("column_name") if isinstance(column, dict) else None + agg = m.get("aggregate") + if col_name and agg: + return f"{agg}({col_name})" + return None + + +def _dimension_key(d: Any) -> Key | None: + """Natural key for a groupby/columns element — usually a bare string.""" + if isinstance(d, str): + return d + if isinstance(d, dict): + return d.get("label") or d.get("column_name") + return None + + +def _coerce_params(p: Any) -> dict[str, Any]: + """Decode ``Slice.params`` which is stored as a JSON string.""" + if p is None: + return {} + if isinstance(p, str): + try: + decoded = _json.loads(p) + except _json.JSONDecodeError: + return {} + return decoded if isinstance(decoded, dict) else {} + if isinstance(p, dict): + return p + return {} + + +def diff_slice_params( + from_params: Any, + to_params: Any, +) -> list[ChangeRecord]: + """Diff the ``Slice.params`` JSON blob, promoting known keys to kinds.""" + from_p = _coerce_params(from_params) + to_p = _coerce_params(to_params) + records: list[ChangeRecord] = [] + all_keys = (set(from_p) | set(to_p)) - _CHART_PARAMS_AUDIT_KEYS + for key in sorted(all_keys): + from_v = from_p.get(key) + to_v = to_p.get(key) + if _values_equivalent(from_v, to_v): + continue + kind = _CHART_PARAMS_KIND_BY_KEY.get(key) + if kind == "filter" and isinstance(from_v, list) and isinstance(to_v, list): + records.extend( + _diff_list_by_natural_key( + "filter", + ["params", "adhoc_filters"], + from_v, + to_v, + _filter_key, + ) + ) + elif kind == "metric" and isinstance(from_v, list) and isinstance(to_v, list): + records.extend( + _diff_list_by_natural_key( + "metric", + ["params", "metrics"], + from_v, + to_v, + _metric_key, + ) + ) + elif ( + kind == "dimension" and isinstance(from_v, list) and isinstance(to_v, list) + ): + records.extend( + _diff_list_by_natural_key( + "dimension", + ["params", key], + from_v, + to_v, + _dimension_key, + ) + ) + elif kind: + # scalar first-class kind (time_range, color_palette). + # For genuinely scalar values the recursion emits one leaf + # record exactly as before; for the unusual case of a dict + # value (custom viz params) it recurses to the leaf. + records.extend( + _recursive_leaf_diff( + kind=kind, + path_prefix=["params", key], + pre=from_v, + post=to_v, + max_depth=_SLICE_PARAMS_DIFF_DEPTH, + ) + ) + else: + # unknown params sub-key: generic field change, recursed + # to the leaf so a deep custom-viz option doesn't ship its + # whole sub-tree on both sides. + records.extend( + _recursive_leaf_diff( + kind="field", + path_prefix=["params", key], + pre=from_v, + post=to_v, + max_depth=_SLICE_PARAMS_DIFF_DEPTH, + ) + ) + return records + + +def diff_scalar_fields( + pre: dict[str, Any], + post: dict[str, Any], + *, + fields: Iterable[str], +) -> list[ChangeRecord]: + """Emit one ``kind="field"`` record per differing field in ``fields``. + + The ``fields`` iterable is supplied by the caller — typically + :func:`scalar_fields_for` at listener wiring time. Keeping the + field list outside this function means adding a new column to a + model does not require a matching edit here. + """ + records: list[ChangeRecord] = [] + for field in sorted(fields): + record = _diff_scalar(field, pre.get(field), post.get(field)) + if record is not None: + records.append(record) + return records + + +def diff_slice( + pre: dict[str, Any], + post: dict[str, Any], + *, + fields: Iterable[str], +) -> list[ChangeRecord]: + """Full Slice (chart) diff — scalars plus params classification. + + Pass ``fields=scalar_fields_for(Slice, special=frozenset({"params"}))`` + to get the ``params``-excluded scalar set; ``Slice.params`` is diffed + separately by :func:`diff_slice_params` for kind promotion. + """ + records = diff_scalar_fields(pre, post, fields=fields) + records.extend(diff_slice_params(pre.get("params"), post.get("params"))) + return records + + +def diff_json_field( + field_name: str, + from_value: Any, + to_value: Any, + *, + exclude_keys: frozenset[str] = frozenset(), + max_depth: int = _JSON_METADATA_DIFF_DEPTH, +) -> list[ChangeRecord]: + """Diff a TEXT column that stores a JSON dict, emitting one record + per changed leaf. + + Used for ``Dashboard.json_metadata`` (``position_json`` has its + own structural diff via :func:`diff_dashboard_layout`). Saving the + blob verbatim into ``from_value`` / ``to_value`` would swamp the + change log with multi-KB strings on every save; recursing into the + parsed dict reduces noise to "exactly which leaf changed". + + *exclude_keys* names sub-keys that are frontend-derived / + auto-stamped on save and don't carry user-authored signal. Same + rationale as the ``audit`` parameter on + :func:`scalar_fields_for` for the parent-column level. + + Path is ``[field_name, key, ...]`` for leaf records, mirroring + :func:`diff_slice_params`'s ``["params", key, ...]`` shape so + renderers can use a single addressing scheme across the chart + and dashboard sides. + """ + from_p = _coerce_params(from_value) + to_p = _coerce_params(to_value) + records: list[ChangeRecord] = [] + for key in sorted(set(from_p) | set(to_p)): + if key in exclude_keys: + continue + records.extend( + _recursive_leaf_diff( + kind="field", + path_prefix=[field_name, key], + pre=from_p.get(key), + post=to_p.get(key), + max_depth=max_depth, + ) + ) + return records + + +# json_metadata sub-keys that the frontend auto-stamps / auto-derives +# on save. They mirror dashboard membership and chart inventory, not +# user-authored content, so they noise up the change log without +# carrying intent. The records produced for these keys can be ~50KB +# (full label-colour dict) for a one-chart save. +# +# chart_configuration: per-chart cross-filter scope state, +# re-derived when charts are added/removed. +# global_chart_configuration: dashboard-wide filter scope; the +# ``chartsInScope`` list mirrors live +# dashboard membership. +# map_label_colors: label → colour map, re-stamped on save +# from currently-visible filter values. +# show_chart_timestamps: frontend toggle, defaults applied on +# save when missing. +# color_namespace: scoped colour-scheme namespace, frontend- +# derived from the chart set. +DASHBOARD_JSON_METADATA_AUDIT_KEYS: frozenset[str] = frozenset( + { + "chart_configuration", + "global_chart_configuration", + "map_label_colors", + "show_chart_timestamps", + "color_namespace", + } +) + + +# Layout component types and how they map to record ``kind`` strings. +# ``HEADER_ID`` is excluded — that's the dashboard's title bar, mirrored +# from ``dashboard_title``. ``ROOT_ID`` and ``GRID_ID`` are structural +# singletons whose only deltas are children lists, which we infer from +# the moves of the children themselves. +_LAYOUT_TYPE_TO_KIND: dict[str, str] = { + "CHART": "chart", + "ROW": "row", + "COLUMN": "column", + "TAB": "tab", + "TABS": "tabs", + "HEADER": "header", + "MARKDOWN": "markdown", + "DIVIDER": "divider", +} + +# Layout components we never emit records for: ROOT_ID is the layout +# root (always present, never moves); GRID_ID is the singleton vertical +# stack inside ROOT_ID; HEADER_ID is the dashboard's title bar (already +# covered by the ``dashboard_title`` scalar field). +_LAYOUT_SUPPRESSED_IDS: frozenset[str] = frozenset({"ROOT_ID", "GRID_ID", "HEADER_ID"}) + + +def _layout_component_label(node: dict[str, Any]) -> str | None: + """Extract a human-readable label from a layout node, when one + exists. Used to build the ``from_value`` / ``to_value`` payload so + the UI can render messages like "Added chart 'Foo'" without + needing to fetch related entities. + """ + meta = node.get("meta") or {} + if not isinstance(meta, dict): + return None + for key in ("sliceName", "label", "text"): + value = meta.get(key) + if isinstance(value, str) and value.strip(): + return value + return None + + +def _layout_node_payload(node: dict[str, Any]) -> dict[str, Any]: + """Minimal payload describing a layout node — enough for the UI + to render the change without dragging the full layout snippet + (which can be ~1KB per row when CHART nodes carry colour configs). + """ + meta = node.get("meta") or {} + if not isinstance(meta, dict): + meta = {} + payload: dict[str, Any] = {"id": node.get("id"), "type": node.get("type")} + if (label := _layout_component_label(node)) is not None: + payload["name"] = label + if (chart_id := meta.get("chartId")) is not None: + payload["chartId"] = chart_id + # ``uuid`` (slice uuid for CHART nodes) lets the M2M-vs-layout + # dedupe in :func:`fold_dashboard_layout_with_chart_changes` + # match on the same key — :func:`diff_dashboard_slices` keys its + # records by uuid, not chartId. + if (slice_uuid := meta.get("uuid")) is not None: + payload["uuid"] = slice_uuid + return payload + + +def _layout_parent_id(node: dict[str, Any]) -> Any: + """The immediate-parent node id for a layout component — the last + entry in ``parents``. Used to detect moves: same id, different + parent.""" + parents = node.get("parents") or [] + if not isinstance(parents, list) or not parents: + return None + return parents[-1] + + +def _meta_excluding_position(node: dict[str, Any]) -> dict[str, Any]: + """Meta dict with ``parents``-equivalent positional bits removed + so two nodes that differ ONLY in where they sit compare equal at + the meta level. Move detection uses ``parents`` directly; this is + for "edit" (meta change) detection.""" + meta = node.get("meta") or {} + return dict(meta) if isinstance(meta, dict) else {} + + +def _diff_layout_node( + node_id: str, + pre_node: dict[str, Any] | None, + post_node: dict[str, Any] | None, +) -> list[ChangeRecord]: + """Diff one component slot in the layout dict and return records for + the logical action — add, remove, move, edit. + + add / remove / move emit a single record carrying the minimal node + payload (so the renderer can describe the affected component). + edit recurses into the node's ``meta`` dict and emits one record per + changed leaf, capped at ``_LAYOUT_META_DIFF_DEPTH``. + + Returns an empty list when the slot is unchanged or holds an unknown + component type. + """ + node_for_kind = post_node or pre_node or {} + kind = _LAYOUT_TYPE_TO_KIND.get(node_for_kind.get("type") or "") + if kind is None: + return [] # unknown component type — skip rather than emit garbage + + if pre_node is None and post_node is not None: + return [ + ChangeRecord( + kind=kind, + operation="add", + path=[node_id], + from_value=None, + to_value=_layout_node_payload(post_node), + ) + ] + if post_node is None and pre_node is not None: + return [ + ChangeRecord( + kind=kind, + operation="remove", + path=[node_id], + from_value=_layout_node_payload(pre_node), + to_value=None, + ) + ] + + # Both present — check move first, then edit. + assert pre_node is not None + assert post_node is not None + pre_parent = _layout_parent_id(pre_node) + if pre_parent != (post_parent := _layout_parent_id(post_node)): + return [ + ChangeRecord( + kind=kind, + operation="move", + path=[node_id], + from_value={**_layout_node_payload(pre_node), "parent": pre_parent}, + to_value={**_layout_node_payload(post_node), "parent": post_parent}, + ) + ] + + # Edit: recurse into meta and emit one record per changed leaf. + # Path shape ``[node_id, , ...]``. The verb (operation) is + # derived per-leaf by the recursion via ``_operation_from_values``; + # a leaf added inside an existing node gets ``add`` and so on. The + # node-level "this was an edit" fact is implicit in the path shape + # carrying segments after ``node_id``. + return _recursive_leaf_diff( + kind=kind, + path_prefix=[node_id], + pre=_meta_excluding_position(pre_node), + post=_meta_excluding_position(post_node), + max_depth=_LAYOUT_META_DIFF_DEPTH, + ) + + +def diff_dashboard_layout( + pre: Any, + post: Any, +) -> list[ChangeRecord]: + """Structural diff of a dashboard's ``position_json``, emitting one + record per logical layout action. + + Walks both sides keyed on the component ``id`` (e.g. + ``"CHART-mkPZLOnWCElgL0Udp1gVK"``): + + * id present only in *post* → ``op=add``, ``from_value=None``, + ``to_value=`` + * id present only in *pre* → ``op=remove``, payload swapped + * id in both, ``parents`` differs → ``op=move``, payloads carry + old + new parent + * id in both, parents equal, ``meta`` differs → ``op=edit``, + payloads carry old + new meta + * id in both, equal → no record + + The verb lives in each record's ``operation`` field + (``add`` / ``remove`` / ``move`` / ``edit``); ``path`` locates the + component as ``[]`` (``[, , …]`` + for an ``edit`` that recurses into ``meta``). Paths no longer carry + the verb — see :func:`_layout_chart_uuids_by_verb`. + + ``ROOT_ID`` / ``GRID_ID`` / ``HEADER_ID`` are suppressed (see + :data:`_LAYOUT_SUPPRESSED_IDS`). + """ + pre_nodes = _layout_nodes(pre) + post_nodes = _layout_nodes(post) + records: list[ChangeRecord] = [] + for node_id in sorted(set(pre_nodes) | set(post_nodes)): + records.extend( + _diff_layout_node(node_id, pre_nodes.get(node_id), post_nodes.get(node_id)) + ) + return records + + +def _layout_nodes(raw: Any) -> dict[str, dict[str, Any]]: + """Coerce *raw* (a ``position_json`` blob or already-parsed dict) into + the ``{node_id: node_dict}`` shape used by the layout diff, filtering + out non-dict values and the always-present root/grid/header singletons. + """ + parsed = _coerce_params(raw) + return { + k: v + for k, v in parsed.items() + if isinstance(v, dict) and k not in _LAYOUT_SUPPRESSED_IDS + } + + +def diff_dashboard( + pre: dict[str, Any], + post: dict[str, Any], + *, + fields: Iterable[str], +) -> list[ChangeRecord]: + """Dashboard diff: scalar fields plus structural diff of + ``json_metadata`` and ``position_json``. + + Promoting ``position_json`` to ``kind="layout"`` or + ``json_metadata.native_filter_configuration`` to ``kind="filter"`` + is deferred to Phase 2 alongside the UI that would render them + (spec Clarifications §Session 2026-04-24); until then, both fields + fall through to ``kind="field"`` records keyed by sub-key. + """ + records = diff_scalar_fields(pre, post, fields=fields) + records.extend( + diff_json_field( + "json_metadata", + pre.get("json_metadata"), + post.get("json_metadata"), + exclude_keys=DASHBOARD_JSON_METADATA_AUDIT_KEYS, + ) + ) + records.extend( + diff_dashboard_layout(pre.get("position_json"), post.get("position_json")) + ) + return records + + +def _layout_chart_uuids_by_verb( + records: list[ChangeRecord], +) -> tuple[set[Any], set[Any]]: + """Scan *records* for layout ``add``/``remove`` records on charts and + return ``(added_uuids, removed_uuids)`` sets. + + Keys off ``operation`` (the explicit verb column) rather than + ``path[0]`` — paths no longer carry the verb. + """ + added: set[Any] = set() + removed: set[Any] = set() + for r in records: + if r.kind != "chart": + continue + # Layout chart records have ``path = [node_id]`` (length 1) for + # add/remove/move and ``[node_id, ...leaf]`` for edits. We only + # care about the structural add/remove cases here. + if len(r.path) != 1: + continue + if r.operation == "add" and isinstance(r.to_value, dict): + uuid_ = r.to_value.get("uuid") + if uuid_ is not None: + added.add(uuid_) + elif r.operation == "remove" and isinstance(r.from_value, dict): + uuid_ = r.from_value.get("uuid") + if uuid_ is not None: + removed.add(uuid_) + return added, removed + + +def _is_redundant_m2m_chart_record( + r: ChangeRecord, added_uuids: set[Any], removed_uuids: set[Any] +) -> bool: + """Return ``True`` when *r* is an M2M-style slice record that + duplicates an already-captured layout add/remove for the same uuid. + + M2M slice records have path ``["slices", uuid]`` (length 2); their + info is strictly less than the corresponding layout record's + (no name, no parent), so the layout side wins on dedup. + """ + if r.kind != "chart" or len(r.path) != 2 or r.path[0] != "slices": + return False + slice_uuid = r.path[1] + if r.from_value is None and r.to_value is not None: + return slice_uuid in added_uuids + if r.to_value is None and r.from_value is not None: + return slice_uuid in removed_uuids + return False + + +def fold_dashboard_layout_with_chart_changes( + records: list[ChangeRecord], +) -> list[ChangeRecord]: + """When a dashboard save adds/removes charts, the ``slices`` M2M + diff and the layout diff each emit a record for the same logical + action. Drop the M2M ``kind="chart"`` records — the layout-side + record carries more information (chart name, parent container). + + The matching is by slice uuid: ``diff_dashboard_slices`` produces + records with path ``["slices", ]``; the layout + payloads carry the same uuid (sourced from + ``position_json.CHART-x.meta.uuid``). We dedupe on that key. + + Called from the change-records listener after the M2M and layout + diffs are both merged into the per-entity buffer. + """ + added_uuids, removed_uuids = _layout_chart_uuids_by_verb(records) + return [ + r + for r in records + if not _is_redundant_m2m_chart_record(r, added_uuids, removed_uuids) + ] + + +def diff_dataset( + pre: dict[str, Any], + post: dict[str, Any], + *, + fields: Iterable[str], +) -> list[ChangeRecord]: + """SqlaTable scalar-field diff. All paths emit ``kind="field"``. + + Children (columns, metrics) are diffed separately via + :func:`diff_dataset_columns` / :func:`diff_dataset_metrics`. The + listener reads them from Continuum shadow tables + (``table_columns_version`` / ``sql_metrics_version``) rather than + walking the ORM collection. + """ + return diff_scalar_fields(pre, post, fields=fields) + + +def diff_dataset_columns( + from_columns: list[dict[str, Any]] | None, + to_columns: list[dict[str, Any]] | None, +) -> list[ChangeRecord]: + """Child-collection diff on TableColumn rows, keyed by column_name.""" + return _diff_list_by_natural_key( + kind="column", + path_prefix=["columns"], + from_list=from_columns, + to_list=to_columns, + key_fn=lambda c: c.get("column_name") if isinstance(c, dict) else None, + ) + + +def diff_dataset_metrics( + from_metrics: list[dict[str, Any]] | None, + to_metrics: list[dict[str, Any]] | None, +) -> list[ChangeRecord]: + """Child-collection diff on SqlMetric rows, keyed by metric_name.""" + return _diff_list_by_natural_key( + kind="metric", + path_prefix=["metrics"], + from_list=from_metrics, + to_list=to_metrics, + key_fn=lambda m: m.get("metric_name") if isinstance(m, dict) else None, + ) + + +def diff_dashboard_slices( + from_slice_uuids: list[str] | None, + to_slice_uuids: list[str] | None, +) -> list[ChangeRecord]: + """Diff a dashboard's chart membership, keyed by slice uuid. + + Pure set-diff: added uuids get ``from_value=None, to_value=uuid``; + removed uuids get the inverse. No "changed" case because chart + associations are identity-only (the list element IS the uuid). + """ + from_set = set(from_slice_uuids or []) + to_set = set(to_slice_uuids or []) + records: list[ChangeRecord] = [] + for uuid_ in sorted(from_set - to_set): + records.append( + ChangeRecord( + kind="chart", + operation="remove", + path=["slices", uuid_], + from_value=uuid_, + to_value=None, + ) + ) + for uuid_ in sorted(to_set - from_set): + records.append( + ChangeRecord( + kind="chart", + operation="add", + path=["slices", uuid_], + from_value=None, + to_value=uuid_, + ) + ) + return records diff --git a/superset/versioning/etag.py b/superset/versioning/etag.py new file mode 100644 index 000000000000..643e1a95dadc --- /dev/null +++ b/superset/versioning/etag.py @@ -0,0 +1,75 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""ETag header emission for the entity-versioning API surface.""" + +from __future__ import annotations + +from typing import TYPE_CHECKING +from uuid import UUID + +import sqlalchemy as sa +from flask_appbuilder import Model + +from superset.extensions import db + +if TYPE_CHECKING: + from flask import Response + + +def set_version_etag(response: "Response", version_uuid: UUID | None) -> "Response": + """Attach ``ETag: ""`` to *response*. + + Uses RFC 7232 strong-validator form (no leading ``W/``); the response + header value is wrapped in double quotes per the spec. No-op when + *version_uuid* is ``None`` (entity has no version rows yet). + """ + if version_uuid is not None: + response.headers["ETag"] = f'"{version_uuid}"' + return response + + +def set_version_etag_by_uuid( + response: "Response", + model_cls: type[Model], + entity_uuid: UUID, + *, + entity_id: int | None = None, +) -> "Response": + """Attach ``ETag`` derived from *entity_uuid*'s current live version. + + If *entity_id* is provided the helper uses it directly; otherwise it + runs ``SELECT id WHERE uuid = ?`` to resolve it. Pass *entity_id* + from call sites that already have the entity in hand (e.g. via + :func:`superset.versioning.api_helpers.resolve_endpoint_path_entity`) + so the lookup doesn't fire twice — every list/get versions request + previously cost an extra round-trip here on top of the resolve. + + No-op when the entity is missing or has no version rows yet. + """ + # pylint: disable=import-outside-toplevel + from superset.daos.version import VersionDAO + + if entity_id is None: + entity_id = db.session.scalar( + sa.select(model_cls.id).where(model_cls.uuid == entity_uuid) + ) + if entity_id is None: + return response + return set_version_etag( + response, + VersionDAO.current_live_version_uuid(model_cls, entity_id, entity_uuid), + ) diff --git a/superset/versioning/factory.py b/superset/versioning/factory.py new file mode 100644 index 000000000000..1de6bec83902 --- /dev/null +++ b/superset/versioning/factory.py @@ -0,0 +1,307 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +from __future__ import annotations + +import logging +from collections.abc import Callable +from typing import Any + +import sqlalchemy as sa +import sqlalchemy.orm as sa_orm +from sqlalchemy_continuum import is_modified, version_class +from sqlalchemy_continuum.operation import Operation +from sqlalchemy_continuum.plugins.base import Plugin +from sqlalchemy_continuum.plugins.flask import FlaskPlugin +from sqlalchemy_continuum.transaction import TransactionFactory +from sqlalchemy_continuum.utils import versioned_column_properties + +from superset.utils import json +from superset.versioning.diff import DASHBOARD_JSON_METADATA_AUDIT_KEYS + +logger = logging.getLogger(__name__) + + +def _normalize_dashboard_json_metadata(value: Any) -> Any: + """Parse ``dashboards.json_metadata`` and drop frontend-stamped audit + sub-keys so a save that only re-stamps ``map_label_colors`` (etc.) + compares equal to its predecessor. + + ``map_label_colors`` is regenerated client-side from the + ``LabelsColorMap`` singleton on every save (see + ``saveDashboardRequest`` in + ``superset-frontend/src/dashboard/actions/dashboardState.ts``). + The singleton's contents depend on which charts have rendered in + the page session, so two saves with no user-authored change produce + different bytes. The diff engine ignores the same audit sub-keys + (``DASHBOARD_JSON_METADATA_AUDIT_KEYS`` in + ``superset/versioning/diff.py``); aligning the skip-plugin's + comparison with that filter keeps the two paths consistent. + """ + if value is None or value == "": + return value + try: + parsed = json.loads(value) + except (TypeError, ValueError): + return value + if not isinstance(parsed, dict): + return parsed + return { + k: v for k, v in parsed.items() if k not in DASHBOARD_JSON_METADATA_AUDIT_KEYS + } + + +# Per-class column normalizers, keyed on (class_name, column_name). Class +# name is used (rather than class itself) so importing the model classes +# at module load is unnecessary — keeps the plugin importable before +# ``make_versioned()`` has registered the version classes. +_COLUMN_NORMALIZERS: dict[tuple[str, str], Callable[[Any], Any]] = { + ("Dashboard", "json_metadata"): _normalize_dashboard_json_metadata, +} + + +def _normalize_for_compare(target: Any, col_name: str, value: Any) -> Any: + """Return *value* run through any per-class column normalizer registered + in ``_COLUMN_NORMALIZERS``, else *value* unchanged. + """ + normalizer = _COLUMN_NORMALIZERS.get((type(target).__name__, col_name)) + return normalizer(value) if normalizer is not None else value + + +def _has_dirty_versioned_children(target: Any, uow: Any) -> bool: + """Return ``True`` when *uow* contains an operation for a versioned + child of *target* (e.g. a ``TableColumn`` whose ``table`` is *target*). + + Used by :meth:`SkipUnmodifiedPlugin._is_no_op_update` so a parent + UPDATE that was force-flagged by + :func:`baseline.force_parent_dirty_on_child_change` is preserved + even though the parent's own scalars match the previous version. + """ + # pylint: disable=import-outside-toplevel + from superset.versioning.baseline import child_to_parent_registry + + child_map = child_to_parent_registry() + target_cls = type(target) + for _key, op in uow.operations.items(): + entry = child_map.get(type(op.target)) + if entry is None: + continue + parent_attr, parent_cls = entry + if parent_cls is not target_cls: + continue + parent = getattr(op.target, parent_attr, None) + if parent is target: + return True + return False + + +class VersionTransactionFactory(TransactionFactory): + """TransactionFactory that renames the transaction table and adds a bare + ``user_id`` integer column so the FlaskPlugin can record the acting user + without requiring a FK relationship to ``ab_user``. + + Continuum only adds ``user_id`` when ``user_cls`` is set on the manager. + We add it unconditionally (no FK) so that both the FlaskPlugin's + ``transaction_args()`` and our ``baseline.py`` direct inserts can record + which user triggered the version event. + """ + + def create_class(self, manager: Any) -> Any: + cls = super().create_class(manager) + cls.__table__.name = "version_transaction" + # Rename the PostgreSQL sequence for consistent naming. + for col in cls.__table__.columns: + if col.name == "id" and col.default is not None: + col.default.name = "version_transaction_id_seq" + # Add user_id INTEGER (no FK) for user tracking. The mapper has not + # been configured yet at this point, so append_column + add_property + # is safe here. + user_id_col = sa.Column("user_id", sa.Integer, nullable=True) + cls.__table__.append_column(user_id_col) + cls.__mapper__.add_property("user_id", sa_orm.column_property(user_id_col)) + # ``action_kind`` — high-level avenue that produced this commit + # (``restore`` / ``import`` / ``clone`` / ``NULL`` for ordinary + # saves). The DDL is in the consolidated Alembic migration; we + # also declare it on the SQLAlchemy Table here so consumers + # like ``superset.versioning.activity._select_change_rows_for_kinds`` + # can reference ``versioning_manager.transaction_cls.__table__ + # .c.action_kind`` at runtime, and so the change-record + # listener's ``sa.update()`` stamp emits the correctly-quoted + # identifier per dialect. + action_kind_col = sa.Column("action_kind", sa.String(32), nullable=True) + cls.__table__.append_column(action_kind_col) + cls.__mapper__.add_property( + "action_kind", sa_orm.column_property(action_kind_col) + ) + return cls + + +class VersioningFlaskPlugin(FlaskPlugin): + """FlaskPlugin subclass that uses Superset's :func:`get_user_id` (which + reads ``g.user``) instead of Flask-Login's ``current_user``. Superset's + JWT auth for API routes populates ``g.user`` but leaves + ``flask_login.current_user`` anonymous, so the upstream plugin would + record ``user_id=NULL`` on version_transaction rows created by API + calls. Returns an empty dict (so the transaction row is written + anyway) when no user is available — e.g. CLI, Celery, import/export. + """ + + def transaction_args(self, uow: Any, session: Any) -> dict[str, Any]: + # pylint: disable=import-outside-toplevel + from flask import has_request_context, request + + from superset.utils.core import get_user_id + + user_id = get_user_id() + if user_id is None: + return {} + + remote_addr: str | None + try: + remote_addr = request.remote_addr if has_request_context() else None + except RuntimeError: + remote_addr = None + + return {"user_id": user_id, "remote_addr": remote_addr} + + +class SkipUnmodifiedPlugin(Plugin): + """Skip creating version rows for UPDATE operations whose post-flush + column values are byte-identical to the previous live version row. + + Continuum creates a version row for every entity in ``session.dirty``, + including saves where the SQLAlchemy ORM marked a column dirty (because + Superset re-serialised ``json_metadata`` via ``json.dumps`` on the save + path, or AuditMixin auto-bumped ``changed_on``) but the resulting value + is unchanged from the previous version. Those rows pollute the version + history with no-op entries. + + ``is_modified()`` from Continuum is not enough: it consults SQLAlchemy's + attribute history, which is "did setattr produce a different value?", + not "did the final stored value change?". So we compare each + non-excluded versioned column on ``operation.target`` against the + previous live version row's value; if all are equal, the operation + is marked ``processed`` and Continuum skips it (see + ``UnitOfWork.create_version_objects``). + + The associated transaction is not removed; if every operation is a + no-op the transaction becomes an orphan in ``version_transaction`` + and is swept by the retention task at cutoff. Deleting the row + inline (in this hook) was considered and rejected: it would couple + this plugin to the change-records listener's buffer state — both + would have to agree that the flush produced nothing before we + could safely DROP the tx row, since ``version_changes.transaction_id`` + has an ON DELETE CASCADE FK that would silently drop any buffered + diff records the listener was about to insert. The orphan's storage + cost (~40 bytes/row) is small enough that the coordination isn't + worth it; retention handles the cleanup correctly by construction + (orphans have no parent shadow → they're never "preserved" by the + "preserve transactions whose shadow has the live row" rule and + age out with the rest of the history). + """ + + def before_create_version_objects(self, uow: Any, session: Any) -> None: + # ``uow.operations`` is a custom Continuum ``Operations`` collection; + # use its ``.items()`` method (not ``.values()``) to iterate. + # INSERTs always create a row (no prior to compare against); + # DELETEs can't be no-ops. Only UPDATE operations are candidates. + for _key, operation in uow.operations.items(): + if operation.processed or operation.type != Operation.UPDATE: + continue + try: + if self._is_no_op_update(operation.target, session, uow): + operation.processed = True + except Exception: # pylint: disable=broad-except + # Defensive — if introspection fails for any reason, fall + # back to creating the version row. + logger.exception( + "SkipUnmodifiedPlugin: skip-check raised for %s", + type(operation.target).__name__, + ) + + @classmethod + def _is_no_op_update(cls, target: Any, session: Any, uow: Any) -> bool: + """Return ``True`` when this UPDATE produces no observable change to + any non-excluded versioned column **and** no versioned children of + *target* are being modified in this flush. + + Stages: + + 1. If any versioned child (e.g. a ``TableColumn`` whose ``table`` + is *target*) has an operation in ``uow.operations``, the parent + is being force-touched by + ``baseline.force_parent_dirty_on_child_change`` to anchor the + child changes against a parent shadow row. Keep the row. + 2. ``is_modified(target)`` — cheap SQLAlchemy attribute-history + check. Returns ``False`` when only excluded columns/relationships + (``owners``, ``changed_on``, …) are dirty. This is the common + case (every save auto-bumps ``changed_on``); short-circuiting + here saves the DB round-trip in stage 3. + 3. Compare post-flush column values against the previous live + version row's stored values. Catches the case where SQLAlchemy + sees a column as dirty (e.g. ``set_dash_metadata`` re-serialised + ``json_metadata`` to a different byte sequence) but the + resulting parsed content matches the prior version. + """ + if _has_dirty_versioned_children(target, uow): + return False + if not is_modified(target): + return True + return cls._matches_previous_version(target, session) + + @staticmethod + def _matches_previous_version(target: Any, session: Any) -> bool: + """Return ``True`` when every non-excluded versioned column on + *target* matches the value stored in its previous live version row + (i.e., the row with ``end_transaction_id IS NULL``). + + Returns ``False`` for entities with no prior version row — letting + Continuum create the first one. In practice this case is rare: + ``register_baseline_listener`` (in ``superset.versioning.baseline``) + runs ahead of Continuum's ``before_flush`` and inserts a baseline + row for any entity being saved for the first time, so the second + save (and beyond) is what flows through this path. + """ + cls = type(target) + try: + ver_cls = version_class(cls) + except Exception: # pylint: disable=broad-except + return False + ver_table = ver_cls.__table__ + + col_keys = [prop.key for prop in versioned_column_properties(target)] + if not col_keys: + return False + + select_stmt = ( + sa.select(*[ver_table.c[c] for c in col_keys]) + .where(ver_table.c.id == target.id) + .where(ver_table.c.end_transaction_id.is_(None)) + .order_by(ver_table.c.transaction_id.desc()) + .limit(1) + ) + row = session.connection().execute(select_stmt).first() + if row is None: + return False # no previous version → let Continuum create one + + for col_name, prev_value in zip(col_keys, row, strict=False): + post = _normalize_for_compare( + target, col_name, getattr(target, col_name, None) + ) + pre = _normalize_for_compare(target, col_name, prev_value) + if post != pre: + return False + return True diff --git a/superset/versioning/queries.py b/superset/versioning/queries.py new file mode 100644 index 000000000000..747dd18d6849 --- /dev/null +++ b/superset/versioning/queries.py @@ -0,0 +1,512 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Read-side queries for the entity-versioning API. + +Pure-read helpers that translate Continuum shadow rows and +``version_changes`` records into the shapes the API endpoints return. +The corresponding write side (restore) lives in +:mod:`superset.versioning.restore`. The backward-compat ``VersionDAO`` +façade in :mod:`superset.daos.version` re-exports both. + +Also exposes the deterministic version-UUID derivation +(:data:`VERSION_UUID_NAMESPACE` + :func:`derive_version_uuid`) used by +both the read endpoints and the ETag emission path in +:mod:`superset.versioning.etag`. +""" + +from __future__ import annotations + +import uuid +from typing import Any +from uuid import UUID + +import sqlalchemy as sa +from flask_appbuilder import Model +from sqlalchemy_continuum import version_class + +from superset.extensions import db +from superset.versioning.baseline import CONTINUUM_BOOKKEEPING_COLUMNS + +# Fixed UUIDv5 namespace under which per-(entity, transaction) version UUIDs +# are derived. Never change this constant — changing it invalidates every +# version_uuid that clients may have cached, bookmarked, or stored. +VERSION_UUID_NAMESPACE = UUID("7a6f5d9b-4c3b-5d8e-9a1c-0e2b4c6d8f10") + +# Continuum's integer ``operation_type`` mapped to the string the API +# returns. Kept short and stable for downstream tooling consuming the +# raw response. Continuum guarantees 0/1/2; anything else is a Continuum +# version mismatch and surfaces as ``str(int)`` rather than crashing. +_OP_TYPE_LABELS: dict[int, str] = {0: "baseline", 1: "update", 2: "delete"} + + +def derive_version_uuid(entity_uuid: UUID, transaction_id: int) -> UUID: + """Derive a deterministic UUIDv5 identifying one version row. + + The UUID is a function of the owning entity's UUID and the Continuum + ``transaction_id`` of the version row, so it is stable across retention + pruning (which never changes ``transaction_id``) and portable across + replicas. It is not randomly generated — two Supersets with identical + ``(entity.uuid, transaction_id)`` will compute the same version_uuid. + """ + return uuid.uuid5(VERSION_UUID_NAMESPACE, f"{entity_uuid}:{transaction_id}") + + +def _resolve_version_tables( + model_cls: type[Model], +) -> tuple[sa.Table, sa.Table, sa.Table]: + """Return the (version, transaction, user) ``Table`` objects used by the + listing and snapshot queries. + + All three lookups happen inside this module on every read; centralising + the trio (a) keeps the imports in one place and (b) makes the join helper + below take a uniform signature. + """ + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import versioning_manager + + from superset import security_manager + + ver_tbl = version_class(model_cls).__table__ + tx_tbl = versioning_manager.transaction_cls.__table__ + user_tbl = security_manager.user_model.__table__ + return ver_tbl, tx_tbl, user_tbl + + +def _version_with_tx_user_join( + ver_tbl: sa.Table, tx_tbl: sa.Table, user_tbl: sa.Table +) -> Any: + """Build the version → transaction → user left-join used by both + :func:`list_versions` and :func:`get_version`. The user-side join is + a left-outer so saves with no Flask user context (CLI, Celery, import) + still surface in the result with ``changed_by = None``. + """ + return ver_tbl.join(tx_tbl, ver_tbl.c.transaction_id == tx_tbl.c.id).outerjoin( + user_tbl, tx_tbl.c.user_id == user_tbl.c.id + ) + + +def _baseline_first_ordering(ver_tbl: sa.Table) -> tuple[Any, ...]: + """Order ``(operation_type != 0).asc(), transaction_id.asc()`` so any + op=0 row — Continuum's INSERT or our synthetic baseline — sorts to + position 0 regardless of its transaction_id. A single entity never has + more than one op=0 row (Continuum tracks one creation per live entity; + our baseline listener only fires when no prior version rows exist), so + this gives a stable chronological order with the "original" version + always first. + """ + return ( + (ver_tbl.c.operation_type != 0).asc(), + ver_tbl.c.transaction_id.asc(), + ) + + +def _user_select_cols(user_tbl: sa.Table) -> list[Any]: + """Columns to select from ``user_tbl`` to build a ``changed_by`` dict. + Labels ``user_tbl.c.id`` as ``"user_id"`` so callers can read the row + by a stable key regardless of whether they also select the version + table's ``id`` column. + """ + return [ + user_tbl.c.id.label("user_id"), + user_tbl.c.username, + user_tbl.c.first_name, + user_tbl.c.last_name, + ] + + +def _changed_by_from_row(row: Any) -> dict[str, Any] | None: + """Project the user columns from a query row onto the API's + ``changed_by`` shape, or ``None`` for saves with no Flask user context + (CLI / Celery / import / unauthenticated). Expects the user columns to + have been selected via :func:`_user_select_cols` so the row keys are + ``user_id`` / ``username`` / ``first_name`` / ``last_name``. + """ + if row["user_id"] is None: + return None + return { + "id": row["user_id"], + "username": row["username"], + "first_name": row["first_name"], + "last_name": row["last_name"], + } + + +def _entity_kind_for(model_cls: type[Model]) -> str | None: + """Return the ``version_changes.entity_kind`` value for *model_cls*, or + ``None`` when the class isn't in the change-records taxonomy.""" + # pylint: disable=import-outside-toplevel + from superset.versioning.changes import ENTITY_KIND_BY_CLASS_NAME + + return ENTITY_KIND_BY_CLASS_NAME.get(model_cls.__name__) + + +def find_active_by_uuid(model_cls: type[Model], entity_uuid: UUID) -> Any | None: + """Return the live entity matching *entity_uuid*, or None if not found.""" + return ( + db.session.query(model_cls) + .filter(model_cls.uuid == entity_uuid) # type: ignore[attr-defined] + .one_or_none() + ) + + +def _get_version_count(model_cls: type[Model], entity_id: int) -> int: + """Return the number of historical version rows for *entity_id*.""" + ver_cls = version_class(model_cls) + return ( + db.session.query(sa.func.count()) + .select_from(ver_cls) + .filter(ver_cls.id == entity_id) + .scalar() + or 0 + ) + + +def current_version_number(model_cls: type[Model], entity_id: int) -> int | None: + """Return the 0-based ``version_number`` of the live row for *entity_id* + — equivalent to the index of the most recent entry that + :func:`list_versions` would return, or ``None`` when the entity has no + version rows yet. + + Note: this index is *unstable under retention pruning*. The scheduled + :func:`prune_old_versions` task drops shadow rows whose owning + ``version_transaction`` is older than + :envvar:`SUPERSET_VERSION_HISTORY_RETENTION_DAYS`, so the same integer + can refer to different rows before and after a prune cycle. Use + :func:`current_live_transaction_id` for a stable identifier. + """ + count = _get_version_count(model_cls, entity_id) + return count - 1 if count > 0 else None + + +def current_live_transaction_id(model_cls: type[Model], entity_id: int) -> int | None: + """Return the Continuum ``transaction_id`` of the live row for + *entity_id* — stable across retention pruning, unlike the index + returned by :func:`current_version_number`. + """ + ver_cls = version_class(model_cls) + row = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == entity_id) + .filter(ver_cls.end_transaction_id.is_(None)) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .first() + ) + return row[0] if row else None + + +def current_live_version_uuid( + model_cls: type[Model], entity_id: int, entity_uuid: UUID +) -> UUID | None: + """Return the deterministic ``version_uuid`` of the live row, or + ``None`` when the entity has no version rows yet.""" + tx_id = current_live_transaction_id(model_cls, entity_id) + if tx_id is None: + return None + return derive_version_uuid(entity_uuid, tx_id) + + +def list_change_records_batch( + entity_kind: str, + entity_id: int, + transaction_ids: list[int], +) -> dict[int, list[dict[str, Any]]]: + """Return ``version_changes`` rows keyed by ``transaction_id``. + + Batches the lookup across multiple transactions with a single + ``WHERE transaction_id IN (...) AND entity_kind = ? AND entity_id = ?`` + query so the list endpoint avoids N+1 round-trips. Rows are + distributed into per-tx lists sorted by ``sequence`` ascending + (matching the replay order the diff engine emits). Missing + transactions are represented by an empty list in the result so + callers can use ``result.get(tx_id, [])`` without guarding. + + If the ``version_changes`` table is missing (pre-migration or + freshly downgraded), returns an empty dict rather than propagating + the error — consistent with this being a descriptive layer that + should not break the list endpoint. + """ + # pylint: disable=import-outside-toplevel + from superset.versioning.changes import version_changes_table + + if not transaction_ids: + return {} + + try: + rows = ( + db.session.connection() + .execute( + sa.select( + version_changes_table.c.transaction_id, + version_changes_table.c.sequence, + version_changes_table.c.kind, + version_changes_table.c.path, + version_changes_table.c.from_value, + version_changes_table.c.to_value, + ) + .where( + version_changes_table.c.entity_kind == entity_kind, + version_changes_table.c.entity_id == entity_id, + version_changes_table.c.transaction_id.in_(transaction_ids), + ) + .order_by( + version_changes_table.c.transaction_id.asc(), + version_changes_table.c.sequence.asc(), + ) + ) + .mappings() + .all() + ) + except sa.exc.OperationalError: + return {} + + grouped: dict[int, list[dict[str, Any]]] = {tx: [] for tx in transaction_ids} + for row in rows: + grouped[row["transaction_id"]].append( + { + "kind": row["kind"], + "path": row["path"], + "from_value": row["from_value"], + "to_value": row["to_value"], + } + ) + return grouped + + +def list_versions( + model_cls: type[Model], + entity_uuid: UUID, + *, + entity: Any | None = None, +) -> list[dict[str, Any]] | None: + """Return the version history for the entity identified by *entity_uuid*. + + Returns ``None`` when no active entity matches the UUID — callers should + translate that into a 404. Returns an empty list when the entity exists + but has no version rows yet (pre-migration, or never edited). + + The list is ordered by ``transaction_id`` ascending and each entry is + assigned a 0-based sequential ``version_number``. ``operation_type`` is + mapped from Continuum's integer constants to a string (``0`` → baseline, + ``1`` → update, ``2`` → delete). ``changed_by`` is the User row keyed + off ``version_transaction.user_id``, or ``None`` when the save had no + Flask user context (CLI, import, etc.). + + Pass *entity* to skip the ``find_active_by_uuid`` lookup when the + caller has already resolved the entity (API handlers do this to enforce + ``raise_for_ownership`` before calling here). The skip saves one + ``WHERE uuid = ?`` query — that lookup isn't identity-map-cacheable + because ``uuid`` is a unique non-PK column. + """ + if entity is None: + entity = find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + return None + + ver_tbl, tx_tbl, user_tbl = _resolve_version_tables(model_cls) + stmt = ( + sa.select( + ver_tbl.c.transaction_id, + ver_tbl.c.operation_type, + tx_tbl.c.issued_at, + *_user_select_cols(user_tbl), + ) + .select_from(_version_with_tx_user_join(ver_tbl, tx_tbl, user_tbl)) + .where(ver_tbl.c.id == entity.id) + .order_by(*_baseline_first_ordering(ver_tbl)) + ) + rows = db.session.execute(stmt).mappings().all() + + # Batch-load change records for every listed transaction in one query + # (T050). ``entity_kind`` is derived from the model class so the API + # filter ``WHERE entity_kind = 'chart' AND entity_id = ?`` can be + # precise when multiple versioned entities share a flush. + changes_by_tx: dict[int, list[dict[str, Any]]] = {} + if (entity_kind := _entity_kind_for(model_cls)) is not None: + tx_ids = [row["transaction_id"] for row in rows] + changes_by_tx = list_change_records_batch(entity_kind, entity.id, tx_ids) + + return [ + { + "version_uuid": derive_version_uuid(entity_uuid, row["transaction_id"]), + "version_number": version_number, + "transaction_id": row["transaction_id"], + "operation_type": _OP_TYPE_LABELS.get( + row["operation_type"], str(row["operation_type"]) + ), + "issued_at": row["issued_at"], + "changed_by": _changed_by_from_row(row), + "changes": changes_by_tx.get(row["transaction_id"], []), + } + for version_number, row in enumerate(rows) + ] + + +def resolve_version_uuid( + model_cls: type[Model], + entity_uuid: UUID, + version_uuid: UUID, + *, + entity: Any | None = None, +) -> int | None: + """Translate a ``version_uuid`` into the 0-based ``version_number`` that + :func:`superset.versioning.restore.restore_version` accepts, or ``None`` + when the UUID does not match any version row of the given entity. + + Ordering matches :func:`list_versions` — op=0 rows first, then by + transaction_id — so the version_number returned here is the same index + a client would see in the list response. + + Implementation note: the loop re-derives ``version_uuid`` per + transaction in Python because there's no portable SQL form for a + UUIDv5 derivation across PostgreSQL / MySQL / SQLite (Postgres has + ``uuid_generate_v5``; the other two do not). The iteration count is + bounded by ``SUPERSET_VERSION_HISTORY_RETENTION_DAYS`` worth of + edits — the retention task ages older shadow rows out — so the + practical N is at most a few hundred. If retention is ever + disabled (``= 0``) on a heavily-edited entity, this loop is the + place to revisit. + + Pass *entity* to skip the ``find_active_by_uuid`` lookup; see + :func:`list_versions` for the rationale. + """ + if entity is None: + entity = find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + return None + + ver_cls = version_class(model_cls) + tx_ids = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == entity.id) + .order_by( + (ver_cls.operation_type != 0).asc(), + ver_cls.transaction_id.asc(), + ) + .all() + ) + for version_number, (tx_id,) in enumerate(tx_ids): + if derive_version_uuid(entity_uuid, tx_id) == version_uuid: + return version_number + return None + + +def get_version( + model_cls: type[Model], + entity_uuid: UUID, + version_uuid: UUID, + *, + entity: Any | None = None, +) -> dict[str, Any] | None: + """Return the entity's state at the specified version as a dict. + + Read-only — nothing in the live database is modified. The returned + shape is intended to mirror a regular single-entity GET response + (scalar columns plus restored ``columns`` / ``metrics`` lists for + ``SqlaTable``), with a ``_version`` key holding the version-level + metadata (uuid, transaction_id, operation_type, issued_at, + changed_by) so callers can tell which version they're looking at. + + Returns ``None`` when either *entity_uuid* or *version_uuid* does not + match — callers should translate to 404. + + Pass *entity* to skip the ``find_active_by_uuid`` lookup; see + :func:`list_versions` for the rationale. The same *entity* is threaded + into :func:`resolve_version_uuid` to eliminate a second redundant + lookup on the same request. + """ + # pylint: disable=import-outside-toplevel + from superset.connectors.sqla.models import SqlaTable + + if entity is None: + entity = find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + return None + + version_num = resolve_version_uuid( + model_cls, entity_uuid, version_uuid, entity=entity + ) + if version_num is None: + return None + + ver_tbl, tx_tbl, user_tbl = _resolve_version_tables(model_cls) + stmt = ( + sa.select( + ver_tbl, + tx_tbl.c.issued_at, + *_user_select_cols(user_tbl), + ) + .select_from(_version_with_tx_user_join(ver_tbl, tx_tbl, user_tbl)) + .where(ver_tbl.c.id == entity.id) + .order_by(*_baseline_first_ordering(ver_tbl)) + .offset(version_num) + .limit(1) + ) + row = db.session.execute(stmt).mappings().first() + if row is None: + return None + + # Project the entity's own scalar fields, skipping versioning + # metadata columns. + result: dict[str, Any] = {} + for col in ver_tbl.columns: + if col.name in CONTINUUM_BOOKKEEPING_COLUMNS: + continue + value = row[col.name] + # uuid columns come back as UUID instances; make them JSON-safe. + if isinstance(value, UUID): + value = str(value) + result[col.name] = value + + changes: list[dict[str, Any]] = [] + if (entity_kind := _entity_kind_for(model_cls)) is not None: + changes = list_change_records_batch( + entity_kind, entity.id, [row["transaction_id"]] + ).get(row["transaction_id"], []) + + result["_version"] = { + "version_uuid": str(version_uuid), + "version_number": version_num, + "transaction_id": row["transaction_id"], + "operation_type": _OP_TYPE_LABELS.get( + row["operation_type"], str(row["operation_type"]) + ), + "issued_at": row["issued_at"], + "changed_by": _changed_by_from_row(row), + "changes": changes, + } + + # For datasets, attach the columns/metrics as they were at this + # transaction by reading from Continuum's child shadow tables + # (``table_columns_version`` / ``sql_metrics_version``). Empty lists + # when the dataset had no children at this tx. + if model_cls is SqlaTable: + # pylint: disable=import-outside-toplevel + from superset.connectors.sqla.models import SqlMetric, TableColumn + from superset.versioning.changes import shadow_rows_valid_at + + target_tx = row["transaction_id"] + cols_tbl = version_class(TableColumn).__table__ + metrics_tbl = version_class(SqlMetric).__table__ + result["columns"] = shadow_rows_valid_at( + db.session, cols_tbl, "table_id", entity.id, target_tx + ) + result["metrics"] = shadow_rows_valid_at( + db.session, metrics_tbl, "table_id", entity.id, target_tx + ) + + return result diff --git a/superset/versioning/restore.py b/superset/versioning/restore.py new file mode 100644 index 000000000000..01d18ebfd7de --- /dev/null +++ b/superset/versioning/restore.py @@ -0,0 +1,150 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Write-side: restore a versioned entity to an earlier state. + +Companion to :mod:`superset.versioning.queries`. The +``BaseRestoreVersionCommand`` in :mod:`superset.commands.version_restore` +is the only intended caller; the backward-compat ``VersionDAO`` façade +in :mod:`superset.daos.version` re-exports ``restore_version`` for +existing call sites. +""" + +from __future__ import annotations + +import logging +from datetime import datetime +from typing import Any +from uuid import UUID + +from flask_appbuilder import Model +from sqlalchemy_continuum import version_class + +from superset.extensions import db +from superset.utils.core import get_user_id +from superset.versioning.queries import find_active_by_uuid +from superset.versioning.utils import single_flush_scope + +logger = logging.getLogger(__name__) + + +# Per-model relationships that Continuum's Reverter recurses into during a +# restore. Each restore replays the listed relationships from the version- +# side shadow tables onto the live entity. Children versioned through +# Continuum (``TableColumn`` / ``SqlMetric`` on ``SqlaTable``; +# ``dashboard_slices`` M2M on ``Dashboard``) come back automatically; +# ``Slice`` has no child collections to recurse into so its list is empty. +_RESTORE_RELATIONS: dict[str, list[str]] = { + "SqlaTable": ["columns", "metrics"], + "Dashboard": ["slices"], + "Slice": [], +} + + +def restore_version( + model_cls: type[Model], + entity_uuid: UUID, + version_num: int, +) -> Any | None: + """Restore the entity identified by *entity_uuid* to the state captured + by *version_num* (0-based, as returned by + :func:`superset.versioning.queries.list_versions`). + + Returns the live entity after the restore, or ``None`` when either the + UUID does not match an active entity or ``version_num`` is out of + range — callers should translate both to a 404. + + Uses SQLAlchemy-Continuum's native ``version_obj.revert(relations=...)`` + and delegates commit to the caller (expected to be a command decorated + with ``@transaction()``). The ``relations`` list depends on the model + type and is looked up in :data:`_RESTORE_RELATIONS`. + + After the revert, ``changed_on`` / ``changed_by_fk`` are re-stamped + with the current time and the restoring user's id (see + :func:`_stamp_audit_fields_for_restore`) so the new version row + produced by the restoring commit reflects who clicked Restore, not + the original author. ``created_on`` / ``created_by_fk`` are left + alone. + """ + entity = find_active_by_uuid(model_cls, entity_uuid) + if entity is None: + return None + + ver_cls = version_class(model_cls) + + # version_num is a 0-based positional index, matching what + # ``list_versions`` emits. Ordering keeps op=0 rows first so position 0 + # is always the baseline/INSERT. + target_version = ( + db.session.query(ver_cls) + .filter(ver_cls.id == entity.id) + .order_by( + (ver_cls.operation_type != 0).asc(), + ver_cls.transaction_id.asc(), + ) + .offset(version_num) + .limit(1) + .first() + ) + if target_version is None: + return None + + # Run the whole multi-relationship revert inside a single flush scope + # so SQLAlchemy-Continuum's ``Reverter`` can iterate relations without + # tripping its autoflush race, and so the change-records listener sees + # the complete shadow state in one ``after_flush`` pass. See + # ``single_flush_scope`` for the full rationale. + relations = _RESTORE_RELATIONS.get(model_cls.__name__, []) + try: + with single_flush_scope(db.session): + target_version.revert(relations=relations) + except Exception: + logger.exception( + "Continuum revert() failed for %s id=%s tx=%s relations=%s", + model_cls.__name__, + entity.id, + target_version.transaction_id, + relations, + ) + raise + + _stamp_audit_fields_for_restore(entity) + return entity + + +def _stamp_audit_fields_for_restore(entity: Any) -> None: + """Overwrite ``changed_on`` / ``changed_by_fk`` on *entity* with the + current time and current user id, so that the restore is attributed + to the restoring user rather than the version snapshot's original + author. + + Charts additionally carry ``last_saved_at`` / ``last_saved_by_fk`` + columns (stamped by ``UpdateChartCommand`` on ordinary saves and + surfaced in the Charts list page's "last edited by" column). Without + overwriting these, the chart list still shows the snapshot's + original author after a restore, contradicting the user-visible + timeline. + """ + now = datetime.utcnow() + user_id = get_user_id() + if hasattr(entity, "changed_on"): + entity.changed_on = now + if hasattr(entity, "changed_by_fk"): + entity.changed_by_fk = user_id + if hasattr(entity, "last_saved_at"): + entity.last_saved_at = now + if hasattr(entity, "last_saved_by_fk"): + entity.last_saved_by_fk = user_id diff --git a/superset/versioning/schemas.py b/superset/versioning/schemas.py new file mode 100644 index 000000000000..4c5cccd6b30b --- /dev/null +++ b/superset/versioning/schemas.py @@ -0,0 +1,411 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Shared Marshmallow schemas for entity version history endpoints. + +Consumed by ChartRestApi, DashboardRestApi, and DatasetRestApi — the response +shape is identical across all three resources, so the schemas live here to +avoid triplicated definitions. +""" + +from __future__ import annotations + +from marshmallow import fields, Schema, validate + +from superset.versioning.changes import ACTION_KINDS + + +class VersionChangedBySchema(Schema): + """Subset of the User model included in each version history entry.""" + + id = fields.Integer() + username = fields.String() + first_name = fields.String() + last_name = fields.String() + + +class VersionChangeRecordSchema(Schema): + """One field-level diff hunk from ``version_changes``. + + The frontend renders human-readable prose from (``kind``, + ``from_value``, ``to_value``) via Flask-Babel. Server-side the + shape is deliberately machine-readable only — see spec FR-019. + """ + + kind = fields.String( + metadata={ + "description": ( + "Semantic category of the change. First-class values in V1: " + "'filter', 'metric', 'dimension', 'column', 'chart', " + "'time_range', 'color_palette'. Falls back to 'field' for " + "generic scalar changes that don't map to a named kind." + ) + }, + ) + path = fields.Raw( + metadata={ + "description": ( + "Array of segments locating the change in the entity's state. " + "Example: ['params', 'adhoc_filters', 'country']." + ) + }, + ) + from_value = fields.Raw( + allow_none=True, + metadata={ + "description": ( + "Value at path before the save; null when the field did not exist." + ), + }, + ) + to_value = fields.Raw( + allow_none=True, + metadata={ + "description": ( + "Value at path after the save; null when the field was removed." + ), + }, + ) + + +class VersionListItemSchema(Schema): + """A single version row in the version history response.""" + + version_number = fields.Integer( + metadata={"description": "0-based position in the history, oldest first"}, + ) + transaction_id = fields.Integer( + metadata={"description": "Underlying Continuum transaction id"}, + ) + operation_type = fields.String( + metadata={ + "description": ( + "One of 'baseline', 'update', or 'delete', derived from the " + "Continuum integer constant. Restore is not a distinct " + "operation_type: a restore surfaces as 'update' carrying " + "``action_kind='restore'`` (see ACTIVITY_ACTION_KINDS)." + ) + }, + ) + issued_at = fields.DateTime( + metadata={"description": "UTC timestamp of the commit that produced the row"}, + ) + changed_by = fields.Nested( + VersionChangedBySchema, + allow_none=True, + metadata={ + "description": ( + "User who produced the version, or null when the commit had no " + "authenticated Flask user (CLI, Celery, import)." + ) + }, + ) + changes = fields.List( + fields.Nested(VersionChangeRecordSchema), + metadata={ + "description": ( + "Structured diff records describing the atomic field-level " + "changes at this version, ordered by emission sequence. " + "Empty for baseline (op=0) transactions per spec M4." + ) + }, + ) + + +class VersionListResponseSchema(Schema): + """Envelope for version list responses.""" + + result = fields.List(fields.Nested(VersionListItemSchema)) + count = fields.Integer() + + +# ---- Cross-entity activity view (sc-107283) ------------------------------- + +#: Allowed values for ``ActivityRecordSchema.entity_kind``. User-facing +#: lowercase strings; the activity layer's internal kind dispatch keys off +#: ``model_cls.__name__`` (``Dashboard`` / ``Slice`` / ``SqlaTable``) and +#: translates to these labels at the JSON boundary in +#: :func:`superset.versioning.activity.apply_record_decoration`. +ACTIVITY_ENTITY_KINDS: tuple[str, ...] = ("dashboard", "chart", "dataset") + +#: Allowed values for ``ActivityRecordSchema.source`` (spec AV-013). +ACTIVITY_SOURCES: tuple[str, ...] = ("self", "related") + +#: Allowed values for ``ActivityRecordSchema.entity_deletion_state``. +#: Hard-delete is communicated separately via ``entity_deleted=true``; +#: the remaining state is the soft-delete sentinel (sc-103157). +ACTIVITY_DELETION_STATES: tuple[str, ...] = ("soft_deleted",) + +#: Allowed values for ``ActivityRecordSchema.kind`` — mirrors the +#: change-record taxonomy from sc-103156 FR-016. ``"field"`` is the +#: fallback for scalar changes without a more specific category. +#: +#: ``"restore"`` (previously the synthetic kind for restore events) is +#: removed: restores now produce regular field-level records plus +#: ``version_transaction.action_kind="restore"`` (see ACTIVITY_ACTION_KINDS). +ACTIVITY_CHANGE_KINDS: tuple[str, ...] = ( + "filter", + "metric", + "dimension", + "column", + "chart", + "row", + "tab", + "tabs", + "header", + "markdown", + "divider", + "time_range", + "color_palette", + "field", +) + +#: Allowed values for ``ActivityRecordSchema.operation`` — the per-record +#: verb. ``move`` only fires for layout records; ``add`` / ``remove`` / +#: ``edit`` apply across every emit site. +ACTIVITY_CHANGE_OPERATIONS: tuple[str, ...] = ( + "add", + "remove", + "move", + "edit", +) + +#: Allowed values for ``ActivityRecordSchema.action_kind`` — the +#: transaction-level avenue. ``null`` (omitted from the enum, signalled +#: by ``allow_none``) means "ordinary save". Sourced from +#: :data:`superset.versioning.changes.ACTION_KINDS` so a future +#: addition (e.g. ``"thumbnail_warm"``) only has to update that +#: constant; the schema picks it up automatically. +ACTIVITY_ACTION_KINDS: tuple[str, ...] = tuple(sorted(ACTION_KINDS)) + + +class ActivityChangedBySchema(Schema): + """User attribution for an activity record. + + The activity-view payload exposes only the display fields + (``id`` + given/family name); ``username`` is omitted by design (see + data-model.md §"ActivityRecord DTO"). ``null`` when the saving user + has been deleted from ``ab_user`` (sc-103156 §Session 2026-05-18 + clarification). + """ + + id = fields.Integer() + first_name = fields.String() + last_name = fields.String() + + +class ActivityImpactSchema(Schema): + """Dependent-count summary attached to ``source='related'`` records. + + Synthesized server-side at the time of the activity query — it counts + siblings affected by the same upstream change at the same transaction + (e.g., how many charts on the requested dashboard pointed at the + dataset whose edit this record represents). + """ + + charts = fields.Integer( + metadata={ + "description": ( + "Number of sibling charts on the path entity affected by " + "the same related-record change at this transaction." + ) + }, + ) + + +class ActivityRecordSchema(Schema): + """One change record in the activity stream. + + One record per atomic field-level change. Fields mirror + data-model.md §"``ActivityRecord`` DTO" — see that doc for source + and required/optional details. + """ + + version_uuid = fields.String( + metadata={ + "description": ( + "Stable UUIDv5 identifier for the source version " + "(``derive_version_uuid(entity_uuid, transaction_id)``). " + "Identical to what ``/versions//`` would " + "return for the same change." + ) + }, + ) + entity_kind = fields.String( + validate=validate.OneOf(ACTIVITY_ENTITY_KINDS), + metadata={ + "description": ( + "User-facing kind of the source entity: one of " + '``"dashboard"`` / ``"chart"`` / ``"dataset"``.' + ) + }, + ) + entity_uuid = fields.String( + allow_none=True, + metadata={ + "description": ( + "UUID of the source entity; ``null`` only when " + "``entity_deleted: true`` (the entity has been hard-deleted " + "since the change was recorded)." + ) + }, + ) + entity_name = fields.String( + metadata={ + "description": ( + "Name of the source entity *at the time of the change* — " + "denormalized from the validity-strategy shadow row. " + "Survives entity rename / delete." + ) + }, + ) + entity_deleted = fields.Boolean( + metadata={ + "description": ( + "True iff the source entity is hard-deleted " + "(no live row by ``entity_id``). False for live and " + "soft-deleted entities." + ) + }, + ) + entity_deletion_state = fields.String( + allow_none=True, + validate=validate.OneOf(ACTIVITY_DELETION_STATES), + metadata={ + "description": ( + "Present when the source entity has non-null ``deleted_at`` " + "(sc-103157). Absent or ``null`` otherwise." + ) + }, + ) + source = fields.String( + validate=validate.OneOf(ACTIVITY_SOURCES), + metadata={ + "description": ( + '``"self"`` if ``(entity_kind, entity_id)`` matches the ' + 'path entity; else ``"related"``. Drives the frontend\'s ' + "no-group-under-save rendering rule (AV-013)." + ) + }, + ) + transaction_id = fields.Integer( + metadata={"description": "Stable secondary ordering key; never reused."}, + ) + issued_at = fields.DateTime( + metadata={"description": "UTC timestamp; primary ordering key (DESC)."}, + ) + changed_by = fields.Nested( + ActivityChangedBySchema, + allow_none=True, + metadata={ + "description": ( + "User who produced the change, or ``null`` when the saving " + "user no longer exists in ``ab_user``." + ) + }, + ) + kind = fields.String( + validate=validate.OneOf(ACTIVITY_CHANGE_KINDS), + metadata={ + "description": ( + "Content category — what kind of thing changed. " + "``field`` is the fallback for scalar changes without a " + "more specific category. Per-record." + ) + }, + ) + operation = fields.String( + validate=validate.OneOf(ACTIVITY_CHANGE_OPERATIONS), + metadata={ + "description": ( + "Per-record verb: ``add`` / ``remove`` / ``move`` / " + "``edit``. Explicit instead of inferred from " + "``from_value`` / ``to_value`` null-tests. ``move`` only " + "fires for layout records." + ) + }, + ) + action_kind = fields.String( + validate=validate.OneOf(ACTIVITY_ACTION_KINDS), + allow_none=True, + metadata={ + "description": ( + "Transaction-level avenue that produced this record's " + "batch: ``restore`` / ``import`` / ``clone``. ``null`` " + "for ordinary saves. All records sharing a " + "``transaction_id`` share the same action_kind. The " + "schema's third ``*_kind`` column (entity_kind / kind / " + "action_kind), at transaction scope." + ) + }, + ) + path = fields.List( + fields.String(), + metadata={ + "description": ( + "Pure navigation address — no verb or kind embedded. " + "Examples: ``['slice_name']``, ``['params', " + "'adhoc_filters', 'country']``, ``['CHART-x']`` for a " + "layout add/remove/move, ``['HEADER-y', 'text']`` for a " + "layout edit leaf. The verb lives in ``operation``, the " + "element type in ``kind``." + ) + }, + ) + from_value = fields.Raw( + allow_none=True, + metadata={"description": "Prior value; ``null`` = didn't exist."}, + ) + to_value = fields.Raw( + allow_none=True, + metadata={"description": "New value; ``null`` = removed."}, + ) + summary = fields.String( + metadata={ + "description": ( + 'Synthesized headline for ``source: "related"`` records — ' + 'e.g., ``"Dataset updated: Sales Transactions"`` ' + '(AV-012). Absent for ``source: "self"`` records.' + ) + }, + ) + impact = fields.Nested( + ActivityImpactSchema, + allow_none=True, + metadata={ + "description": ( + 'Optional dependent-count for ``source: "related"`` ' + 'records — e.g., ``{"charts": 4}`` for a dataset edit ' + "that affected 4 charts on the path dashboard at the " + 'change\'s transaction. Absent for ``source: "self"`` ' + "records and for related records without dependents." + ) + }, + ) + + +class ActivityResponseSchema(Schema): + """Envelope for activity-view responses.""" + + result = fields.List(fields.Nested(ActivityRecordSchema)) + count = fields.Integer( + metadata={ + "description": ( + "Total record count across all pages (the filtered + " + "denormalized stream), not just the current page." + ) + }, + ) diff --git a/superset/versioning/utils.py b/superset/versioning/utils.py new file mode 100644 index 000000000000..e09f133bf1dd --- /dev/null +++ b/superset/versioning/utils.py @@ -0,0 +1,81 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Shared session helpers used by the entity-versioning machinery.""" + +from __future__ import annotations + +from collections.abc import Iterator +from contextlib import contextmanager +from typing import Any + +import sqlalchemy as sa +from sqlalchemy.orm import Session + + +@contextmanager +def single_flush_scope(session: Session) -> Iterator[None]: + """Suppress autoflushes inside the block, flush once on clean exit. + + Intended for operations that (a) make multiple mutations across + relationships and (b) issue intermediate queries which would + otherwise autoflush. Iterating from one relationship to another + inside SQLAlchemy-Continuum's ``Reverter`` is the canonical case: + a mid-iteration autoflush transitions pending DELETEs to + ``state.deleted=True``, and the subsequent + ``session.add(version_parent)`` cascade walk trips on the + deleted-state instances with ``InvalidRequestError``. Wrapping the + whole revert keeps marked-for-deletion instances in + ``state.persistent`` until the trailing flush drains DELETEs + + INSERTs in one atomic step. That single flush is also load-bearing + for the ``after_flush`` change-records listener — splitting the + work across multiple flushes would split it across multiple + Continuum transactions, and the listener's tx-dedup guard would + silently drop the second pass's records. + + On exception, the trailing flush is skipped — the session's normal + rollback flow handles cleanup, and flushing a partially-mutated + state would be wrong. + """ + with session.no_autoflush: + yield + session.flush() + + +def read_row_outside_flush( + session: Session, table: sa.Table, entity_id: int +) -> dict[str, Any] | None: + """Read the row with ``id == entity_id`` from *table* without triggering + an autoflush. Returns the row as a plain dict, or ``None`` when no row + matches. + + The companion read primitive to :func:`single_flush_scope`. Listeners + that need pre-flush state (the row as it existed *before* the in-flight + edit was staged) use this — without ``no_autoflush``, the + ``session.connection().execute(...)`` would itself trigger a flush of + the pending edit, leaving "pre" and "post" indistinguishable. + + Returns ``dict[str, Any]`` rather than ``RowMapping`` so callers don't + accidentally hold a cursor-bound object past the listener boundary. + """ + with session.no_autoflush: + result = ( + session.connection() + .execute(sa.select(table).where(table.c.id == entity_id)) + .mappings() + .one_or_none() + ) + return dict(result) if result else None diff --git a/superset/views/activity_debug.py b/superset/views/activity_debug.py new file mode 100644 index 000000000000..589a1e3c88b4 --- /dev/null +++ b/superset/views/activity_debug.py @@ -0,0 +1,51 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +# Throwaway: sc-107283 activity-view debug UI. This Flask view exists +# only to serve the React shell on a fresh page-load of +# /activity-debug//. Without it, Flask returns its +# API-style 404 because the SPA doesn't have a true catch-all — each +# React route needs a corresponding render_app_template call. Delete +# this file + the AppBuilder registration in +# ``superset.initialization`` when the activity-view feature ships +# (or when you no longer need the debug UI). + +from flask_appbuilder import expose + +from superset.superset_typing import FlaskResponse +from superset.views.base import BaseSupersetView + + +class ActivityDebugView(BaseSupersetView): + """Serves the React shell for the throwaway activity-view debug page. + + No auth decorator on the shell itself — the shell page exposes no + data of its own. The React component renders inside it and fires + calls to ``/api/v1/{resource}/{uuid}/activity/`` which gate access + via ``raise_for_ownership`` on the path entity. Anonymous users + who somehow land here will see the React UI and the API errors + surface inline as "error: 401 ...". That's a fine UX for a debug + tool — and avoids the FAB ``@has_access`` redirect-to-home + behavior that masked real failures earlier. + """ + + route_base = "/activity-debug" + + @expose("///") + @expose("//") + def show(self, resource: str, uuid: str) -> FlaskResponse: # noqa: ARG002 + return super().render_app_template() diff --git a/tests/integration_tests/charts/version_history_tests.py b/tests/integration_tests/charts/version_history_tests.py new file mode 100644 index 000000000000..0f46b2e589cf --- /dev/null +++ b/tests/integration_tests/charts/version_history_tests.py @@ -0,0 +1,634 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Integration tests for chart (Slice) version history capture. + +T014 — chart version capture +T017 — baseline row capture +T018 (partial) — retention pruning (chart side) +T026 — chart version list endpoint +""" + +from __future__ import annotations + +from typing import Any + +import pytest +from sqlalchemy_continuum import version_class + +from superset.extensions import db +from superset.models.slice import Slice +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME, ALPHA_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _get_version_rows(chart: Slice) -> list[Any]: + ver_cls = version_class(Slice) + return ( + db.session.query(ver_cls) + .filter(ver_cls.id == chart.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + + +def _persist_fixture_state() -> None: + """Force fixture's pending INSERTs to commit in their own transaction. + + The birth_names fixture stages charts and the dashboard via session.add() + but does not commit. Without this, the test's first commit batches the + INSERTs and UPDATEs into the same Continuum transaction, causing the + existing version row to be updated in place instead of a new one being + created. + """ + db.session.commit() + + +class TestChartVersionCapture(SupersetTestCase): + """T014 — version rows are created on save; no spurious extra rows.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def test_single_save_creates_one_version_row(self) -> None: + """Saving a chart for the first time creates exactly one version row.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + + # Trigger a save (update a scalar field) + original_name = chart.slice_name + chart_id = chart.id + + try: + chart.slice_name = "Girls (edited)" + db.session.commit() + + rows = _get_version_rows(chart) + # Two rows: baseline (operation_type=0) + edit (operation_type=1) + assert len(rows) == 2, f"Expected 2 version rows, got {len(rows)}" + assert rows[0].operation_type == 0 # baseline + assert rows[1].operation_type == 1 # update + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_two_saves_create_exactly_two_version_rows_after_baseline(self) -> None: + """Second save adds exactly one more version row (no duplicate rows).""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + + original_name = chart.slice_name + chart_id = chart.id + + try: + chart.slice_name = "Boys v1" + db.session.commit() + rows_after_first = _get_version_rows(chart) + # baseline + v1 = 2 rows + assert len(rows_after_first) == 2 + + chart.slice_name = "Boys v2" + db.session.commit() + rows_after_second = _get_version_rows(chart) + # baseline + v1 + v2 = 3 rows + assert len(rows_after_second) == 3 + assert rows_after_second[-1].slice_name == "Boys v2" + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + +class TestChartBaselineCapture(SupersetTestCase): + """T017 — the baseline listener inserts a pre-edit snapshot row (operation_type=0).""" # noqa: E501 + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def test_baseline_row_has_pre_edit_state(self) -> None: + """The baseline row captures the field value *before* the first edit.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice) + .filter(Slice.slice_name == "Top 10 Girl Name Share") + .first() + ) + assert chart is not None + + pre_edit_name = chart.slice_name + chart_id = chart.id + + try: + chart.slice_name = "Top 10 Girl Name Share (baseline test)" + db.session.commit() + + rows = _get_version_rows(chart) + assert rows[0].operation_type == 0 # baseline row + assert rows[0].slice_name == pre_edit_name # pre-edit name preserved + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = pre_edit_name + db.session.commit() + + def test_baseline_row_is_at_position_zero_for_preexisting_entity(self) -> None: + """When an entity has zero Continuum history (e.g. created before + versioning was enabled), our baseline listener must produce a row + that sorts to version_number 0 — i.e. its transaction_id must be + strictly less than the UPDATE row Continuum writes in the same + commit.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Participants").first() + ) + assert chart is not None + chart_id = chart.id + original_name = chart.slice_name + + try: + # Wipe this chart's Continuum history so our baseline listener has + # count==0 on the next save — simulating a pre-existing entity. + ver_cls = version_class(Slice) + db.session.query(ver_cls).filter(ver_cls.id == chart_id).delete( + synchronize_session=False + ) + db.session.commit() + + chart.slice_name = "Participants (preexisting baseline test)" + db.session.commit() + + rows = _get_version_rows(chart) + pairs = [(r.operation_type, r.transaction_id) for r in rows] + assert len(rows) == 2, f"Expected baseline + update; got {pairs}" + assert rows[0].operation_type == 0, ( + f"Position 0 should be the baseline (op=0); got " + f"op={rows[0].operation_type} at tx={rows[0].transaction_id}" + ) + assert rows[0].slice_name == original_name, ( + "The baseline row must carry the pre-edit slice_name" + ) + assert rows[0].transaction_id < rows[1].transaction_id, ( + "Baseline's transaction_id must be less than the update's so it " + "sorts to position 0" + ) + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_no_duplicate_baseline_on_subsequent_saves(self) -> None: + """Subsequent saves do NOT add a second baseline row.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice) + .filter(Slice.slice_name == "Top 10 Boy Name Share") + .first() + ) + assert chart is not None + original_name = chart.slice_name + chart_id = chart.id + + try: + chart.slice_name = "Top 10 Boy Name Share v1" + db.session.commit() + + chart.slice_name = "Top 10 Boy Name Share v2" + db.session.commit() + + baseline_rows = [ + r for r in _get_version_rows(chart) if r.operation_type == 0 + ] + assert len(baseline_rows) == 1, "Should have exactly one baseline row" + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + +class TestChartVersionListApi(SupersetTestCase): + """T026 — GET /api/v1/chart//versions/ endpoint.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _list_versions(self, chart_uuid: str) -> Any: + return self.client.get(f"/api/v1/chart/{chart_uuid}/versions/") + + def test_list_versions_returns_ordered_sequence(self) -> None: + """Three saves produce three rows in ascending version_number order.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + original_name = chart.slice_name + chart_uuid = str(chart.uuid) + chart_id = chart.id + + try: + for i in range(3): + chart.slice_name = f"Girls v{i}" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._list_versions(chart_uuid) + assert rv.status_code == 200 + + body = _json.loads(rv.data.decode("utf-8")) + # Baseline + three updates = 4 rows; we only need to check the last 3 + # are the updates we just made in order. + assert body["count"] == len(body["result"]) + assert len(body["result"]) >= 3 + for idx, entry in enumerate(body["result"]): + assert entry["version_number"] == idx + assert entry["issued_at"] is not None + # Timestamps are monotonically non-decreasing. + timestamps = [e["issued_at"] for e in body["result"]] + assert timestamps == sorted(timestamps) + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_list_versions_empty_for_untouched_entity(self) -> None: + """A chart with no version rows returns [] (not 404).""" + _persist_fixture_state() + # Create a chart without subsequently editing it. + chart = Slice( + slice_name="Untouched chart for version list test", + datasource_type="table", + viz_type="table", + ) + db.session.add(chart) + db.session.commit() + chart_uuid = str(chart.uuid) + chart_id = chart.id + + try: + # Purge the INSERT version row so the history is genuinely empty. + ver_cls = version_class(Slice) + db.session.query(ver_cls).filter(ver_cls.id == chart_id).delete( + synchronize_session=False + ) + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._list_versions(chart_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["count"] == 0 + assert body["result"] == [] + finally: + db.session.rollback() + stale = db.session.query(Slice).filter(Slice.id == chart_id).one_or_none() + if stale is not None: + db.session.delete(stale) + db.session.commit() + + def test_list_versions_returns_404_for_unknown_uuid(self) -> None: + """An unknown UUID returns 404.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_list_versions_returns_400_for_invalid_uuid(self) -> None: + """A malformed UUID string is rejected with 400.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("not-a-uuid") + assert rv.status_code == 400 + + def test_list_versions_denies_non_owner(self) -> None: + """T056 — Alpha has ``can_write`` on Chart but doesn't own the + admin-owned fixture, so the row-level ownership check rejects.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + chart_uuid = str(chart.uuid) + + self.login(ALPHA_USERNAME) + rv = self._list_versions(chart_uuid) + assert rv.status_code == 403 + + def test_list_versions_admin_sees_all_entities(self) -> None: + """FR-013: workspace admin can list versions for any entity.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + chart_uuid = str(chart.uuid) + + self.login(ADMIN_USERNAME) + rv = self._list_versions(chart_uuid) + assert rv.status_code == 200 + + +class TestChartRestoreApi(SupersetTestCase): + """T037 — POST /api/v1/chart//versions//restore.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _restore(self, chart_uuid: str, version_uuid: str) -> Any: + return self.client.post( + f"/api/v1/chart/{chart_uuid}/versions/{version_uuid}/restore" + ) + + def _list(self, chart_uuid: str) -> Any: + return self.client.get(f"/api/v1/chart/{chart_uuid}/versions/") + + def test_restore_applies_scalar_field_from_target_version(self) -> None: + """Restoring version 0 puts the slice_name back to its pre-edit value + and appends a new version entry.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_uuid = str(chart.uuid) + chart_id = chart.id + original_name = chart.slice_name + + try: + # Produce two additional saves so version history is 0/1/2. + chart.slice_name = "Girls v1" + db.session.commit() + chart.slice_name = "Girls v2" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv_list = self._list(chart_uuid) + assert rv_list.status_code == 200 + listing = _json.loads(rv_list.data.decode("utf-8")) + initial_count = listing["count"] + assert initial_count >= 3 + target_uuid = listing["result"][0]["version_uuid"] + + # Restore to the first version (the original "Girls" name). + rv = self._restore(chart_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + # Live state matches the restored snapshot. + db.session.expire_all() + chart = db.session.query(Slice).filter(Slice.uuid == chart.uuid).one() + assert chart.slice_name == original_name + + # A new version row was recorded (non-destructive). + rv_list2 = self._list(chart_uuid) + body = _json.loads(rv_list2.data.decode("utf-8")) + assert body["count"] == initial_count + 1 + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_restore_returns_404_for_unknown_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._restore( + "00000000-0000-0000-0000-000000000000", + "00000000-0000-0000-0000-000000000001", + ) + assert rv.status_code == 404 + + def test_restore_returns_404_for_unknown_version_uuid(self) -> None: + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + self.login(ADMIN_USERNAME) + rv = self._restore(str(chart.uuid), "00000000-0000-0000-0000-000000000099") + assert rv.status_code == 404 + + def test_restore_returns_400_for_invalid_entity_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._restore("not-a-uuid", "00000000-0000-0000-0000-000000000001") + assert rv.status_code == 400 + + def test_restore_returns_400_for_invalid_version_uuid(self) -> None: + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + self.login(ADMIN_USERNAME) + rv = self._restore(str(chart.uuid), "not-a-uuid") + assert rv.status_code == 400 + + def test_get_version_returns_historical_snapshot(self) -> None: + """GET /versions// returns the chart's fields at that version + without modifying live state.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_uuid = str(chart.uuid) + chart_id = chart.id + original_name = chart.slice_name + + try: + chart.slice_name = "Girls (v1)" + db.session.commit() + + self.login(ADMIN_USERNAME) + listing = _json.loads(self._list(chart_uuid).data.decode("utf-8")) + assert listing["count"] >= 2 + # The earliest entry should still hold the original slice_name. + first_version_uuid = listing["result"][0]["version_uuid"] + + rv = self.client.get( + f"/api/v1/chart/{chart_uuid}/versions/{first_version_uuid}/" + ) + assert rv.status_code == 200, rv.data + body = _json.loads(rv.data.decode("utf-8"))["result"] + assert body["slice_name"] == original_name + assert body["_version"]["version_uuid"] == first_version_uuid + assert body["_version"]["version_number"] == 0 + # Live row unchanged. + db.session.expire_all() + live = db.session.query(Slice).filter(Slice.uuid == chart.uuid).one() + assert live.slice_name == "Girls (v1)" + finally: + db.session.rollback() + live = db.session.query(Slice).filter(Slice.id == chart_id).one() + live.slice_name = original_name + db.session.commit() + + def test_get_version_returns_404_for_unknown_entity(self) -> None: + self.login(ADMIN_USERNAME) + rv = self.client.get( + "/api/v1/chart/00000000-0000-0000-0000-000000000000" + "/versions/00000000-0000-0000-0000-000000000001/" + ) + assert rv.status_code == 404 + + def test_get_version_returns_400_for_invalid_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self.client.get( + "/api/v1/chart/not-a-uuid/versions/00000000-0000-0000-0000-000000000001/" + ) + assert rv.status_code == 400 + + def test_restore_stamps_changed_by_with_restoring_user(self) -> None: + """After a restore, changed_by_fk on the live entity must point at + the restoring user (not at whoever authored the version being + restored). created_by_fk stays unchanged. The new version row + produced by the restore also carries the restoring user in its + changed_by metadata. + """ + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + self.login(ADMIN_USERNAME) + admin_id = self.get_user(ADMIN_USERNAME).id + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_id = chart.id + chart_uuid = str(chart.uuid) + entity_uuid = chart.uuid + original_name = chart.slice_name + original_created_by = chart.created_by_fk + before_changed_on = chart.changed_on + + try: + # Produce a second version to restore to. + chart.slice_name = "Girls v1" + db.session.commit() + + ver_cls = version_class(Slice) + first_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart_id) + .order_by(ver_cls.transaction_id.asc()) + .limit(1) + .scalar() + ) + assert first_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, first_tx)) + + rv = self.client.post( + f"/api/v1/chart/{chart_uuid}/versions/{target_uuid}/restore" + ) + assert rv.status_code == 200, rv.data + + db.session.expire_all() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + + # Live entity checks. + assert chart.slice_name == original_name + assert chart.created_by_fk == original_created_by + assert chart.changed_by_fk == admin_id, ( + f"Expected changed_by_fk to be restoring user id={admin_id}, " + f"got {chart.changed_by_fk}" + ) + if before_changed_on is not None and chart.changed_on is not None: + assert chart.changed_on >= before_changed_on + + # The new version row produced by the restore must attribute the + # change to the restoring user. + rv_list = self.client.get(f"/api/v1/chart/{chart_uuid}/versions/") + assert rv_list.status_code == 200 + body = _json.loads(rv_list.data.decode("utf-8")) + latest_entry = body["result"][-1] + assert latest_entry["changed_by"] is not None, ( + "New version row should have a changed_by" + ) + assert latest_entry["changed_by"]["id"] == admin_id + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_put_response_returns_old_and_new_version_numbers(self) -> None: + """PUT /api/v1/chart/ response must include old_version and + new_version matching the list-versions ordering.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_id = chart.id + original_name = chart.slice_name + + try: + ver_cls = version_class(Slice) + count_before = ( + db.session.query(ver_cls).filter(ver_cls.id == chart_id).count() + ) + expected_old = count_before - 1 if count_before > 0 else None + + self.login(ADMIN_USERNAME) + rv = self.client.put( + f"/api/v1/chart/{chart_id}", + json={"slice_name": "put-response-version-test"}, + ) + assert rv.status_code == 200, rv.data + body = _json.loads(rv.data.decode("utf-8")) + assert body["id"] == chart_id + assert body["old_version"] == expected_old + assert body["new_version"] is not None + assert "old_transaction_id" in body + assert "new_transaction_id" in body + if body["old_transaction_id"] is not None: + assert body["new_transaction_id"] != body["old_transaction_id"] + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_restore_denies_non_owner(self) -> None: + """T056 — Alpha has ``can_write`` on Chart but isn't an owner of + the admin-owned fixture, so ``BaseRestoreVersionCommand.validate`` + rejects with 403.""" + _persist_fixture_state() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + ) + assert chart is not None + chart_uuid = str(chart.uuid) + + self.login(ALPHA_USERNAME) + rv = self._restore(chart_uuid, "00000000-0000-0000-0000-000000000001") + assert rv.status_code == 403 diff --git a/tests/integration_tests/dashboards/superset_factory_util.py b/tests/integration_tests/dashboards/superset_factory_util.py index b569bc72d68b..5a3f5f7e165e 100644 --- a/tests/integration_tests/dashboards/superset_factory_util.py +++ b/tests/integration_tests/dashboards/superset_factory_util.py @@ -25,7 +25,6 @@ from superset.models.core import Database from superset.models.dashboard import ( Dashboard, - dashboard_slices, dashboard_user, DashboardRoles, ) @@ -234,9 +233,15 @@ def delete_dashboard_roles_associations(dashboard: Dashboard) -> None: def delete_dashboard_slices_associations(dashboard: Dashboard) -> None: - db.session.execute( - dashboard_slices.delete().where(dashboard_slices.c.dashboard_id == dashboard.id) - ) + # Use ORM-level reassignment instead of `db.session.execute(table.delete())`. + # SQLAlchemy-Continuum's M2M tracker needs row-level visibility to record + # shadow entries; a bulk DELETE via Core bypasses the ORM and produces a + # malformed INSERT into `dashboard_slices_version` (missing the composite-PK + # columns), which fails under MySQL strict mode and produces dead rows on + # Postgres. Mirrors the precedent set by ``DatasetDAO.update_columns`` + # being rewritten to ORM-level ``session.delete()`` for the same reason. + dashboard.slices = [] + db.session.flush() def delete_all_inserted_slices(): diff --git a/tests/integration_tests/dashboards/version_history_tests.py b/tests/integration_tests/dashboards/version_history_tests.py new file mode 100644 index 000000000000..97234f35dae4 --- /dev/null +++ b/tests/integration_tests/dashboards/version_history_tests.py @@ -0,0 +1,671 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Integration tests for Dashboard version history capture. + +T015 — dashboard version capture (single version per save; no extra rows from + process_tab_diff) +T018 — retention pruning (drop rows older than SUPERSET_VERSION_HISTORY_RETENTION_DAYS) +T027 — dashboard version list endpoint +""" + +from __future__ import annotations + +from typing import Any + +import pytest +from sqlalchemy_continuum import version_class + +from superset.extensions import db +from superset.models.dashboard import Dashboard +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME, ALPHA_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _get_version_rows(dashboard: Dashboard) -> list[Any]: + ver_cls = version_class(Dashboard) + return ( + db.session.query(ver_cls) + .filter(ver_cls.id == dashboard.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + + +def _persist_fixture_state() -> None: + """Force fixture's pending INSERTs to commit in their own transaction. + + The birth_names fixture stages charts and the dashboard via session.add() + but does not commit. Without this, the test's first commit batches the + INSERTs and UPDATEs into the same Continuum transaction, causing the + existing version row to be updated in place instead of a new one being + created. + """ + db.session.commit() + + +class TestDashboardVersionCapture(SupersetTestCase): + """T015 — one version row per save; no multiple rows from tab/filter diff processing.""" # noqa: E501 + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def test_single_save_creates_one_version_row(self) -> None: + """Saving a dashboard title creates exactly one update version row.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + + original_title = dashboard.dashboard_title + dashboard_id = dashboard.id + + try: + # Capture tx IDs that exist before this save — we'll verify that + # exactly ONE new tx_id with operation_type=1 appears after the save + # (comparing by tx_id makes the test robust against retention + # pruning of older rows). + tx_ids_before = {r.transaction_id for r in _get_version_rows(dashboard)} + + dashboard.dashboard_title = "USA Births Names (edited)" + db.session.commit() + + rows_after = _get_version_rows(dashboard) + new_update_rows = [ + r + for r in rows_after + if r.operation_type == 1 and r.transaction_id not in tx_ids_before + ] + assert len(new_update_rows) == 1, ( + f"Expected 1 new update row from this save, got {len(new_update_rows)}" # noqa: E501 + " — possible no_autoflush regression" + ) + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + def test_second_save_adds_one_row(self) -> None: + """Each subsequent save adds exactly one more version row.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + + original_title = dashboard.dashboard_title + dashboard_id = dashboard.id + + try: + # Track tx IDs across saves; compare by tx_id to sidestep retention + # pruning of older rows. + tx_before_v1 = {r.transaction_id for r in _get_version_rows(dashboard)} + dashboard.dashboard_title = "USA Births Names v1" + db.session.commit() + tx_after_v1 = {r.transaction_id for r in _get_version_rows(dashboard)} + new_txs_v1 = tx_after_v1 - tx_before_v1 + assert len(new_txs_v1) == 1, ( + f"Expected 1 new tx from v1 save, got {len(new_txs_v1)}" + ) + + dashboard.dashboard_title = "USA Births Names v2" + db.session.commit() + tx_after_v2 = {r.transaction_id for r in _get_version_rows(dashboard)} + new_txs_v2 = tx_after_v2 - tx_after_v1 + assert len(new_txs_v2) == 1, ( + f"Expected 1 new tx from v2 save, got {len(new_txs_v2)}" + ) + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + +class TestDashboardVersionRetention(SupersetTestCase): + """T018 — retention pruning drops shadow rows older than SUPERSET_VERSION_HISTORY_RETENTION_DAYS.""" # noqa: E501 + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def test_retention_prunes_old_rows(self) -> None: + """``prune_old_versions`` removes shadow rows whose owning + ``version_transaction.issued_at`` is older than the retention + window, while preserving the live row and the baseline.""" + from datetime import datetime, timedelta + + import sqlalchemy as sa + + from superset.extensions import db as _db + from superset.tasks.version_history_retention import ( + _prune_old_versions_impl, + ) + + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + + original_title = dashboard.dashboard_title + + try: + # Force a few saves so we have ≥ 2 closed shadow rows plus + # a baseline plus the live row. + for i in range(3): + dashboard.dashboard_title = f"USA Births Names retention test {i}" + db.session.commit() + + rows_before = _get_version_rows(dashboard) + assert len(rows_before) >= 3, "Expected at least 3 version rows" + + # Backdate every version_transaction row by 100 days so the + # prune sees them as old. Skip baseline+live rows; the prune + # itself preserves them. + from sqlalchemy_continuum import versioning_manager + + tx_table = versioning_manager.transaction_cls.__table__ + with _db.engine.begin() as conn: + conn.execute( + sa.update(tx_table).values( + issued_at=datetime.utcnow() - timedelta(days=100) + ) + ) + + stats = _prune_old_versions_impl(retention_days=30) + assert stats.get("pruned_transactions", 0) >= 1, stats + + rows_after = _get_version_rows(dashboard) + # Live row must still exist (this is the only preservation rule) + live_rows = [r for r in rows_after if r.end_transaction_id is None] + assert len(live_rows) >= 1, "Live row must never be pruned" + # Some rows should have been pruned. Closed historical rows — + # including the synthetic baseline (operation_type=0) — are + # subject to retention like everything else. + assert len(rows_after) < len(rows_before), ( + f"Expected fewer rows after prune; before={len(rows_before)} " + f"after={len(rows_after)}" + ) + + finally: + dashboard.dashboard_title = original_title + db.session.commit() + + def test_retention_retries_on_serialization_failure(self) -> None: + """A transient ``OperationalError`` from the SERIALIZABLE pass + triggers an inline retry; the prune completes on the second + attempt and the stats dict records the retry count.""" + from datetime import datetime, timedelta + from unittest.mock import patch + + import sqlalchemy as sa + from sqlalchemy.exc import OperationalError + + from superset.tasks import version_history_retention + from superset.tasks.version_history_retention import ( + _prune_old_versions_impl, + ) + + # Backdate transactions so the prune has work to do. + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + original_title = dashboard.dashboard_title + try: + for i in range(3): + dashboard.dashboard_title = f"USA Births Names retry test {i}" + db.session.commit() + + from sqlalchemy_continuum import versioning_manager + + tx_table = versioning_manager.transaction_cls.__table__ + from superset.extensions import db as _db + + with _db.engine.begin() as conn: + conn.execute( + sa.update(tx_table).values( + issued_at=datetime.utcnow() - timedelta(days=100) + ) + ) + + original_run = version_history_retention._run_prune_pass + calls: list[int] = [] + + def flaky_run(*args: Any, **kwargs: Any) -> dict[str, Any]: + calls.append(1) + if len(calls) == 1: + raise OperationalError( + "SELECT 1", {}, Exception("could not serialize access") + ) + return original_run(*args, **kwargs) + + with patch.object( + version_history_retention, "_run_prune_pass", side_effect=flaky_run + ): + # Patch sleep so the test doesn't actually wait through + # the backoff. + with patch.object(version_history_retention.time, "sleep"): + stats = _prune_old_versions_impl(retention_days=30) + + assert len(calls) == 2, ( + f"Expected 2 _run_prune_pass calls (1 failure + 1 retry), " + f"got {len(calls)}" + ) + assert stats.get("retried") == 1, stats + assert stats.get("pruned_transactions", 0) >= 1, stats + finally: + dashboard.dashboard_title = original_title + db.session.commit() + + def test_retention_gives_up_after_max_attempts(self) -> None: + """When every attempt hits ``OperationalError``, the function + re-raises after the retry cap so the outer Celery wrapper logs + + returns ``{"error": 1}``.""" + from unittest.mock import patch + + from sqlalchemy.exc import OperationalError + + from superset.tasks import version_history_retention + from superset.tasks.version_history_retention import ( + _MAX_RETRY_ATTEMPTS, + _prune_old_versions_impl, + ) + + def always_fail(*args: Any, **kwargs: Any) -> dict[str, Any]: + raise OperationalError( + "SELECT 1", {}, Exception("could not serialize access") + ) + + call_count = 0 + + def counting_fail(*args: Any, **kwargs: Any) -> dict[str, Any]: + nonlocal call_count + call_count += 1 + return always_fail(*args, **kwargs) + + with patch.object( + version_history_retention, + "_run_prune_pass", + side_effect=counting_fail, + ): + with patch.object(version_history_retention.time, "sleep"): + with pytest.raises(OperationalError): + _prune_old_versions_impl(retention_days=30) + + assert call_count == _MAX_RETRY_ATTEMPTS, ( + f"Expected exactly {_MAX_RETRY_ATTEMPTS} attempts; got {call_count}" + ) + + +class TestDashboardVersionListApi(SupersetTestCase): + """T027 — GET /api/v1/dashboard//versions/ endpoint.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _list_versions(self, dashboard_uuid: str) -> Any: + return self.client.get(f"/api/v1/dashboard/{dashboard_uuid}/versions/") + + def test_list_versions_returns_ordered_sequence(self) -> None: + """Saving a dashboard three times extends the version list by three.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + original_title = dashboard.dashboard_title + dashboard_id = dashboard.id + dashboard_uuid = str(dashboard.uuid) + + try: + self.login(ADMIN_USERNAME) + rv = self._list_versions(dashboard_uuid) + assert rv.status_code == 200 + assert "count" in _json.loads(rv.data.decode("utf-8")) + + for i in range(3): + dashboard.dashboard_title = f"USA Births Names v{i}" + db.session.commit() + + rv = self._list_versions(dashboard_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + # Delta-based assertion — retention pruning from other tests can lower + # the absolute count, but each of our three saves must produce exactly + # one new entry. We compare by transaction_id instead. + assert len(body["result"]) == body["count"] + for idx, entry in enumerate(body["result"]): + assert entry["version_number"] == idx + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + def test_list_versions_empty_for_untouched_entity(self) -> None: + """A dashboard with no version rows returns [] (not 404).""" + _persist_fixture_state() + dashboard = Dashboard(dashboard_title="Untouched dashboard", slug="untouched") + db.session.add(dashboard) + db.session.commit() + dashboard_uuid = str(dashboard.uuid) + dashboard_id = dashboard.id + + try: + ver_cls = version_class(Dashboard) + db.session.query(ver_cls).filter(ver_cls.id == dashboard_id).delete( + synchronize_session=False + ) + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._list_versions(dashboard_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["count"] == 0 + assert body["result"] == [] + finally: + db.session.rollback() + stale = ( + db.session.query(Dashboard) + .filter(Dashboard.id == dashboard_id) + .one_or_none() + ) + if stale is not None: + db.session.delete(stale) + db.session.commit() + + def test_list_versions_returns_404_for_unknown_uuid(self) -> None: + """An unknown UUID returns 404.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_list_versions_returns_400_for_invalid_uuid(self) -> None: + """A malformed UUID string is rejected with 400.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("not-a-uuid") + assert rv.status_code == 400 + + def test_list_versions_denies_non_owner(self) -> None: + """T056 — Alpha has ``can_write`` on Dashboard but doesn't own the + admin-owned fixture, so the row-level ownership check rejects.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + + self.login(ALPHA_USERNAME) + rv = self._list_versions(dashboard_uuid) + assert rv.status_code == 403 + + def test_list_versions_admin_sees_all_entities(self) -> None: + """FR-013: workspace admin can list versions for any entity.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + + self.login(ADMIN_USERNAME) + rv = self._list_versions(dashboard_uuid) + assert rv.status_code == 200 + + +class TestDashboardRestoreApi(SupersetTestCase): + """T038 — POST /api/v1/dashboard//versions//restore.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _restore(self, dashboard_uuid: str, version_uuid: str) -> Any: + return self.client.post( + f"/api/v1/dashboard/{dashboard_uuid}/versions/{version_uuid}/restore" + ) + + def test_restore_applies_scalar_field(self) -> None: + """Restore a dashboard title edit.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + original_title = dashboard.dashboard_title + dashboard_id = dashboard.id + entity_uuid = dashboard.uuid + + try: + # Make two more edits so we have a known non-trivial history to + # navigate: [initial, v1, v2]. + dashboard.dashboard_title = "USA Births Names v1" + db.session.commit() + dashboard.dashboard_title = "USA Births Names v2" + db.session.commit() + + ver_cls = version_class(Dashboard) + rows = ( + db.session.query( + ver_cls.transaction_id, + ver_cls.operation_type, + ver_cls.dashboard_title, + ver_cls.end_transaction_id, + ) + .filter(ver_cls.id == dashboard_id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + # Find the version whose snapshot has the original title. Skip DELETE + # rows (operation_type=2) — the integration DB may carry shadow rows + # from prior fixture teardown cycles, and restoring to a DELETE state + # would re-delete the live entity. + target_row = next( + ( + row + for row in rows + if row.dashboard_title == original_title and row.operation_type != 2 + ), + None, + ) + assert target_row is not None, ( + f"Expected at least one version row with original title; rows={rows}" + ) + target_uuid = str( + derive_version_uuid(entity_uuid, target_row.transaction_id) + ) + + self.login(ADMIN_USERNAME) + rv = self._restore(dashboard_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + db.session.expire_all() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + assert dashboard.dashboard_title == original_title, ( + f"Restore did not revert title; rows={rows}" + ) + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + def test_restore_reattaches_chart_removed_after_snapshot(self) -> None: + """After the target snapshot is captured, detaching a chart and saving + must be undone by restore — the chart comes back on dashboard_slices.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + dashboard_id = dashboard.id + entity_uuid = dashboard.uuid + + original_slice_ids = sorted(s.id for s in dashboard.slices) + assert len(original_slice_ids) >= 2, ( + f"fixture expected to attach >= 2 charts; got {original_slice_ids}" + ) + slice_to_drop = dashboard.slices[0] + drop_id = slice_to_drop.id + + # Touch the dashboard so a snapshot row is captured at a known tx. + dashboard.dashboard_title = "USA Births Names — snapshot point" + db.session.commit() + + ver_cls = version_class(Dashboard) + target_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == dashboard_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + assert target_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, target_tx)) + + # Detach the chart and commit — moves history forward. + dashboard.slices.remove(slice_to_drop) + db.session.commit() + + db.session.expire_all() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + live_ids = {s.id for s in dashboard.slices} + assert drop_id not in live_ids, "pre-restore: dropped chart should be detached" + + self.login(ADMIN_USERNAME) + rv = self._restore(dashboard_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + db.session.expire_all() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + restored_ids = sorted(s.id for s in dashboard.slices) + assert restored_ids == original_slice_ids, ( + f"restore did not re-attach chart: expected {original_slice_ids}, " + f"got {restored_ids}" + ) + + def test_restore_returns_404_for_unknown_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._restore( + "00000000-0000-0000-0000-000000000000", + "00000000-0000-0000-0000-000000000001", + ) + assert rv.status_code == 404 + + def test_restore_returns_404_for_unknown_version_uuid(self) -> None: + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._restore(str(dashboard.uuid), "00000000-0000-0000-0000-000000000099") + assert rv.status_code == 404 + + def test_put_response_returns_old_and_new_version_numbers(self) -> None: + """PUT /api/v1/dashboard/ response must include old_version and + new_version matching the list-versions ordering.""" + _persist_fixture_state() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + dashboard_id = dashboard.id + original_title = dashboard.dashboard_title + + try: + ver_cls = version_class(Dashboard) + count_before = ( + db.session.query(ver_cls).filter(ver_cls.id == dashboard_id).count() + ) + expected_old = count_before - 1 if count_before > 0 else None + + self.login(ADMIN_USERNAME) + rv = self.client.put( + f"/api/v1/dashboard/{dashboard_id}", + json={"dashboard_title": "put-response-version-test"}, + ) + assert rv.status_code == 200, rv.data + body = _json.loads(rv.data.decode("utf-8")) + assert body["id"] == dashboard_id + assert body["old_version"] == expected_old + assert body["new_version"] is not None + assert "old_transaction_id" in body + assert "new_transaction_id" in body + if body["old_transaction_id"] is not None: + assert body["new_transaction_id"] != body["old_transaction_id"] + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() diff --git a/tests/integration_tests/datasets/version_history_tests.py b/tests/integration_tests/datasets/version_history_tests.py new file mode 100644 index 000000000000..8382da13832b --- /dev/null +++ b/tests/integration_tests/datasets/version_history_tests.py @@ -0,0 +1,706 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Integration tests for Dataset (SqlaTable) version history capture. + +T016 — dataset column and metric version rows are created via ORM (not bulk) ops +T028 — dataset version list endpoint +""" + +from __future__ import annotations + +from typing import Any + +import pytest +import sqlalchemy as sa +from sqlalchemy_continuum import version_class + +from superset.connectors.sqla.models import SqlaTable, SqlMetric, TableColumn +from superset.extensions import db +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ( + ADMIN_USERNAME, + ALPHA_USERNAME, + GAMMA_USERNAME, +) +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _get_table_column_version_rows(column: TableColumn) -> list[Any]: + ver_cls = version_class(TableColumn) + return ( + db.session.query(ver_cls) + .filter(ver_cls.id == column.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + + +def _get_sql_metric_version_rows(metric: SqlMetric) -> list[Any]: + ver_cls = version_class(SqlMetric) + return ( + db.session.query(ver_cls) + .filter(ver_cls.id == metric.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + + +def _get_table_version_rows(table: SqlaTable) -> list[Any]: + ver_cls = version_class(SqlaTable) + return ( + db.session.query(ver_cls) + .filter(ver_cls.id == table.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + + +def _persist_fixture_state() -> None: + """Force fixture's pending INSERTs to commit in their own transaction. + + The birth_names fixture stages charts and the dashboard via session.add() + but does not commit. Without this, the test's first commit batches the + INSERTs and UPDATEs into the same Continuum transaction, causing the + existing version row to be updated in place instead of a new one being + created. + """ + db.session.commit() + + +class TestDatasetVersionListApi(SupersetTestCase): + """T028 — GET /api/v1/dataset//versions/ endpoint.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _list_versions(self, dataset_uuid: str) -> Any: + return self.client.get(f"/api/v1/dataset/{dataset_uuid}/versions/") + + def test_list_versions_returns_ordered_sequence(self) -> None: + """Editing a dataset produces ascending version_number entries.""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + original_description = table.description + table_uuid = str(table.uuid) + table_id = table.id + + try: + for i in range(3): + table.description = f"Test description v{i}" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._list_versions(table_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["count"] == len(body["result"]) + for idx, entry in enumerate(body["result"]): + assert entry["version_number"] == idx + assert entry["issued_at"] is not None + # issued_at is an RFC-1123 HTTP date ("Wed, 22 Apr 2026 …"); parse + # before checking monotonic order rather than sorting strings, + # which would reorder incorrectly across day-of-week boundaries. + from email.utils import parsedate_to_datetime + + parsed = [parsedate_to_datetime(e["issued_at"]) for e in body["result"]] + assert parsed == sorted(parsed) + finally: + # Restore fixture state even if an assertion above failed (otherwise + # the polluted description cascades to later tests in the suite). + db.session.rollback() + table = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + table.description = original_description + db.session.commit() + + def test_list_versions_empty_for_untouched_entity(self) -> None: + """A dataset with no version rows returns [] (not 404).""" + _persist_fixture_state() + table = SqlaTable( + table_name="__untouched_table_for_version_list__", + database_id=1, + ) + db.session.add(table) + db.session.commit() + table_uuid = str(table.uuid) + table_id = table.id + + try: + ver_cls = version_class(SqlaTable) + db.session.query(ver_cls).filter(ver_cls.id == table_id).delete( + synchronize_session=False + ) + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._list_versions(table_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["count"] == 0 + assert body["result"] == [] + finally: + db.session.rollback() + stale = ( + db.session.query(SqlaTable) + .filter(SqlaTable.id == table_id) + .one_or_none() + ) + if stale is not None: + db.session.delete(stale) + db.session.commit() + + def test_list_versions_returns_404_for_unknown_uuid(self) -> None: + """An unknown UUID returns 404.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_list_versions_returns_400_for_invalid_uuid(self) -> None: + """A malformed UUID string is rejected with 400.""" + self.login(ADMIN_USERNAME) + rv = self._list_versions("not-a-uuid") + assert rv.status_code == 400 + + def test_list_versions_denies_without_write_permission(self) -> None: + """Gamma is read-only on Dataset — 403 on list_versions.""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + + self.login(GAMMA_USERNAME) + rv = self._list_versions(table_uuid) + assert rv.status_code == 403 + + def test_list_versions_denies_non_owner(self) -> None: + """T056 — Alpha has ``can_write`` on Dataset but isn't an owner of + the admin-owned fixture, so the row-level ownership check rejects. + Exercises the row-level branch specifically (the Gamma test above + only proves model-level denial via ``@protect()``).""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + + self.login(ALPHA_USERNAME) + rv = self._list_versions(table_uuid) + assert rv.status_code == 403 + + def test_list_versions_admin_sees_all_entities(self) -> None: + """FR-013: workspace admin can list versions for any entity.""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + + self.login(ADMIN_USERNAME) + rv = self._list_versions(table_uuid) + assert rv.status_code == 200 + + +class TestDatasetRestoreApi(SupersetTestCase): + """T039 — POST /api/v1/dataset//versions//restore.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def setUp(self) -> None: + # Reset session state before each test in this class so the restore + # path is exercised against a clean identity map rather than whatever + # half-flushed state a previous test in the full-suite run may have + # left behind. Specifically: a Postgres-only multi-test cascade (see + # the sc-103156 follow-up note) can leave Continuum's shadow-table + # session attributes in a state where the restore command's + # ``@transaction`` boundary unexpectedly raises and surfaces as 422 + # "Dataset could not be updated." Rolling back + expiring all clears + # the cascade for this class' tests without modifying the upstream + # tests that cause it. + super().setUp() + db.session.rollback() + db.session.expire_all() + + def tearDown(self) -> None: + db.session.rollback() + db.session.expire_all() + super().tearDown() + + def _restore(self, dataset_uuid: str, version_uuid: str) -> Any: + return self.client.post( + f"/api/v1/dataset/{dataset_uuid}/versions/{version_uuid}/restore" + ) + + def test_restore_applies_scalar_field(self) -> None: + """Restore a dataset's description edit.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + entity_uuid = table.uuid + table_id = table.id + original_description = table.description + + try: + # Two more edits to produce a non-trivial history. + table.description = "restore-test v1" + db.session.commit() + table.description = "restore-test v2" + db.session.commit() + + ver_cls = version_class(SqlaTable) + rows = ( + db.session.query( + ver_cls.transaction_id, + ver_cls.operation_type, + ver_cls.description, + ) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + # Skip DELETE rows (operation_type=2) — the integration DB may carry + # shadow rows from prior fixture teardown cycles, and restoring to a + # DELETE state would re-delete the live entity (same fix as the + # dashboard restore test). + target_row = next( + ( + row + for row in rows + if row.description == original_description + and row.operation_type != 2 + ), + None, + ) + assert target_row is not None, ( + f"No version with original description; rows={rows}" + ) + target_uuid = str( + derive_version_uuid(entity_uuid, target_row.transaction_id) + ) + + self.login(ADMIN_USERNAME) + rv = self._restore(table_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + db.session.expire_all() + table = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + assert table.description == original_description + finally: + # Cleanup — guard fixture state against assertion failures cascading + # to later tests in the suite (saw this manifest on Postgres CI's + # full-suite ordering: a failure here left ``description="restore-test + # v2"`` on birth_names and polluted downstream tests). + db.session.rollback() + table = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + table.description = original_description + db.session.commit() + + def test_restore_with_column_edits_reverts_columns(self) -> None: + """After editing a column's description, restoring an earlier version + reverts the column.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + entity_uuid = table.uuid + table_id = table.id + + col = table.columns[0] + col_name = col.column_name + original_col_description = col.description + + try: + # Snapshot target version before our column edit. + ver_cls = version_class(SqlaTable) + last_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + assert last_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, last_tx)) + + col.description = "restore-test column edit" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._restore(table_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + # JSON-snapshot restore reassigns child PKs, so look up by natural + # key (column_name) rather than the old id. + db.session.expire_all() + col = ( + db.session.query(TableColumn) + .filter(TableColumn.table_id == table_id) + .filter(TableColumn.column_name == col_name) + .one() + ) + assert col.description == original_col_description + finally: + db.session.rollback() + col = ( + db.session.query(TableColumn) + .filter(TableColumn.table_id == table_id) + .filter(TableColumn.column_name == col_name) + .one_or_none() + ) + if col is not None: + col.description = original_col_description + db.session.commit() + + def test_restore_adds_back_removed_column_and_drops_added_one(self) -> None: + """After a snapshot is taken, removing an existing column and adding + a new one, restoring the snapshot must undo both operations.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_id = table.id + table_uuid = str(table.uuid) + entity_uuid = table.uuid + + original_col_names = sorted(c.column_name for c in table.columns) + removed_name = table.columns[0].column_name + + # Capture a snapshot tx point by touching the dataset. + table.description = "snapshot before column-swap" + db.session.commit() + + ver_cls = version_class(SqlaTable) + target_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + assert target_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, target_tx)) + + # Remove a column, add a new one, commit (moves history forward). + db.session.delete(table.columns[0]) + db.session.add( + TableColumn( + table_id=table_id, + column_name="__restore_test_calc__", + expression="1", + ) + ) + db.session.commit() + + assert removed_name not in {c.column_name for c in table.columns} + assert "__restore_test_calc__" in {c.column_name for c in table.columns} + + self.login(ADMIN_USERNAME) + rv = self._restore(table_uuid, target_uuid) + assert rv.status_code == 200, rv.data + + db.session.expire_all() + table = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + restored_names = sorted(c.column_name for c in table.columns) + assert restored_names == original_col_names + + def test_restore_emits_full_child_diff_in_one_transaction(self) -> None: + """A restore that re-adds one column and drops another MUST write + *both* change records under the same transaction. Under the prior + per-relation flush loop the first flush emitted only the + easier-to-detect change (the modification of a surviving + column), the listener's tx-dedup guard then suppressed the + second pass, and the addition record was silently lost from + ``version_changes`` — the dropdown rendered the restore as an + empty "Baseline" entry. Locks in the single-flush restore + behavior in ``VersionDAO.restore_version``. + """ + from superset.daos.version import derive_version_uuid + from superset.versioning.changes import version_changes_table + + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_id = table.id + table_uuid = str(table.uuid) + entity_uuid = table.uuid + removed_name = table.columns[0].column_name + added_name = "__restore_full_diff_test__" + + # Snapshot point captures the baseline. + table.description = "snapshot before full-diff column swap" + db.session.commit() + + ver_cls = version_class(SqlaTable) + target_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + assert target_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, target_tx)) + + db.session.delete(table.columns[0]) + db.session.add( + TableColumn(table_id=table_id, column_name=added_name, expression="1") + ) + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._restore(table_uuid, target_uuid) + assert rv.status_code == 200, rv.data + db.session.expire_all() + + restore_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + rows = ( + db.session.connection() + .execute( + sa.select( + version_changes_table.c.kind, + version_changes_table.c.path, + ).where( + version_changes_table.c.transaction_id == restore_tx, + version_changes_table.c.entity_kind == "dataset", + version_changes_table.c.entity_id == table_id, + ) + ) + .all() + ) + paths = {tuple(row.path) for row in rows} + assert ("columns", added_name) in paths, ( + f"restore tx {restore_tx} did not emit removal record for " + f"the added-then-restored-away column {added_name!r}; " + f"observed paths={paths}" + ) + assert ("columns", removed_name) in paths, ( + f"restore tx {restore_tx} did not emit addition record for " + f"the deleted-then-restored column {removed_name!r}; " + f"observed paths={paths}" + ) + + def test_restore_returns_404_for_unknown_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._restore( + "00000000-0000-0000-0000-000000000000", + "00000000-0000-0000-0000-000000000001", + ) + assert rv.status_code == 404 + + def test_restore_returns_404_for_unknown_version_uuid(self) -> None: + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + self.login(ADMIN_USERNAME) + rv = self._restore(str(table.uuid), "00000000-0000-0000-0000-000000000099") + assert rv.status_code == 404 + + def test_restore_returns_400_for_invalid_entity_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._restore("not-a-uuid", "00000000-0000-0000-0000-000000000001") + assert rv.status_code == 400 + + def test_restore_returns_400_for_invalid_version_uuid(self) -> None: + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + self.login(ADMIN_USERNAME) + rv = self._restore(str(table.uuid), "not-a-uuid") + assert rv.status_code == 400 + + def test_get_version_returns_historical_snapshot_with_children(self) -> None: + """GET /versions// on a dataset returns scalar fields and + reconstructed columns/metrics, without modifying live state.""" + from superset.daos.version import derive_version_uuid + + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_id = table.id + table_uuid = str(table.uuid) + entity_uuid = table.uuid + original_description = table.description + original_col_names = sorted(c.column_name for c in table.columns) + + try: + # Capture a snapshot point now; make a change after. + ver_cls = version_class(SqlaTable) + target_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == table_id) + .order_by(ver_cls.transaction_id.desc()) + .limit(1) + .scalar() + ) + assert target_tx is not None + target_uuid = str(derive_version_uuid(entity_uuid, target_tx)) + + table.description = "edited after snapshot" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self.client.get( + f"/api/v1/dataset/{table_uuid}/versions/{target_uuid}/" + ) + assert rv.status_code == 200, rv.data + body = _json.loads(rv.data.decode("utf-8"))["result"] + + # Scalar fields reflect the snapshot, not the live edit. + assert body["description"] == original_description + assert body["_version"]["version_uuid"] == target_uuid + + # Columns list matches original set. + snapshot_col_names = sorted(c["column_name"] for c in body["columns"]) + assert snapshot_col_names == original_col_names + + # Metrics reconstructed. + assert isinstance(body["metrics"], list) + assert all("metric_name" in m for m in body["metrics"]) + + # Live row remains in its edited state. + db.session.expire_all() + live = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + assert live.description == "edited after snapshot" + finally: + db.session.rollback() + live = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + live.description = original_description + db.session.commit() + + def test_put_response_returns_old_and_new_version_numbers(self) -> None: + """PUT /api/v1/dataset/ should include old_version and new_version + fields that match the list-versions endpoint's version_number values.""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_id = table.id + original_description = table.description + + try: + ver_cls = version_class(SqlaTable) + count_before = ( + db.session.query(ver_cls).filter(ver_cls.id == table_id).count() + ) + expected_old = count_before - 1 if count_before > 0 else None + + self.login(ADMIN_USERNAME) + rv = self.client.put( + f"/api/v1/dataset/{table_id}", + json={"description": "version-number response test"}, + ) + assert rv.status_code == 200, rv.data + body = _json.loads(rv.data.decode("utf-8")) + assert body["id"] == table_id + assert "old_version" in body + assert "new_version" in body + assert "old_transaction_id" in body + assert "new_transaction_id" in body + assert body["old_version"] == expected_old + # new_version points to the live row post-commit. It is usually + # old_version + 1, but can equal old_version when retention pruning + # removed an older closed row in the same commit. + assert body["new_version"] is not None + assert body["new_version"] >= 0 + # Transaction ids are stable identifiers, so a successful update + # always produces a new_transaction_id distinct from the previous + # one (when old_transaction_id is known). + if body["old_transaction_id"] is not None: + assert body["new_transaction_id"] != body["old_transaction_id"] + finally: + db.session.rollback() + table = db.session.query(SqlaTable).filter(SqlaTable.id == table_id).one() + table.description = original_description + db.session.commit() + + def test_restore_denies_without_write_permission(self) -> None: + """Gamma is read-only on Dataset — 403 on restore.""" + _persist_fixture_state() + table: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert table is not None + table_uuid = str(table.uuid) + + self.login(GAMMA_USERNAME) + rv = self._restore(table_uuid, "00000000-0000-0000-0000-000000000001") + assert rv.status_code == 403 diff --git a/tests/integration_tests/migrations/composite_pk_association_tables__tests.py b/tests/integration_tests/migrations/composite_pk_association_tables__tests.py new file mode 100644 index 000000000000..52b1942bdb24 --- /dev/null +++ b/tests/integration_tests/migrations/composite_pk_association_tables__tests.py @@ -0,0 +1,131 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Schema-shape assertion tests for the composite-PK association-tables +migration (revision 2bee73611e32). + +Builds the pre-migration shape against an isolated in-memory SQLite engine, +runs the migration's ``upgrade()``, and asserts the resulting shape matches +the data-model.md "After" specification: no ``id`` column, composite PK on +the two FK columns, and no redundant ``UNIQUE(fk1, fk2)`` on the two tables +that previously carried one. + +Continuum-restore verification is OUT OF SCOPE; that work lives in the +versioning epic (sc-103156). Cross-backend verification (PostgreSQL, MySQL) +is handled by the CI matrix (T034a). +""" + +from importlib import import_module + +import pytest +import sqlalchemy as sa +from alembic.migration import MigrationContext +from alembic.operations import Operations +from sqlalchemy import inspect + +# Import the migration module under test. +_migration = import_module( + "superset.migrations.versions." + "2026-05-01_23-36_2bee73611e32_composite_pk_association_tables" +) +AFFECTED_TABLES = _migration.AFFECTED_TABLES +TABLES_WITH_PRE_EXISTING_UNIQUE = _migration.TABLES_WITH_PRE_EXISTING_UNIQUE + + +@pytest.fixture(scope="module") +def post_upgrade_engine() -> sa.engine.Engine: + """An isolated in-memory SQLite engine with the migration applied to a + pre-migration-shaped seed schema. Used by the post-upgrade assertions + below. Module-scoped so the upgrade only runs once per test session.""" + engine = sa.create_engine("sqlite:///:memory:") + md = sa.MetaData() + for t in AFFECTED_TABLES: + cols: list[sa.SchemaItem] = [ + sa.Column("id", sa.Integer, primary_key=True), + sa.Column(t.fk1, sa.Integer, nullable=False), + sa.Column(t.fk2, sa.Integer, nullable=False), + ] + constraints: list[sa.SchemaItem] = [] + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + constraints.append(sa.UniqueConstraint(t.fk1, t.fk2)) + sa.Table(t.name, md, *cols, *constraints) + md.create_all(engine) + + # Apply the migration's upgrade() against this engine via Alembic's + # MigrationContext, patching the migration module's ``op`` reference. + with engine.connect() as conn: + ctx = MigrationContext.configure(conn) + ops = Operations(ctx) + original_op = _migration.op + _migration.op = ops # type: ignore[attr-defined] + try: + _migration.upgrade() + finally: + _migration.op = original_op # type: ignore[attr-defined] + return engine + + +@pytest.mark.parametrize("t", AFFECTED_TABLES, ids=lambda t: t.name) +def test_no_id_column(post_upgrade_engine: sa.engine.Engine, t) -> None: + """The synthetic ``id`` column is gone from each affected table.""" + insp = inspect(post_upgrade_engine) + column_names = {c["name"] for c in insp.get_columns(t.name)} + assert "id" not in column_names, ( + f"{t.name} still has an 'id' column after migration; " + f"composite-PK conversion incomplete" + ) + + +@pytest.mark.parametrize("t", AFFECTED_TABLES, ids=lambda t: t.name) +def test_primary_key_is_composite_fks(post_upgrade_engine: sa.engine.Engine, t) -> None: + """The primary key of each affected table is exactly ``(fk1, fk2)``.""" + insp = inspect(post_upgrade_engine) + pk_cols = set(insp.get_pk_constraint(t.name).get("constrained_columns", [])) + assert pk_cols == {t.fk1, t.fk2}, ( + f"{t.name} primary key is {pk_cols}, expected {{{t.fk1}, {t.fk2}}}" + ) + + +@pytest.mark.parametrize( + "t", + [t for t in AFFECTED_TABLES if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE], + ids=lambda t: t.name, +) +def test_redundant_unique_dropped(post_upgrade_engine: sa.engine.Engine, t) -> None: + """For the two tables that previously carried a UNIQUE(fk1, fk2), that + constraint is now subsumed by the composite PK and must not appear + separately in the unique-constraint list.""" + insp = inspect(post_upgrade_engine) + redundant_pair = {t.fk1, t.fk2} + for uc in insp.get_unique_constraints(t.name): + cols = set(uc.get("column_names", [])) + assert cols != redundant_pair, ( + f"{t.name} still carries a redundant UniqueConstraint over " + f"{redundant_pair} (name={uc.get('name')!r}); " + f"composite-PK conversion incomplete" + ) + + +@pytest.mark.parametrize("t", AFFECTED_TABLES, ids=lambda t: t.name) +def test_fk_columns_not_null(post_upgrade_engine: sa.engine.Engine, t) -> None: + """PK promotion implicitly tightens the FK columns to NOT NULL.""" + insp = inspect(post_upgrade_engine) + cols_by_name = {c["name"]: c for c in insp.get_columns(t.name)} + for col in (t.fk1, t.fk2): + assert col in cols_by_name, f"{t.name} missing column {col}" + assert cols_by_name[col].get("nullable") is False, ( + f"{t.name}.{col} is nullable; expected NOT NULL after PK promotion" + ) diff --git a/tests/integration_tests/migrations/composite_pk_round_trip__tests.py b/tests/integration_tests/migrations/composite_pk_round_trip__tests.py new file mode 100644 index 000000000000..d83c9d113c3f --- /dev/null +++ b/tests/integration_tests/migrations/composite_pk_round_trip__tests.py @@ -0,0 +1,168 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Schema round-trip tests for the composite-PK association-tables migration +(revision 2bee73611e32). Builds the pre-migration shape against an in-memory +SQLite engine, runs the migration's ``upgrade()``, asserts the post-upgrade +shape, runs ``downgrade()``, asserts the prior shape is restored (modulo the +documented FK NOT NULL asymmetry), and re-runs ``upgrade()`` to verify +idempotency. + +This is run against an isolated in-memory engine via Alembic's +``MigrationContext`` so the test does not perturb the project's test DB. + +Cross-backend verification of the same migration against PostgreSQL and +MySQL is delegated to the CI matrix (see T034a in tasks.md) and to the +quickstart.md verification (T033). This file covers the SQLite slice. +""" + +from importlib import import_module +from typing import Any + +import pytest +import sqlalchemy as sa +from alembic.migration import MigrationContext +from alembic.operations import Operations +from sqlalchemy import inspect + +# Import the migration module under test. +_migration = import_module( + "superset.migrations.versions." + "2026-05-01_23-36_2bee73611e32_composite_pk_association_tables" +) +AFFECTED_TABLES = _migration.AFFECTED_TABLES +TABLES_WITH_PRE_EXISTING_UNIQUE = _migration.TABLES_WITH_PRE_EXISTING_UNIQUE + + +def _build_pre_migration_schema(engine: sa.engine.Engine) -> None: + """Recreate the eight tables in their pre-migration shape (surrogate + ``id INTEGER PRIMARY KEY`` plus an optional ``UNIQUE(fk1, fk2)`` on the + two tables that previously carried one). FKs to parent tables are + omitted to keep the test self-contained — we're testing schema + transformations, not FK enforcement.""" + md = sa.MetaData() + for t in AFFECTED_TABLES: + cols: list[sa.Column] = [ + sa.Column("id", sa.Integer, primary_key=True), + sa.Column(t.fk1, sa.Integer, nullable=False), + sa.Column(t.fk2, sa.Integer, nullable=False), + ] + constraints: list[sa.SchemaItem] = [] + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + constraints.append(sa.UniqueConstraint(t.fk1, t.fk2)) + sa.Table(t.name, md, *cols, *constraints) + md.create_all(engine) + + +def _shape(engine: sa.engine.Engine, table: str) -> dict[str, Any]: + """Return a structural summary for asserting equality across runs.""" + insp = inspect(engine) + pk = insp.get_pk_constraint(table).get("constrained_columns", []) + columns = sorted(c["name"] for c in insp.get_columns(table)) + uniques = sorted( + tuple(sorted(uc.get("column_names", []))) + for uc in insp.get_unique_constraints(table) + ) + return {"columns": columns, "pk": sorted(pk), "uniques": uniques} + + +def _run_with_alembic_context(engine: sa.engine.Engine, fn) -> None: + """Run ``fn()`` (the migration's upgrade/downgrade body) inside a fresh + Alembic ``MigrationContext`` bound to ``engine``. Patches the + migration module's ``op`` to point at this context so its + ``op.get_bind()`` and ``op.batch_alter_table`` calls execute against + the in-memory engine.""" + with engine.connect() as conn: + ctx = MigrationContext.configure(conn) + ops = Operations(ctx) + original_op = _migration.op + _migration.op = ops # type: ignore[attr-defined] + try: + fn() + finally: + _migration.op = original_op # type: ignore[attr-defined] + + +def test_round_trip_against_in_memory_sqlite() -> None: + """Round-trip: pre-migration → upgrade → downgrade → upgrade again. + + Asserts: + - Post-upgrade shape: no ``id``, composite PK on (fk1, fk2), no + UNIQUE(fk1, fk2) on the two tables that previously carried one. + - Post-downgrade shape: ``id`` restored, PK back on (id), UNIQUE + re-added on the two tables. (FK columns remain NOT NULL — the + documented intentional asymmetry.) + - Post-re-upgrade idempotency: shape matches the first post-upgrade. + """ + engine = sa.create_engine("sqlite:///:memory:") + _build_pre_migration_schema(engine) + + pre_shape = {t.name: _shape(engine, t.name) for t in AFFECTED_TABLES} + + _run_with_alembic_context(engine, _migration.upgrade) + + for t in AFFECTED_TABLES: + s = _shape(engine, t.name) + assert "id" not in s["columns"], f"{t.name}: id still present post-upgrade: {s}" + assert s["pk"] == sorted([t.fk1, t.fk2]), ( + f"{t.name}: PK is {s['pk']}, expected {sorted([t.fk1, t.fk2])}" + ) + assert tuple(sorted([t.fk1, t.fk2])) not in s["uniques"], ( + f"{t.name}: redundant UNIQUE not dropped post-upgrade: {s['uniques']}" + ) + + post_upgrade_shape = {t.name: _shape(engine, t.name) for t in AFFECTED_TABLES} + + _run_with_alembic_context(engine, _migration.downgrade) + + for t in AFFECTED_TABLES: + s = _shape(engine, t.name) + assert "id" in s["columns"], f"{t.name}: id not restored post-downgrade: {s}" + assert s["pk"] == ["id"], f"{t.name}: PK is {s['pk']}, expected ['id']" + if t.name in TABLES_WITH_PRE_EXISTING_UNIQUE: + assert tuple(sorted([t.fk1, t.fk2])) in s["uniques"], ( + f"{t.name}: UNIQUE not restored post-downgrade: {s['uniques']}" + ) + + _run_with_alembic_context(engine, _migration.upgrade) + + re_upgrade_shape = {t.name: _shape(engine, t.name) for t in AFFECTED_TABLES} + assert re_upgrade_shape == post_upgrade_shape, ( + "Re-upgrade shape differs from initial upgrade shape — " + "migration is not idempotent. " + f"diff: {set(re_upgrade_shape.items()) ^ set(post_upgrade_shape.items())}" + ) + + # Use pre_shape only to demonstrate it was captured (not asserted against + # because the round-trip downgrade intentionally diverges on FK NOT NULL). + _ = pre_shape + + +def test_migration_module_constants_are_consistent() -> None: + """Sanity-check the migration module's exported constants. Catches + accidental edits that misalign AFFECTED_TABLES with the auxiliary sets.""" + affected_names = {t.name for t in AFFECTED_TABLES} + assert _migration.TABLES_WITH_PRE_EXISTING_UNIQUE.issubset(affected_names) + assert _migration.TABLES_WITH_NULLABLE_FKS.issubset(affected_names) + # Order is alphabetical (deterministic for review/bisection). + assert [t.name for t in AFFECTED_TABLES] == sorted(affected_names) + + +@pytest.mark.skipif(True, reason="placeholder — see test_round_trip above") +def test_placeholder_for_future_postgres_round_trip() -> None: + """Reserved slot for a future Postgres-specific round-trip if local + SQLite divergence ever needs to be cross-checked against the real + backend. Today's CI matrix (T034a) handles this implicitly.""" diff --git a/tests/integration_tests/migrations/versioning_round_trip__tests.py b/tests/integration_tests/migrations/versioning_round_trip__tests.py new file mode 100644 index 000000000000..bef58650b351 --- /dev/null +++ b/tests/integration_tests/migrations/versioning_round_trip__tests.py @@ -0,0 +1,402 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Round-trip tests for the entity-versioning migrations +(``56cd24c07170_add_versioning_tables`` + ``8f3a1b2c4d5e_shadow_live_row_indexes``). + +The migration's ``downgrade()`` is correct against an empty schema, but +the realistic operational scenario is: we deployed, accumulated +versioning rows from real saves over hours or days, then need to roll +back. This file exercises that path against an isolated in-memory +SQLite engine via Alembic's ``MigrationContext``: + +1. Run ``56cd24c07170.upgrade()`` then ``8f3a1b2c4d5e.upgrade()``. +2. Populate every shadow table with a few rows, simulating a day's + worth of save traffic (live transactions + historical transactions + linking to the parent / child / M2M shadows). +3. Run ``8f3a1b2c4d5e.downgrade()`` then ``56cd24c07170.downgrade()``. +4. Assert every created table, index, and (where applicable) sequence + is gone. There must be no orphan rows, no lingering constraints, no + FK violations from a partial drop. +5. Re-run the upgrade and assert the post-second-upgrade shape matches + the first post-upgrade shape (idempotency: the rebuilt schema must + not differ from a brand-new install). + +The migration's ``MEDIUMTEXT`` shadow columns (sqlalchemy-review C1) +are tested implicitly — the upgrade declares them with the right +SQLAlchemy type, and any test that inserts a 64KB+ string into +``dashboards_version.position_json`` would fail on plain ``sa.Text`` +under MySQL ``STRICT_TRANS_TABLES``. SQLite ignores the length cap +(everything is TEXT), so the type-correctness test for MySQL is +delegated to the cross-backend CI matrix. The shape this file pins +is reversibility under load — the schema-correctness slice is +covered separately. + +Cross-backend verification against PostgreSQL (sequence handling) and +MySQL (composite-shadow-index dialect dispatch) is delegated to the +CI matrix. This file covers the SQLite slice — the deployment dialect +most contributors test locally against — so a regression that breaks +``downgrade()`` against populated data fails in pytest before it +fails on a production Postgres cluster at 3am. +""" + +from importlib import import_module +from typing import Any + +import sqlalchemy as sa +from alembic.migration import MigrationContext +from alembic.operations import Operations +from sqlalchemy import inspect + +_base_migration = import_module( + "superset.migrations.versions.2026-05-28_19-50_56cd24c07170_add_versioning_tables" +) +_indexes_migration = import_module( + "superset.migrations.versions.2026-06-03_12-00_8f3a1b2c4d5e_shadow_live_row_indexes" +) + + +# Tables the base migration creates, in dependency order (parents first, +# children last). All test assertions iterate this list so a regression +# that adds or removes a table here surfaces as a single edit. +_VERSIONING_TABLES: tuple[str, ...] = ( + "version_transaction", + "dashboards_version", + "slices_version", + "tables_version", + "version_changes", + "table_columns_version", + "sql_metrics_version", + "dashboard_slices_version", +) + + +# Parent + child shadow tables that carry an ``id`` column (the +# ``8f3a1b2c4d5e`` migration creates a live-row partial index over +# each one). ``dashboard_slices_version`` is intentionally excluded — +# composite PK, no ``id``. +_SHADOW_TABLES_WITH_LIVE_INDEX: tuple[str, ...] = ( + "dashboards_version", + "slices_version", + "tables_version", + "table_columns_version", + "sql_metrics_version", +) + + +# Child shadow tables that additionally carry a ``(table_id, transaction_id)`` +# composite index (the ``8f3a1b2c4d5e`` migration adds it for the dataset +# child-diff access pattern, which filters by parent ``table_id`` plus a +# transaction-range bound). Parent shadows and the M2M shadow are excluded — +# they aren't queried by ``table_id``. +_CHILD_SHADOW_TABLES_WITH_TX_INDEX: tuple[str, ...] = ( + "table_columns_version", + "sql_metrics_version", +) + + +def _run_migration( + engine: sa.engine.Engine, migration_module: Any, direction: str +) -> None: + """Execute *migration_module*'s ``upgrade()`` or ``downgrade()`` body + inside an Alembic ``MigrationContext`` bound to *engine*. + + The migrations call ``op.`` against Alembic's module-level ``op`` + singleton; this helper temporarily redirects that singleton to a + fresh ``Operations`` instance bound to our in-memory engine. Same + pattern as ``composite_pk_round_trip__tests.py`` so the two test + files don't diverge on harness style. + """ + with engine.connect() as conn: + ctx = MigrationContext.configure(conn) + ops = Operations(ctx) + original_op = migration_module.op + migration_module.op = ops # type: ignore[attr-defined] + try: + getattr(migration_module, direction)() + finally: + migration_module.op = original_op # type: ignore[attr-defined] + + +def _shape(engine: sa.engine.Engine) -> dict[str, Any]: + """Return a structural summary of all versioning-table schema + state — used to assert equality across upgrade / re-upgrade.""" + insp = inspect(engine) + all_tables = set(insp.get_table_names()) + out: dict[str, Any] = {} + for tbl in _VERSIONING_TABLES: + if tbl not in all_tables: + out[tbl] = None + continue + out[tbl] = { + "columns": sorted( + (c["name"], str(c["type"])) for c in insp.get_columns(tbl) + ), + "pk": sorted(insp.get_pk_constraint(tbl).get("constrained_columns", [])), + "indexes": sorted( + (ix["name"], tuple(ix.get("column_names", []))) + for ix in insp.get_indexes(tbl) + ), + "fks": sorted( + ( + fk["name"], + tuple(fk.get("constrained_columns", [])), + fk.get("referred_table"), + ) + for fk in insp.get_foreign_keys(tbl) + ), + } + return out + + +def _populate_shadow_rows(engine: sa.engine.Engine) -> None: + """Insert a small batch of rows into every shadow table, simulating + a day's worth of production save traffic. + + Shape: + * 3 ``version_transaction`` rows — TX 1 is "live" (open + ``end_transaction_id``); TX 2 is "closed by TX 3"; TX 3 closes + TX 2 and is itself live. + * For each parent shadow (dashboards / slices / tables): one row + per entity per transaction, with the latest row left open + (``end_transaction_id IS NULL``) and the prior row closed by it. + * One ``version_changes`` row per transaction that's tied to a + content change — the listener's typical write shape. + * One ``table_columns_version`` / ``sql_metrics_version`` row per + dataset-edit transaction. + * Two ``dashboard_slices_version`` rows — chart added + chart + removed in successive transactions. + + The point is volume + every FK exercised, not realistic semantics. + A populated downgrade that succeeds against this set succeeds + against arbitrary production volume too. + """ + with engine.begin() as conn: + # Three transactions. + conn.execute( + sa.text( + "INSERT INTO version_transaction " + "(id, issued_at, remote_addr, user_id, action_kind) VALUES " + "(1, '2026-01-01 00:00:00', NULL, NULL, NULL), " + "(2, '2026-01-02 00:00:00', NULL, NULL, 'restore'), " + "(3, '2026-01-03 00:00:00', NULL, NULL, NULL)" + ) + ) + # Dashboard 100: live row at tx=3, closed at tx=2 by tx=3. + conn.execute( + sa.text( + "INSERT INTO dashboards_version " + "(id, dashboard_title, transaction_id, end_transaction_id, " + " operation_type) VALUES " + "(100, 'Pre-restore', 2, 3, 1), " + "(100, 'Live', 3, NULL, 1)" + ) + ) + # Slice 200: live row at tx=3. + conn.execute( + sa.text( + "INSERT INTO slices_version " + "(id, slice_name, transaction_id, end_transaction_id, " + " operation_type) VALUES " + "(200, 'Live chart', 3, NULL, 1)" + ) + ) + # Dataset 300: live row at tx=3 + a column edit at tx=2. + conn.execute( + sa.text( + "INSERT INTO tables_version " + "(id, table_name, transaction_id, end_transaction_id, " + " operation_type) VALUES " + "(300, 'Pre-edit', 2, 3, 1), " + "(300, 'Live dataset', 3, NULL, 1)" + ) + ) + # Child shadows: column + metric for the dataset edit. + conn.execute( + sa.text( + "INSERT INTO table_columns_version " + "(id, column_name, table_id, transaction_id, " + " end_transaction_id, operation_type) VALUES " + "(400, 'col_a', 300, 2, 3, 1), " + "(400, 'col_a', 300, 3, NULL, 1)" + ) + ) + conn.execute( + sa.text( + "INSERT INTO sql_metrics_version " + "(id, metric_name, table_id, transaction_id, " + " end_transaction_id, operation_type) VALUES " + "(500, 'count', 300, 3, NULL, 1)" + ) + ) + # M2M: slice attached to dashboard at tx=1, still live. + conn.execute( + sa.text( + "INSERT INTO dashboard_slices_version " + "(dashboard_id, slice_id, transaction_id, " + " end_transaction_id, operation_type) VALUES " + "(100, 200, 1, NULL, 1)" + ) + ) + # Field-level change records spanning the transactions. + conn.execute( + sa.text( + "INSERT INTO version_changes " + "(transaction_id, entity_kind, entity_id, sequence, " + " kind, operation, path, from_value, to_value) VALUES " + "(2, 'dashboard', 100, 0, 'field', 'edit', " + "'[\"dashboard_title\"]', '\"Pre-restore\"', '\"Edited\"'), " + "(3, 'dashboard', 100, 0, 'field', 'edit', " + "'[\"dashboard_title\"]', '\"Edited\"', '\"Live\"'), " + "(3, 'dataset', 300, 0, 'field', 'edit', " + "'[\"table_name\"]', '\"Pre-edit\"', '\"Live dataset\"')" + ) + ) + + +def test_round_trip_against_populated_shadow_tables() -> None: + """Upgrade → populate → downgrade → upgrade-again, all against + in-memory SQLite. + + Asserts: + 1. Post-first-upgrade: all 8 versioning tables exist + all 5 partial + indexes from ``8f3a1b2c4d5e`` are present. + 2. Population step writes successfully (no FK violations against + the just-created schema). + 3. Post-downgrade: every versioning table is gone. No orphan rows + (the ``version_changes.transaction_id`` CASCADE FK does its job; + the parent/child shadows drop with their tables). + 4. Post-second-upgrade: shape matches post-first-upgrade + byte-for-byte. The migration is idempotent under round-trip; + a future operator who downgrades + re-upgrades does not end up + with a subtly different schema. + """ + engine = sa.create_engine("sqlite:///:memory:") + + # 1. Upgrade. + _run_migration(engine, _base_migration, "upgrade") + _run_migration(engine, _indexes_migration, "upgrade") + + first_upgrade_shape = _shape(engine) + for tbl in _VERSIONING_TABLES: + assert first_upgrade_shape[tbl] is not None, ( + f"Expected {tbl} to exist after upgrade; got None" + ) + + # The live-row indexes from 8f3a1b2c4d5e exist on every parent + + # child shadow (M2M shadow excluded by design). + for tbl in _SHADOW_TABLES_WITH_LIVE_INDEX: + index_names = {ix[0] for ix in first_upgrade_shape[tbl]["indexes"]} + expected = f"ix_{tbl}_live_id" + assert expected in index_names, ( + f"Expected live-id partial index {expected!r} on {tbl} after " + f"8f3a1b2c4d5e upgrade; got {sorted(index_names)}" + ) + # The M2M shadow must NOT have the live-id index. + m2m_indexes = { + ix[0] for ix in first_upgrade_shape["dashboard_slices_version"]["indexes"] + } + assert "ix_dashboard_slices_version_live_id" not in m2m_indexes, ( + "M2M shadow shouldn't get the live-id partial index (no id column)" + ) + + # The child-diff composite index exists on each child shadow. Parent + # shadows aren't queried by table_id, so they must NOT carry it. + for tbl in _CHILD_SHADOW_TABLES_WITH_TX_INDEX: + index_names = {ix[0] for ix in first_upgrade_shape[tbl]["indexes"]} + expected = f"ix_{tbl}_table_id_transaction_id" + assert expected in index_names, ( + f"Expected child-diff composite index {expected!r} on {tbl} after " + f"8f3a1b2c4d5e upgrade; got {sorted(index_names)}" + ) + for tbl in ("dashboards_version", "slices_version", "tables_version"): + parent_indexes = {ix[0] for ix in first_upgrade_shape[tbl]["indexes"]} + assert f"ix_{tbl}_table_id_transaction_id" not in parent_indexes, ( + f"Parent shadow {tbl} shouldn't get the child-diff table_id index" + ) + + # 2. Populate. + _populate_shadow_rows(engine) + + # Sanity-check: rows actually landed. + with engine.connect() as conn: + for tbl in _VERSIONING_TABLES: + # S608 justification: ``tbl`` comes from the hardcoded + # ``_VERSIONING_TABLES`` tuple in this module, never user input. + count = conn.execute( + sa.text(f"SELECT COUNT(*) FROM {tbl}") # noqa: S608 + ).scalar_one() + assert count > 0, f"Expected rows in {tbl} after population; got 0" + + # 3. Downgrade in reverse migration order. + _run_migration(engine, _indexes_migration, "downgrade") + _run_migration(engine, _base_migration, "downgrade") + + insp = inspect(engine) + surviving = set(insp.get_table_names()) + leftover = [t for t in _VERSIONING_TABLES if t in surviving] + assert not leftover, ( + f"Versioning tables survived downgrade: {leftover}. The downgrade() " + f"is supposed to drop every table created by upgrade(). Any " + f"survivor here means an operator who rolls back will be left with " + f"orphan shadow data the next upgrade attempt will collide with." + ) + + # 4. Re-upgrade and compare shapes. + _run_migration(engine, _base_migration, "upgrade") + _run_migration(engine, _indexes_migration, "upgrade") + + second_upgrade_shape = _shape(engine) + assert second_upgrade_shape == first_upgrade_shape, ( + "Schema after downgrade + re-upgrade differs from first upgrade. " + "The migration is not idempotent under round-trip; an operator " + "rolling forward after a rollback would end up with a subtly " + "different schema." + ) + + +def test_downgrade_against_empty_schema_is_safe() -> None: + """Sanity belt-and-braces: a downgrade run immediately after upgrade + (no population) must also drop everything cleanly. This catches the + case where the population step somehow influenced the downgrade path + (it should not — drops are unconditional).""" + engine = sa.create_engine("sqlite:///:memory:") + _run_migration(engine, _base_migration, "upgrade") + _run_migration(engine, _indexes_migration, "upgrade") + _run_migration(engine, _indexes_migration, "downgrade") + _run_migration(engine, _base_migration, "downgrade") + + insp = inspect(engine) + leftover = [t for t in _VERSIONING_TABLES if t in insp.get_table_names()] + assert not leftover, f"Empty-schema downgrade left {leftover}" + + +def test_indexes_migration_downgrade_is_idempotent() -> None: + """``8f3a1b2c4d5e.downgrade()`` uses ``if_exists=True`` on every + ``op.drop_index`` so a repeat call doesn't raise on missing indexes. + Operators who interrupt a downgrade mid-sequence and re-run it + rely on this property.""" + engine = sa.create_engine("sqlite:///:memory:") + _run_migration(engine, _base_migration, "upgrade") + _run_migration(engine, _indexes_migration, "upgrade") + + # Tear the indexes down once. + _run_migration(engine, _indexes_migration, "downgrade") + # Run downgrade a second time — must be a no-op, not an error. + _run_migration(engine, _indexes_migration, "downgrade") + + # Cleanup so the engine is releasable. + _run_migration(engine, _base_migration, "downgrade") diff --git a/tests/integration_tests/versioning/__init__.py b/tests/integration_tests/versioning/__init__.py new file mode 100644 index 000000000000..13a83393a912 --- /dev/null +++ b/tests/integration_tests/versioning/__init__.py @@ -0,0 +1,16 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. diff --git a/tests/integration_tests/versioning/activity_view_tests.py b/tests/integration_tests/versioning/activity_view_tests.py new file mode 100644 index 000000000000..84054389d73a --- /dev/null +++ b/tests/integration_tests/versioning/activity_view_tests.py @@ -0,0 +1,1056 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Integration tests for the cross-entity activity-view API (sc-107283). + +US1 — dashboard activity stream: ``GET /api/v1/dashboard//activity/``. +Tests for US2 (chart activity) and US3 (dataset activity) come in later +phases. + +Per spec T053 / sc-103156 T062, every test that mutates a fixture entity +wraps the test body in ``try``/``finally`` with +``metadata_db.session.rollback()`` in the ``finally``. The rationale is +documented in the spec — Continuum captures dirty mappers during +autoflush, so leaving an instrumented attribute dirty pollutes +downstream tests via the shadow tables. +""" + +from __future__ import annotations + +from typing import Any + +import pytest + +from superset.connectors.sqla.models import SqlaTable +from superset.extensions import db +from superset.models.dashboard import Dashboard +from superset.models.slice import Slice +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME, ALPHA_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _get_birth_names_dataset() -> SqlaTable: + return ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + + +def _persist_fixture_state() -> None: + """Force the fixture's pending INSERTs to commit so subsequent edits + produce *new* version rows instead of being batched into the + creation transaction. Mirrors the same helper in + ``tests/integration_tests/dashboards/version_history_tests.py``. + """ + db.session.commit() + + +def _get_birth_names_dashboard() -> Dashboard: + return ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + + +class TestDashboardActivityView(SupersetTestCase): + """T017–T026 — ``GET /api/v1/dashboard//activity/`` (US1).""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _activity(self, dashboard_uuid: str, **query: Any) -> Any: + return self.client.get( + f"/api/v1/dashboard/{dashboard_uuid}/activity/", + query_string=query, + ) + + # ---- 4xx boundary cases ---- + + def test_activity_returns_404_for_unknown_uuid(self) -> None: + """AV-009: unknown path entity → 404.""" + self.login(ADMIN_USERNAME) + rv = self._activity("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_activity_returns_400_for_invalid_uuid(self) -> None: + """A malformed UUID is rejected by the endpoint, not by Werkzeug.""" + self.login(ADMIN_USERNAME) + rv = self._activity("not-a-uuid") + assert rv.status_code == 400 + + def test_activity_returns_400_for_invalid_include(self) -> None: + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dashboard.uuid), include="sibling") + assert rv.status_code == 400 + + def test_activity_returns_400_for_invalid_since(self) -> None: + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dashboard.uuid), since="yesterday") + assert rv.status_code == 400 + + def test_activity_denies_non_owner(self) -> None: + """Mirrors sc-103156 T056 — Alpha lacks read access to the + admin-fixture dashboard, so ``raise_for_access(dashboard=)`` + rejects with 403 before the activity layer runs.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + + self.login(ALPHA_USERNAME) + rv = self._activity(dashboard_uuid) + assert rv.status_code == 403 + + # ---- 200 happy paths ---- + + def test_activity_returns_200_with_envelope_shape(self) -> None: + """Smoke test: the endpoint returns the documented envelope shape + (``result`` list + ``count`` integer) even when the dashboard has + no activity yet.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + + self.login(ADMIN_USERNAME) + rv = self._activity(dashboard_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert "result" in body + assert "count" in body + assert isinstance(body["result"], list) + assert isinstance(body["count"], int) + + def test_activity_includes_chart_edit_as_related(self) -> None: + """T018 / AS-1 of US1: editing a chart on the dashboard surfaces + the chart-edit record with ``entity_kind=Slice`` and + ``source=related``.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + chart_on_dashboard = next(iter(dashboard.slices), None) + assert chart_on_dashboard is not None + chart_id = chart_on_dashboard.id + original_name = chart_on_dashboard.slice_name + + try: + chart_on_dashboard.slice_name = f"{original_name} (edited)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dashboard_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + related = [ + r + for r in body["result"] + if r["entity_kind"] == "chart" and r["source"] == "related" + ] + assert related, ( + "Expected at least one Slice/related record from the chart " + "edit; got: " + f"{[(r['entity_kind'], r['source']) for r in body['result']]}" + ) + # Spot-check the carry-through of denormalized fields + sample = related[0] + assert sample["entity_uuid"] is not None + assert "transaction_id" in sample + assert "issued_at" in sample + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_activity_include_self_excludes_related(self) -> None: + """T023 / AV-016: ``?include=self`` filters out related records.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + chart_on_dashboard = next(iter(dashboard.slices), None) + assert chart_on_dashboard is not None + chart_id = chart_on_dashboard.id + original_name = chart_on_dashboard.slice_name + + try: + chart_on_dashboard.slice_name = f"{original_name} (edited self)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dashboard_uuid, include="self") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + for record in body["result"]: + assert record["source"] == "self", ( + f"include=self leaked a non-self record: {record}" + ) + assert record["entity_kind"] == "dashboard" + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_activity_include_related_excludes_self(self) -> None: + """T024 / AV-016: ``?include=related`` returns only related records.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + original_title = dashboard.dashboard_title + dashboard_id = dashboard.id + + try: + # Edit the dashboard's own field so we have a self record to + # filter out, and edit a chart on it so we have a related + # record to keep. + dashboard.dashboard_title = f"{original_title} (edited dash)" + db.session.commit() + chart_on_dashboard = next(iter(dashboard.slices), None) + assert chart_on_dashboard is not None + chart_id = chart_on_dashboard.id + chart_original_name = chart_on_dashboard.slice_name + chart_on_dashboard.slice_name = f"{chart_original_name} (edited chart)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dashboard_uuid, include="related") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + for record in body["result"]: + assert record["source"] == "related", ( + f"include=related leaked a self record: {record}" + ) + assert record["entity_kind"] != "dashboard" + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = chart_original_name + db.session.commit() + + def test_activity_pagination_clamps_oversized_page_size(self) -> None: + """``?page_size=500`` is silently clamped to the contract max + (200) rather than rejected with 400.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dashboard.uuid), page_size="500") + assert rv.status_code == 200 + + def test_activity_ordering_is_stable_by_issued_at_then_transaction_id(self) -> None: + """T040 / AV-006: records are ordered ``(issued_at DESC, + transaction_id DESC)``. When two records share ``issued_at`` the + tie-break is ``transaction_id`` — never random. We verify this by + asserting the result list is monotonically non-increasing on the + composite key, which would only hold under deterministic + ordering.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dashboard.uuid)) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + records = body["result"] + # Each pair of adjacent records must satisfy (prev >= cur) on the + # composite (issued_at, transaction_id) — DESC ordering. + # ``records[1:]`` is intentionally one element shorter than + # ``records``; strict=False is the correct semantic for an + # adjacent-pair iteration. + for prev, cur in zip(records, records[1:], strict=False): + assert (prev["issued_at"], prev["transaction_id"]) >= ( + cur["issued_at"], + cur["transaction_id"], + ), ( + f"Ordering broke at adjacent pair: " + f"prev=({prev['issued_at']}, {prev['transaction_id']}) " + f"cur=({cur['issued_at']}, {cur['transaction_id']})" + ) + + def test_activity_page_size_caps_returned_records_at_200(self) -> None: + """T041: ``?page_size=500`` must return *at most* 200 records. + Pairs with the no-400 check above: that test confirms the + oversized request is accepted, this test confirms the response + is bounded as the contract guarantees (AV-019 / spec + ActivityResponseSchema documentation).""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dashboard.uuid), page_size="500") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert len(body["result"]) <= 200, ( + f"page_size=500 returned {len(body['result'])} records; " + "cap is 200 per the OpenAPI schema" + ) + + def test_activity_marks_hard_deleted_chart_with_tombstone(self) -> None: + """T042 / D-15: when a chart was on the dashboard and has since + been hard-deleted, the chart's historical change records still + surface in the dashboard's activity stream, marked with + ``entity_deleted: true`` and ``entity_uuid: null``. ``entity_name`` + is preserved from the last shadow row so the UI can show + "(deleted) Girls" without a live row to query. + + Hard-delete pattern: edit the chart (creates a Slice change + record), commit, then ``db.session.delete(chart); commit``. + Continuum end-stamps the M2M row but does not cascade-delete + the shadow rows, so the history is still reachable. The + activity-view's tombstone check (``_check_entity_tombstones``) + detects the missing live row and stamps the record.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + chart_to_delete = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart_to_delete is not None + original_name = chart_to_delete.slice_name + + try: + # Step 1: generate a chart-edit change record for "Girls". + chart_to_delete.slice_name = f"{original_name} (pre-delete edit)" + db.session.commit() + + # Step 2: hard-delete the chart. The fixture's _cleanup will + # tolerate this — its `Slice.id.in_(slice_ids)` filter + # silently skips the missing row. + db.session.delete(chart_to_delete) + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dashboard_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + tombstoned = [ + r + for r in body["result"] + if r["entity_kind"] == "chart" and r["entity_deleted"] is True + ] + seen = [ + (r["entity_kind"], r["entity_deleted"]) for r in body["result"][:10] + ] + assert tombstoned, ( + "Expected ≥1 tombstoned Slice record after the chart was " + f"hard-deleted; got entity_deleted values: {seen}" + ) + sample = tombstoned[0] + got_uuid = sample["entity_uuid"] + assert got_uuid is None, ( + f"Hard-deleted entity should have null entity_uuid; got {got_uuid!r}" + ) + assert sample["entity_name"], ( + "entity_name should be recovered from the last shadow row; " + f"got empty: {sample!r}" + ) + finally: + db.session.rollback() + + def test_check_entity_tombstones_handles_multiple_kinds(self) -> None: + """Regression for the v4 indent slip in ``check_entity_tombstones``: + when called with ``distinct_entities`` spanning multiple kinds, + every kind must get its tombstone-state result, not just the + one iterated last. + + Pre-fix, the per-entity result-population block sat outside the + ``for api_kind in by_kind.items():`` loop, so all but the + last-iterated kind silently fell through to the call-site + default ``{"deleted": True}`` in ``render.apply_record_decoration`` + — live entities were rendered as tombstoned in the API response. + The previous tombstone test exercised only one kind, so dict + iteration order made the bug invisible. + """ + # pylint: disable=import-outside-toplevel + from superset.versioning.activity.queries import check_entity_tombstones + + _persist_fixture_state() + chart = db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + dataset = _get_birth_names_dataset() + assert chart is not None + assert dataset is not None + + distinct = {("Slice", chart.id), ("SqlaTable", dataset.id)} + result = check_entity_tombstones(distinct) + + assert ("Slice", chart.id) in result, ( + "Multi-kind call must populate every kind; got keys: " + f"{sorted(result.keys())}" + ) + assert ("SqlaTable", dataset.id) in result + assert result[("Slice", chart.id)] == { + "deleted": False, + "deletion_state": None, + }, f"Live chart should report deleted=False; got {result[('Slice', chart.id)]}" + assert result[("SqlaTable", dataset.id)] == { + "deleted": False, + "deletion_state": None, + }, ( + f"Live dataset should report deleted=False; " + f"got {result[('SqlaTable', dataset.id)]}" + ) + + def test_activity_excludes_records_after_retention_prune(self) -> None: + """T051 / AV-010: retention bounds the activity feed. After + ``_prune_old_versions_impl`` drops shadow / change-record rows + whose ``version_transaction.issued_at`` is older than the + retention cutoff, the activity stream stops surfacing them. + + Test pattern: capture the highest ``version_transaction.id`` + before our edits, edit a chart (creating a new transaction), + backdate that transaction's ``issued_at`` past the retention + cutoff, run the prune, and assert the chart-edit no longer + appears in the activity stream.""" + # pylint: disable=import-outside-toplevel + from datetime import datetime, timedelta + + import sqlalchemy as sa + from sqlalchemy_continuum import versioning_manager + + from superset.tasks.version_history_retention import ( + _prune_old_versions_impl, + ) + + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + chart = db.session.query(Slice).filter(Slice.slice_name == "Boys").first() + assert chart is not None + chart_id = chart.id + original_name = chart.slice_name + + tx_table = versioning_manager.transaction_cls.__table__ + + # Capture pre-edit max tx_id so we can identify the rows produced + # by THIS test (and not backdate anything else). + max_tx_before = ( + db.session.connection() + .execute(sa.select(sa.func.max(tx_table.c.id))) + .scalar() + or 0 + ) + + try: + chart.slice_name = f"{original_name} (retention test)" + db.session.commit() + + # Backdate the new transactions to before the 30-day cutoff. + old_timestamp = datetime.utcnow() - timedelta(days=60) + db.session.connection().execute( + sa.update(tx_table) + .where(tx_table.c.id > max_tx_before) + .values(issued_at=old_timestamp) + ) + db.session.commit() + + # Snapshot the activity-record count BEFORE the prune. With + # ?page_size=200 + the highest possible page coverage, the + # count field is the post-visibility filtered total. + self.login(ADMIN_USERNAME) + rv_before = self._activity(dashboard_uuid, page_size="200") + assert rv_before.status_code == 200 + count_before = _json.loads(rv_before.data.decode("utf-8"))["count"] + + # Run the prune. The backdated tx rows are now > 30 days old + # and should be deleted. AV-010 requires the prune to remove + # at least the backdated transaction(s) we created. + stats = _prune_old_versions_impl(retention_days=30) + assert stats.get("pruned_transactions", 0) >= 1, ( + f"Prune should have removed our backdated tx; stats={stats}" + ) + + # After the prune, the activity endpoint still works and the + # filtered count has DROPPED — change records joined to the + # pruned transactions are no longer in the result set (the + # join in _fetch_change_records drops them). + rv_after = self._activity(dashboard_uuid, page_size="200") + assert rv_after.status_code == 200 + count_after = _json.loads(rv_after.data.decode("utf-8"))["count"] + assert count_after < count_before, ( + f"Activity count should decrease after prune; " + f"before={count_before} after={count_after}" + ) + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_activity_pagination_is_deterministic_and_disjoint(self) -> None: + """T039 / SC-AV-002 (pragmatic interpretation): two consecutive + requests for the same page return identical results, and + consecutive pages do not overlap. + + The spec's stricter "no skip/duplicate under concurrent writes" + is unprovable with offset pagination — new top-inserted records + shift every later page by one. Cursor pagination would solve + this and is deferred per plan §D-10. Under THIS pagination + scheme, the testable guarantees are: (a) the same request fired + twice produces the same page (request determinism), and (b) + page N and page N+1 share no record under the same request + round. Both come from the stable + ``(issued_at DESC, transaction_id DESC, sequence DESC)`` sort. + """ + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + self.login(ADMIN_USERNAME) + + rv1a = self._activity(dashboard_uuid, page="0", page_size="25") + rv1b = self._activity(dashboard_uuid, page="0", page_size="25") + rv2 = self._activity(dashboard_uuid, page="1", page_size="25") + assert rv1a.status_code == 200 + assert rv1b.status_code == 200 + assert rv2.status_code == 200 + + page0_first = _json.loads(rv1a.data.decode("utf-8"))["result"] + page0_second = _json.loads(rv1b.data.decode("utf-8"))["result"] + page1 = _json.loads(rv2.data.decode("utf-8"))["result"] + + # (a) Request determinism: same page twice → same records in same + # order. Use (entity_kind, entity_id_internal_proxy, tx, seq) + # fingerprint — entity_uuid + transaction_id is sufficient + # since entity_id isn't in the API contract. + fingerprint = lambda r: ( # noqa: E731 + r["entity_kind"], + r["entity_uuid"], + r["transaction_id"], + r["kind"], + tuple(r["path"]) if r["path"] else (), + ) + assert [fingerprint(r) for r in page0_first] == [ + fingerprint(r) for r in page0_second + ], "page=0 fired twice returned different records" + + # (b) Page 0 and page 1 are disjoint under one request round. + page0_keys = {fingerprint(r) for r in page0_first} + page1_keys = {fingerprint(r) for r in page1} + overlap = page0_keys & page1_keys + assert not overlap, f"page=0 and page=1 returned overlapping records: {overlap}" + + @pytest.mark.xfail( + reason=( + "AV-015 requires sc-103156's restore code to emit a synthetic " + "change record with kind='restore', path=['__meta__', " + "'restored_from'], and to_value carrying the source version_uuid " + "+ label. sc-103156's restore_version() currently does not emit " + "this — it relies on the diff capture for the field changes the " + "revert produces, which surface as kind='field' records. The " + "activity-view layer correctly passes through whatever kind " + "sc-103156 emits; this test will pass once the upstream " + "emission lands. Tracking via the AV-015 contract in the spec; " + "no code change required on the sc-107283 side." + ), + strict=True, + ) + def test_activity_surfaces_dashboard_restore_event(self) -> None: + """T044 / AV-015: restoring a dashboard to a prior version surfaces + a ``kind='restore'`` record in the dashboard's own activity stream + (``source='self'``). The restore is emitted by sc-103156's restore + path and the activity layer passes it through without special- + casing.""" + _persist_fixture_state() + dashboard = _get_birth_names_dashboard() + assert dashboard is not None + dashboard_uuid = str(dashboard.uuid) + dashboard_id = dashboard.id + original_title = dashboard.dashboard_title + + try: + # Two edits → at least two restorable prior versions. + dashboard.dashboard_title = f"{original_title} v1" + db.session.commit() + dashboard.dashboard_title = f"{original_title} v2" + db.session.commit() + + self.login(ADMIN_USERNAME) + # Find a prior version to restore to (version_number 0 is the + # baseline; we restore to whichever earlier version the list + # endpoint surfaces). + versions_rv = self.client.get( + f"/api/v1/dashboard/{dashboard_uuid}/versions/" + ) + assert versions_rv.status_code == 200, versions_rv.data + versions = _json.loads(versions_rv.data.decode("utf-8"))["result"] + assert len(versions) >= 2, f"expected ≥2 versions, got {versions}" + target_version_uuid = versions[0]["version_uuid"] # earliest + + # Restore. The endpoint commits; finally clean up below. + restore_rv = self.client.post( + f"/api/v1/dashboard/{dashboard_uuid}" + f"/versions/{target_version_uuid}/restore" + ) + assert restore_rv.status_code == 200, restore_rv.data + + # Activity stream should now show a restore record on the + # dashboard itself. + rv = self._activity(dashboard_uuid, include="self") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + restore_records = [ + r + for r in body["result"] + if r["kind"] == "restore" and r["entity_kind"] == "dashboard" + ] + assert restore_records, ( + "Expected at least one kind='restore' Dashboard record; " + f"got kinds: {[r['kind'] for r in body['result'][:10]]}" + ) + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + +class TestChartActivityView(SupersetTestCase): + """T028–T032 — ``GET /api/v1/chart//activity/`` (US2). + + Chart activity = chart's own edits + datasets the chart pointed at + during association. **No** dashboard records — even when the chart + is on a dashboard, sibling-traversal is excluded per the spec's + Relationship Traversal section (T032). + """ + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _activity(self, chart_uuid: str, **query: Any) -> Any: + return self.client.get( + f"/api/v1/chart/{chart_uuid}/activity/", + query_string=query, + ) + + def _get_birth_names_chart(self) -> Slice: + return db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + + # ---- 4xx boundary cases ---- + + def test_chart_activity_returns_404_for_unknown_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._activity("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_chart_activity_returns_400_for_invalid_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._activity("not-a-uuid") + assert rv.status_code == 400 + + def test_chart_activity_returns_400_for_invalid_include(self) -> None: + _persist_fixture_state() + chart = self._get_birth_names_chart() + assert chart is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(chart.uuid), include="upstream") + assert rv.status_code == 400 + + def test_chart_activity_denies_non_owner(self) -> None: + """Same shape as the dashboard endpoint: Alpha lacks read access + to the admin-fixture chart so ``raise_for_access(chart=)`` + returns 403.""" + _persist_fixture_state() + chart = self._get_birth_names_chart() + assert chart is not None + self.login(ALPHA_USERNAME) + rv = self._activity(str(chart.uuid)) + assert rv.status_code == 403 + + # ---- 200 happy paths ---- + + def test_chart_activity_returns_200_with_envelope_shape(self) -> None: + _persist_fixture_state() + chart = self._get_birth_names_chart() + assert chart is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(chart.uuid)) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert isinstance(body["result"], list) + assert isinstance(body["count"], int) + + def test_chart_activity_self_edit_appears_as_self_record(self) -> None: + """Editing the chart itself surfaces a ``source=self``, + ``entity_kind=Slice`` record.""" + _persist_fixture_state() + chart = self._get_birth_names_chart() + assert chart is not None + chart_id = chart.id + chart_uuid = str(chart.uuid) + original_name = chart.slice_name + + try: + chart.slice_name = f"{original_name} (edited self)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(chart_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + self_records = [ + r + for r in body["result"] + if r["entity_kind"] == "chart" and r["source"] == "self" + ] + got = [(r["entity_kind"], r["source"]) for r in body["result"]] + assert self_records, ( + f"Expected ≥1 Slice/self record from the chart edit; got: {got}" + ) + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = original_name + db.session.commit() + + def test_chart_activity_includes_dataset_edit_as_related(self) -> None: + """T030 / AS-1 of US2: editing the chart's dataset surfaces a + ``source=related``, ``entity_kind=SqlaTable`` record.""" + _persist_fixture_state() + chart = self._get_birth_names_chart() + dataset = _get_birth_names_dataset() + assert chart is not None + assert dataset is not None + chart_uuid = str(chart.uuid) + dataset_id = dataset.id + original_description = dataset.description + + try: + dataset.description = "edited for activity-view test" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(chart_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + related = [ + r + for r in body["result"] + if r["entity_kind"] == "dataset" and r["source"] == "related" + ] + assert related, ( + "Expected at least one SqlaTable/related record from the " + "dataset edit; got: " + f"{[(r['entity_kind'], r['source']) for r in body['result']]}" + ) + finally: + db.session.rollback() + dataset = ( + db.session.query(SqlaTable).filter(SqlaTable.id == dataset_id).one() + ) + dataset.description = original_description + db.session.commit() + + def test_chart_activity_excludes_sibling_dashboards(self) -> None: + """T032: Even when the chart is on a dashboard, dashboard edits + do NOT appear in the chart's activity. Per the spec's Relationship + Traversal section: charts don't see "sideways" to the dashboards + they happen to be on.""" + _persist_fixture_state() + chart = self._get_birth_names_chart() + dashboard = _get_birth_names_dashboard() + assert chart is not None + assert dashboard is not None + chart_uuid = str(chart.uuid) + dashboard_id = dashboard.id + original_title = dashboard.dashboard_title + + try: + # Mutate the dashboard the chart is on — that edit MUST NOT + # appear in the chart's activity stream. + dashboard.dashboard_title = f"{original_title} (edited sibling)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(chart_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + for record in body["result"]: + assert record["entity_kind"] != "dashboard", ( + f"Dashboard edit leaked into chart's activity stream: {record}" + ) + finally: + db.session.rollback() + dashboard = ( + db.session.query(Dashboard).filter(Dashboard.id == dashboard_id).one() + ) + dashboard.dashboard_title = original_title + db.session.commit() + + def test_chart_activity_include_self_excludes_related(self) -> None: + """``?include=self`` filters out the dataset records.""" + _persist_fixture_state() + chart = self._get_birth_names_chart() + dataset = _get_birth_names_dataset() + assert chart is not None + assert dataset is not None + chart_uuid = str(chart.uuid) + dataset_id = dataset.id + original_description = dataset.description + + try: + dataset.description = "edited (self filter test)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(chart_uuid, include="self") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + for record in body["result"]: + assert record["source"] == "self" + assert record["entity_kind"] == "chart" + finally: + db.session.rollback() + dataset = ( + db.session.query(SqlaTable).filter(SqlaTable.id == dataset_id).one() + ) + dataset.description = original_description + db.session.commit() + + +class TestDatasetActivityView(SupersetTestCase): + """T033–T036 — ``GET /api/v1/dataset//activity/`` (US3). + + Dataset activity = dataset's own edits only. **No** transitive layer + in V2 (AV-004) — even when charts use the dataset, those chart edits + do NOT appear here. ``?include=related`` and ``?include=all`` + collapse to the same self-only stream as ``?include=self``. + """ + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _activity(self, dataset_uuid: str, **query: Any) -> Any: + return self.client.get( + f"/api/v1/dataset/{dataset_uuid}/activity/", + query_string=query, + ) + + # ---- 4xx boundary cases ---- + + def test_dataset_activity_returns_404_for_unknown_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._activity("00000000-0000-0000-0000-000000000000") + assert rv.status_code == 404 + + def test_dataset_activity_returns_400_for_invalid_uuid(self) -> None: + self.login(ADMIN_USERNAME) + rv = self._activity("not-a-uuid") + assert rv.status_code == 400 + + def test_dataset_activity_returns_400_for_invalid_include(self) -> None: + _persist_fixture_state() + dataset = _get_birth_names_dataset() + assert dataset is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dataset.uuid), include="upstream") + assert rv.status_code == 400 + + def test_dataset_activity_denies_non_owner(self) -> None: + _persist_fixture_state() + dataset = _get_birth_names_dataset() + assert dataset is not None + self.login(ALPHA_USERNAME) + rv = self._activity(str(dataset.uuid)) + assert rv.status_code == 403 + + # ---- 200 happy paths ---- + + def test_dataset_activity_returns_200_with_envelope_shape(self) -> None: + _persist_fixture_state() + dataset = _get_birth_names_dataset() + assert dataset is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dataset.uuid)) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert isinstance(body["result"], list) + assert isinstance(body["count"], int) + + def test_dataset_activity_includes_dataset_self_edits(self) -> None: + """T036: the dataset's own scalar edits appear as ``source=self``, + ``entity_kind=SqlaTable``.""" + _persist_fixture_state() + dataset = _get_birth_names_dataset() + assert dataset is not None + dataset_id = dataset.id + dataset_uuid = str(dataset.uuid) + original_description = dataset.description + + try: + dataset.description = "edited self for dataset activity" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dataset_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + self_records = [ + r + for r in body["result"] + if r["entity_kind"] == "dataset" and r["source"] == "self" + ] + got = [(r["entity_kind"], r["source"]) for r in body["result"]] + assert self_records, ( + f"Expected ≥1 SqlaTable/self record from the dataset edit; got: {got}" + ) + finally: + db.session.rollback() + dataset = ( + db.session.query(SqlaTable).filter(SqlaTable.id == dataset_id).one() + ) + dataset.description = original_description + db.session.commit() + + def test_dataset_activity_excludes_chart_edits(self) -> None: + """T035 / AS-1 / AV-004: When a chart that uses the dataset is + edited, that edit does NOT appear in the dataset's activity stream. + Datasets are read-only upstream in V2.""" + _persist_fixture_state() + dataset = _get_birth_names_dataset() + chart = db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + assert dataset is not None + assert chart is not None + dataset_uuid = str(dataset.uuid) + chart_id = chart.id + chart_original_name = chart.slice_name + + try: + # Edit the chart — generates a Slice change record. The + # dataset's activity MUST NOT surface it. + chart.slice_name = f"{chart_original_name} (edited from dataset test)" + db.session.commit() + + self.login(ADMIN_USERNAME) + rv = self._activity(dataset_uuid) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + for record in body["result"]: + assert record["entity_kind"] == "dataset", ( + "Non-dataset record leaked into dataset's activity " + f"stream: {record}" + ) + assert record["source"] == "self", ( + f"Dataset activity contains a related record: {record}" + ) + finally: + db.session.rollback() + chart = db.session.query(Slice).filter(Slice.id == chart_id).one() + chart.slice_name = chart_original_name + db.session.commit() + + def test_dataset_activity_related_only_returns_empty(self) -> None: + """AV-004: datasets have no transitive layer. ``?include=related`` + returns an empty result list because there are no related entities + to draw from.""" + _persist_fixture_state() + dataset = _get_birth_names_dataset() + assert dataset is not None + self.login(ADMIN_USERNAME) + rv = self._activity(str(dataset.uuid), include="related") + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["result"] == [] + assert body["count"] == 0 + + +class TestActivityOpenApiSpec(SupersetTestCase): + """T049 — confirm the three ``/activity/`` endpoints are surfaced by + FAB-generated OpenAPI at ``/api/v1/_openapi``. + + ``base_api_tests.py::TestOpenApiSpec::test_open_api_spec`` already + validates the full spec's YAML correctness on every CI run. This + class adds activity-specific assertions: the paths exist, are + documented with the expected query parameters, and reference an + ``ActivityResponse``-shaped 200 response. + """ + + def _spec(self) -> dict[str, Any]: + self.login(ADMIN_USERNAME) + rv = self.client.get("/api/v1/_openapi") + assert rv.status_code == 200, rv.status_code + return _json.loads(rv.data.decode("utf-8")) + + def test_three_activity_paths_appear_in_openapi(self) -> None: + """One path per endpoint family. Paths are keyed by the URL + template, not the method name, so the FAB-generated keys are + the ``//activity/`` route templates.""" + spec = self._spec() + paths = spec.get("paths", {}) + # FAB templates the path-arg as ``{uuid_str}`` in the OpenAPI dict. + expected = { + "/api/v1/dashboard/{uuid_str}/activity/", + "/api/v1/chart/{uuid_str}/activity/", + "/api/v1/dataset/{uuid_str}/activity/", + } + missing = expected - paths.keys() + assert not missing, f"missing activity paths in OpenAPI: {missing}" + + def test_activity_endpoints_document_query_params(self) -> None: + """Each endpoint declares since / until / include / page / + page_size as query parameters. Spot-check on the dashboard + endpoint — the YAML docstring is the same shape across all + three so this assertion is sufficient.""" + spec = self._spec() + op = spec["paths"]["/api/v1/dashboard/{uuid_str}/activity/"]["get"] + params = {p["name"]: p for p in op.get("parameters", [])} + for expected in ("since", "until", "include", "page", "page_size"): + assert expected in params, ( + f"query param {expected!r} missing from dashboard /activity/" + ) + # include enum is the published contract — verify it's correct. + include_param = params["include"] + assert include_param["in"] == "query" + assert set(include_param["schema"]["enum"]) == {"self", "related", "all"} + + def test_activity_endpoints_declare_200_response(self) -> None: + """Each endpoint declares a 200 response. The exact schema + reference depends on how FAB resolves ``schema: ActivityResponseSchema`` + in the YAML docstring; here we just confirm the 200 + the 4xx + error responses are all present.""" + spec = self._spec() + op = spec["paths"]["/api/v1/dashboard/{uuid_str}/activity/"]["get"] + responses = op.get("responses", {}) + for code in ("200", "400", "401", "403", "404"): + assert code in responses, ( + f"response code {code} missing on dashboard /activity/" + ) diff --git a/tests/integration_tests/versioning/change_records_tests.py b/tests/integration_tests/versioning/change_records_tests.py new file mode 100644 index 000000000000..7f7ae604cde8 --- /dev/null +++ b/tests/integration_tests/versioning/change_records_tests.py @@ -0,0 +1,650 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Integration tests for ``version_changes`` capture (T052, partial). + +Covers in this file: + (a) saving a chart with three field changes produces three rows + (f) baseline / INSERT transactions produce zero records *for that entity* + + unchanged-save / dashboard / params-classification cases + +Deferred: + (b) ``GET /versions/`` response includes ``changes`` array — lands with + T050 (API integration). + (c) FK cascade — exercisable in principle (the migration declares + ``ON DELETE CASCADE``) but can't be isolated in a unit-style test + because ``version_transaction`` is referenced by non-cascading FKs + from slices_version / dashboards_version / etc. Covered instead + by (d) below once it lands, and by the structural declaration in + T046's migration. + (d) retention prune drops change records alongside the pruned + version — will land when T049 extends ``VersionDAO.prune_versions`` + to include ``version_changes`` alongside the shadow-row delete. + (e) ``kind`` index query plan on Postgres — deferred to T053 perf + validation. +""" + +from __future__ import annotations + +from datetime import datetime, timedelta +from typing import Any + +import pytest +import sqlalchemy as sa +from sqlalchemy_continuum import version_class + +from superset.extensions import db +from superset.models.dashboard import Dashboard +from superset.models.slice import Slice +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + +_VERSION_CHANGES = sa.table( + "version_changes", + sa.column("id"), + sa.column("transaction_id"), + sa.column("entity_kind"), + sa.column("entity_id"), + sa.column("sequence"), + sa.column("kind"), + sa.column("operation"), + sa.column("path"), + sa.column("from_value"), + sa.column("to_value"), +) + +_VERSION_TRANSACTION = sa.table( + "version_transaction", + sa.column("id"), + sa.column("issued_at"), + sa.column("user_id"), + sa.column("action_kind"), +) + + +def _action_kind_for(tx_id: int) -> str | None: + """Read the ``action_kind`` column from the version_transaction row.""" + return ( + db.session.connection() + .execute( + sa.select(_VERSION_TRANSACTION.c.action_kind).where( + _VERSION_TRANSACTION.c.id == tx_id + ) + ) + .scalar() + ) + + +def _change_rows_for( + tx_id: int, + *, + entity_kind: str | None = None, + entity_id: int | None = None, +) -> list[dict[str, Any]]: + """Raw fetch of ``version_changes`` rows for a tx + optional entity filter.""" + query = sa.select(_VERSION_CHANGES).where( + _VERSION_CHANGES.c.transaction_id == tx_id + ) + if entity_kind is not None: + query = query.where(_VERSION_CHANGES.c.entity_kind == entity_kind) + if entity_id is not None: + query = query.where(_VERSION_CHANGES.c.entity_id == entity_id) + query = query.order_by(_VERSION_CHANGES.c.sequence.asc()) + result = db.session.connection().execute(query) + return [dict(row._mapping) for row in result] + + +def _persist_fixture_state() -> None: + """Commit fixture INSERTs so the baseline row exists before the test edits. + + Without this, the test's first commit batches the fixture's pending + INSERTs with the test's UPDATE into a single Continuum transaction + and no diff records are emitted (no pre-state). + """ + db.session.commit() + + +class TestChartChangeRecords(SupersetTestCase): + """Change-record capture for chart (Slice) saves.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: F811, PT004 + pass + + def test_single_scalar_edit_produces_one_change_record(self) -> None: + """(a) — one field changed, one ``version_changes`` row.""" + _persist_fixture_state() + + chart = db.session.query(Slice).first() + assert chart is not None + chart.slice_name = f"{chart.slice_name[:64]}_renamed" + db.session.commit() + + # The save produces one new version row (the UPDATE). Fetch its tx_id. + ver_cls = version_class(Slice) + update_tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + + rows = _change_rows_for(update_tx_id, entity_kind="chart", entity_id=chart.id) + assert len(rows) == 1 + assert rows[0]["kind"] == "field" + path = ( + _json.loads(rows[0]["path"]) + if isinstance(rows[0]["path"], str) + else rows[0]["path"] + ) + assert path == ["slice_name"] + assert rows[0]["sequence"] == 0 + + def test_last_saved_at_is_excluded_as_audit_noise(self) -> None: + """``last_saved_at`` / ``last_saved_by_fk`` are save-side-effect + fields stamped by ``UpdateChartCommand`` and must not produce + change records — same category as ``changed_on``. + + Saving a chart with ONLY a ``last_saved_at`` bump must produce + zero ``version_changes`` rows for that transaction. (Continuum + still records the shadow row; we just don't want to noise up + the per-edit diff log.) + """ + _persist_fixture_state() + + chart = db.session.query(Slice).first() + assert chart is not None + chart.last_saved_at = datetime.now() + timedelta(seconds=1) + db.session.commit() + + ver_cls = version_class(Slice) + latest_tx = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + ) + # If the save produced no version row at all (no actual model + # change beyond the audit field), nothing to assert. If it did, + # there must be no ``last_saved_at`` row in version_changes. + if latest_tx is None: + return + rows = _change_rows_for( + latest_tx.transaction_id, entity_kind="chart", entity_id=chart.id + ) + paths = [ + _json.loads(r["path"]) if isinstance(r["path"], str) else r["path"] + for r in rows + ] + assert ["last_saved_at"] not in paths + assert ["last_saved_by_fk"] not in paths + + def test_three_scalar_edits_produce_three_records_in_sequence(self) -> None: + """(a) — three fields changed, three rows, ``sequence`` 0..2.""" + _persist_fixture_state() + + chart = db.session.query(Slice).first() + assert chart is not None + # Derive from CURRENT values so every run guarantees a real + # change even against a persistent test DB where prior runs + # have already mutated the chart. + chart.slice_name = f"{chart.slice_name[:60]}_x" + chart.description = f"{chart.description or ''}_x" + chart.cache_timeout = (chart.cache_timeout or 0) + 1 + db.session.commit() + + ver_cls = version_class(Slice) + update_tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + rows = _change_rows_for(update_tx_id, entity_kind="chart", entity_id=chart.id) + assert len(rows) == 3 + assert [r["sequence"] for r in rows] == [0, 1, 2] + # Sorted by field name (diff engine emits in sorted field order) + paths = [ + _json.loads(r["path"]) if isinstance(r["path"], str) else r["path"] + for r in rows + ] + assert paths == [["cache_timeout"], ["description"], ["slice_name"]] + + def test_params_filter_add_produces_filter_kind_record(self) -> None: + """(a) — params classification still flows through the listener. + + Adds an adhoc_filter with a natural key (``subject``) derived + from the chart id so it's unique across test runs on a + persistent DB. Whatever was in ``adhoc_filters`` before stays; + we only want to confirm at least one ``kind='filter'`` record + is emitted. + """ + _persist_fixture_state() + + chart = db.session.query(Slice).first() + assert chart is not None + unique_subject = ( + f"col_{chart.id}_{db.session.connection().engine.url.database[-8:]}" + ) + params = _json.loads(chart.params or "{}") + existing = params.get("adhoc_filters", []) or [] + params["adhoc_filters"] = [ + *existing, + { + "subject": unique_subject, + "operator": "==", + "comparator": "x", + "expressionType": "SIMPLE", + }, + ] + chart.params = _json.dumps(params) + db.session.commit() + + ver_cls = version_class(Slice) + update_tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + rows = _change_rows_for(update_tx_id, entity_kind="chart", entity_id=chart.id) + filter_rows = [r for r in rows if r["kind"] == "filter"] + assert len(filter_rows) >= 1, ( + f"expected at least one filter record, got rows: {rows}" + ) + + def test_unchanged_save_produces_zero_change_records(self) -> None: + """An edit that sets fields to identical values emits nothing.""" + _persist_fixture_state() + + chart = db.session.query(Slice).first() + ver_cls = version_class(Slice) + # Capture the latest tx_id BEFORE this test's save so we can + # distinguish "the no-op save produced nothing new" (the intent) + # from "prior tests left tx rows with records on them" (noise). + pre_save_tx_row = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + ) + pre_save_tx_id = pre_save_tx_row.transaction_id if pre_save_tx_row else 0 + + # Touch the object (mark dirty) but assign the same value. + current_name = chart.slice_name + chart.slice_name = current_name + db.session.commit() + + post_save_tx_row = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .filter(ver_cls.transaction_id > pre_save_tx_id) + .order_by(ver_cls.transaction_id.desc()) + .first() + ) + # Either no new tx at all (nothing dirty, best case), or a new + # tx with zero change records for this chart. + if post_save_tx_row is not None: + assert ( + _change_rows_for( + post_save_tx_row.transaction_id, + entity_kind="chart", + entity_id=chart.id, + ) + == [] + ) + + +class TestDashboardChangeRecords(SupersetTestCase): + """Same flow for dashboards — all scalar fields land in ``kind='field'``.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: F811, PT004 + pass + + def test_dashboard_title_edit_produces_field_record(self) -> None: + _persist_fixture_state() + + dashboard = db.session.query(Dashboard).first() + assert dashboard is not None + dashboard.dashboard_title = f"{dashboard.dashboard_title}_rev" + db.session.commit() + + ver_cls = version_class(Dashboard) + update_tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == dashboard.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + rows = _change_rows_for( + update_tx_id, entity_kind="dashboard", entity_id=dashboard.id + ) + assert len(rows) >= 1 + field_rows = [r for r in rows if r["kind"] == "field"] + paths = [ + _json.loads(r["path"]) if isinstance(r["path"], str) else r["path"] + for r in field_rows + ] + assert ["dashboard_title"] in paths + + +class TestDatasetChildChangeRecords(SupersetTestCase): + """T048b — column and metric diff records for dataset saves. + + Two snapshots must exist for any child diff to emit: the prior + save's and the current one. The fixture ``load_birth_names_data`` + has already created the dataset before these tests run; their + first commit produces snapshot #1. The test's edit produces + snapshot #2, and the listener diffs the two. + """ + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: F811, PT004 + pass + + def test_column_description_change_produces_column_record(self) -> None: + # pylint: disable=import-outside-toplevel + from sqlalchemy_continuum import version_class + + from superset.connectors.sqla.models import SqlaTable + + _persist_fixture_state() + + dataset = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert dataset is not None + assert dataset.columns, "birth_names fixture should produce columns" + # First save establishes snapshot #1 (the pre-edit state). + # Scalar + child diffs won't emit anything yet because there's + # no prior snapshot to diff against. + dataset.description = f"{dataset.description or ''}_v1" + db.session.commit() + # Second save: edit a column AND touch a dataset scalar so + # the parent SqlaTable ends up in session.dirty. In real + # flows DatasetDAO.update_columns() marks the parent via its + # individual session.add / session.delete calls (T011); the + # direct-ORM test here needs an explicit parent touch. + column = dataset.columns[0] + column.description = f"{column.description or ''}_edited" + dataset.description = f"{dataset.description}_v2" + db.session.commit() + + ver_cls = version_class(SqlaTable) + latest_tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == dataset.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + rows = _change_rows_for( + latest_tx_id, entity_kind="dataset", entity_id=dataset.id + ) + column_rows = [r for r in rows if r["kind"] == "column"] + assert len(column_rows) >= 1, ( + f"expected at least one kind='column' record, got {rows}" + ) + + +class TestBaselineProducesZeroChangeRecords(SupersetTestCase): + """(f) — operation_type=0 (baseline / INSERT) transactions emit no records.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: F811, PT004 + pass + + def test_baseline_transaction_has_no_change_records_for_this_entity( + self, + ) -> None: + """(f) — baseline tx produces zero records *for that entity*. + + A single transaction can touch multiple entities (fixture loads, + import pipelines). A tx that's a baseline for this chart might + still legitimately carry update records for some *other* entity + that shared the flush. The spec's M4 clarification means: + records filtered to this entity's (tx, entity_kind, entity_id) + are empty for its baseline tx. + """ + _persist_fixture_state() + + chart = db.session.query(Slice).first() + chart.slice_name = f"{chart.slice_name[:64]}_force_baseline" + db.session.commit() + + ver_cls = version_class(Slice) + rows_by_tx = ( + db.session.query(ver_cls.transaction_id, ver_cls.operation_type) + .filter(ver_cls.id == chart.id) + .order_by(ver_cls.transaction_id.asc()) + .all() + ) + baseline_tx_ids = [tx for tx, op in rows_by_tx if op == 0] + assert baseline_tx_ids, "expected at least one baseline version row" + + for tx_id in baseline_tx_ids: + records_for_this_chart = _change_rows_for( + tx_id, entity_kind="chart", entity_id=chart.id + ) + assert records_for_this_chart == [], ( + f"baseline tx {tx_id} unexpectedly has change records for " + f"chart id={chart.id}: {records_for_this_chart}" + ) + + +class TestTransactionActionKindPropagation(SupersetTestCase): + """Confirm ``version_transaction.action_kind`` is stamped when a + command declares one via ``session.info["_versioning_action_kind"]``, + and stays ``NULL`` on ordinary saves.""" + + @pytest.mark.usefixtures("load_birth_names_dashboard_with_slices") + def test_ordinary_save_has_null_action_kind(self) -> None: + """No command sets the key → version_transaction.action_kind + is NULL for a normal user-initiated save.""" + from superset.versioning.changes import ACTION_KIND_KEY + + _persist_fixture_state() + # Sanity: the key shouldn't already be on the session. + assert ACTION_KIND_KEY not in db.session.info + + chart = db.session.query(Slice).first() + assert chart is not None + chart.slice_name = f"{chart.slice_name[:60]}_baseline" + db.session.commit() + + ver_cls = version_class(Slice) + tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + assert _action_kind_for(tx_id) is None + + @pytest.mark.usefixtures("load_birth_names_dashboard_with_slices") + def test_session_info_action_kind_propagates_to_transaction(self) -> None: + """The listener reads ``session.info[ACTION_KIND_KEY]`` and + stamps it on the version_transaction row. Exercises the wiring + directly so we don't need a full end-to-end command run for the + propagation test (the per-command tests below cover the + calling side).""" + from superset.versioning.changes import ACTION_KIND_KEY + + _persist_fixture_state() + chart = db.session.query(Slice).first() + assert chart is not None + + db.session.info[ACTION_KIND_KEY] = "restore" + chart.slice_name = f"{chart.slice_name[:60]}_trig" + db.session.commit() + + ver_cls = version_class(Slice) + tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + assert _action_kind_for(tx_id) == "restore" + + # And: the key is popped — next save resets to NULL action_kind. + assert ACTION_KIND_KEY not in db.session.info + + @pytest.mark.usefixtures("load_birth_names_dashboard_with_slices") + def test_action_kind_pops_so_next_save_is_clean(self) -> None: + """After the listener stamps the action_kind, subsequent saves + on the same session must not carry it forward.""" + from superset.versioning.changes import ACTION_KIND_KEY + + _persist_fixture_state() + chart = db.session.query(Slice).first() + assert chart is not None + + # First save with action_kind. + db.session.info[ACTION_KIND_KEY] = "import" + chart.slice_name = f"{chart.slice_name[:60]}_a" + db.session.commit() + + # Second save without setting the key. + chart.slice_name = f"{chart.slice_name[:60]}_b" + db.session.commit() + + ver_cls = version_class(Slice) + # Get the two most-recent edit tx_ids. + rows = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .limit(2) + .all() + ) + assert len(rows) == 2 + second_tx, first_tx = rows[0].transaction_id, rows[1].transaction_id + + assert _action_kind_for(first_tx) == "import" + assert _action_kind_for(second_tx) is None + + @pytest.mark.usefixtures("load_birth_names_dashboard_with_slices") + def test_action_kind_dropped_on_rollback(self) -> None: + """When a command sets ACTION_KIND_KEY and then an exception + fires before any flush stamps it (e.g. validation error after + the key is set), the value must not leak into the next save on + the same session. Regression for sqlalchemy-review C3.""" + from superset.versioning.changes import ACTION_KIND_KEY + + _persist_fixture_state() + chart = db.session.query(Slice).first() + assert chart is not None + + # Declare an action_kind, then force a rollback before the + # listener's flush stamps it. + db.session.info[ACTION_KIND_KEY] = "restore" + db.session.rollback() + + # The after_rollback listener must have popped the key. + assert ACTION_KIND_KEY not in db.session.info + + # And: a normal save now records NULL action_kind, not "restore". + chart.slice_name = f"{chart.slice_name[:60]}_postrollback" + db.session.commit() + + ver_cls = version_class(Slice) + tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + assert _action_kind_for(tx_id) is None + + @pytest.mark.usefixtures("load_birth_names_dashboard_with_slices") + def test_action_kind_survives_multiple_flushes_in_one_transaction(self) -> None: + """A restore/import can flush more than once before commit (an + explicit ``flush()`` or an autoflush from a mid-commit query). The + listener stamps ``action_kind`` *eagerly* on the first firing and + then dedups later firings of the same Continuum tx via the + ``_PROCESSED_TXS_KEY`` guard. + + This pins the eager-stamp + dedup interaction: across two flushes + in a single transaction the action_kind lands exactly once on the + one tx that carries the change records, the key is popped (so it + can't leak to the next save), and the second flush does not + re-emit records (which would trip the UNIQUE(transaction_id, + entity_kind, entity_id, sequence) constraint). Regression for the + amin-review finding on stamp-before-short-circuit ordering. + """ + from superset.versioning.changes import ACTION_KIND_KEY + + _persist_fixture_state() + chart = db.session.query(Slice).first() + assert chart is not None + + # One transaction, two flushes: declare the action_kind, edit + + # flush (first after_flush firing stamps and persists), then edit + # again + commit (second firing for the same tx must short-circuit). + db.session.info[ACTION_KIND_KEY] = "restore" + chart.slice_name = f"{chart.slice_name[:60]}_f1" + db.session.flush() + chart.slice_name = f"{chart.slice_name[:60]}_f2" + db.session.commit() + + ver_cls = version_class(Slice) + tx_id = ( + db.session.query(ver_cls.transaction_id) + .filter(ver_cls.id == chart.id) + .filter(ver_cls.operation_type == 1) + .order_by(ver_cls.transaction_id.desc()) + .first() + .transaction_id + ) + + # Stamped exactly once on the records-bearing tx, and popped. + assert _action_kind_for(tx_id) == "restore" + assert ACTION_KIND_KEY not in db.session.info + # Records exist for the tx and the dedup guard prevented a + # duplicate-sequence re-emit on the second flush. + rows = _change_rows_for(tx_id, entity_kind="chart", entity_id=chart.id) + assert rows, "expected change records on the multi-flush transaction" + sequences = [r["sequence"] for r in rows] + assert len(sequences) == len(set(sequences)), ( + f"duplicate sequences on tx {tx_id}: {sequences}" + ) diff --git a/tests/integration_tests/versioning/etag_tests.py b/tests/integration_tests/versioning/etag_tests.py new file mode 100644 index 000000000000..155a5debc72d --- /dev/null +++ b/tests/integration_tests/versioning/etag_tests.py @@ -0,0 +1,184 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""T055 — ``ETag`` header emission on entity GETs / PUTs / version endpoints.""" + +from __future__ import annotations + +import pytest + +from superset.connectors.sqla.models import SqlaTable +from superset.daos.version import VersionDAO +from superset.extensions import db +from superset.models.dashboard import Dashboard +from superset.models.slice import Slice +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _expected_etag(model_cls: type, entity_id: int, entity_uuid) -> str: + version_uuid = VersionDAO.current_live_version_uuid( + model_cls, entity_id, entity_uuid + ) + return f'"{version_uuid}"' + + +class TestETagEmission(SupersetTestCase): + """ETag header on entity detail, save response, and version endpoints.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def test_chart_get_emits_etag_matching_current_live_version(self) -> None: + db.session.commit() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + expected = _expected_etag(Slice, chart.id, chart.uuid) + + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/chart/{chart.id}") + assert rv.status_code == 200 + assert rv.headers.get("ETag") == expected + + def test_chart_put_emits_etag_matching_new_live_version(self) -> None: + db.session.commit() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_id = chart.id + original_name = chart.slice_name + + self.login(ADMIN_USERNAME) + rv = self.client.put( + f"/api/v1/chart/{chart_id}", + json={"slice_name": "etag-put-test"}, + ) + assert rv.status_code == 200 + body = _json.loads(rv.data.decode("utf-8")) + assert body["new_version_uuid"] is not None + assert rv.headers.get("ETag") == f'"{body["new_version_uuid"]}"' + + # Cleanup + chart.slice_name = original_name + db.session.commit() + + def test_chart_list_versions_emits_etag(self) -> None: + db.session.commit() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + expected = _expected_etag(Slice, chart.id, chart.uuid) + + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/chart/{chart.uuid}/versions/") + assert rv.status_code == 200 + assert rv.headers.get("ETag") == expected + + def test_chart_get_version_emits_etag(self) -> None: + db.session.commit() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + expected = _expected_etag(Slice, chart.id, chart.uuid) + + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/chart/{chart.uuid}/versions/") + body = _json.loads(rv.data.decode("utf-8")) + version_uuid = body["result"][0]["version_uuid"] + + rv = self.client.get(f"/api/v1/chart/{chart.uuid}/versions/{version_uuid}/") + assert rv.status_code == 200 + # ETag reflects the live version, not the queried version. + assert rv.headers.get("ETag") == expected + + def test_dashboard_get_emits_etag_matching_current_live_version(self) -> None: + db.session.commit() + dashboard: Dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dashboard is not None + expected = _expected_etag(Dashboard, dashboard.id, dashboard.uuid) + + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/dashboard/{dashboard.id}") + assert rv.status_code == 200 + assert rv.headers.get("ETag") == expected + + def test_dataset_get_emits_etag_matching_current_live_version(self) -> None: + db.session.commit() + dataset: SqlaTable = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert dataset is not None + expected = _expected_etag(SqlaTable, dataset.id, dataset.uuid) + + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/dataset/{dataset.id}") + assert rv.status_code == 200 + assert rv.headers.get("ETag") == expected + + def test_etag_absent_when_entity_has_no_version_rows(self) -> None: + """``set_version_etag`` is a no-op when the entity has no version rows.""" + from sqlalchemy_continuum import version_class + + db.session.commit() + chart: Slice = ( + db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + ) + assert chart is not None + chart_id = chart.id + chart_uuid = chart.uuid + + ver_cls = version_class(Slice) + db.session.query(ver_cls).filter(ver_cls.id == chart_id).delete( + synchronize_session=False + ) + db.session.commit() + + try: + self.login(ADMIN_USERNAME) + rv = self.client.get(f"/api/v1/chart/{chart_id}") + assert rv.status_code == 200 + assert rv.headers.get("ETag") is None + finally: + # Always restore the chart's name + version rows so downstream + # tests in this class don't see corrupted fixture state, even + # if the assertions above fail. + self.client.put( + f"/api/v1/chart/{chart_id}", + json={"slice_name": "Girls"}, + ) + + # Sanity-check that version rows came back. + assert ( + VersionDAO.current_live_version_uuid(Slice, chart_id, chart_uuid) + is not None + ) diff --git a/tests/integration_tests/versioning/perf_validation_tests.py b/tests/integration_tests/versioning/perf_validation_tests.py new file mode 100644 index 000000000000..5e792db50c05 --- /dev/null +++ b/tests/integration_tests/versioning/perf_validation_tests.py @@ -0,0 +1,447 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""T044 — Performance validation for entity version history. + +Skipped by default. Run on demand: + + SUPERSET_PERF_VALIDATION=1 pytest \ + tests/integration_tests/versioning/perf_validation_tests.py -v -s + +Measures the three success criteria defined in the spec: + + * SC-002: version list endpoint responds in under 1 second + * SC-003: restore endpoint completes in under 3 seconds + * SC-004: save path p95 overhead under 50 ms with Continuum tracking + on vs. off (FR-014) + +The test prints a summary table suitable for pasting into the PR +description. It also asserts each target so regressions fail loudly +when the harness is re-run. +""" + +from __future__ import annotations + +import os +import statistics +import time +from typing import Any + +import pytest +import sqlalchemy as sa +from sqlalchemy_continuum import version_class, versioning_manager + +from superset.extensions import db +from superset.models.slice import Slice +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + +SKIP_REASON = "Performance validation is manual. Set SUPERSET_PERF_VALIDATION=1 to run." + +# Thresholds from spec.md §Success Criteria. +LIST_ENDPOINT_MAX_MS = 1000 # SC-002 +RESTORE_ENDPOINT_MAX_MS = 3000 # SC-003 +SAVE_OVERHEAD_P95_MAX_MS = 50 # SC-004 + +# Activity-view thresholds (sc-107283 §Success Criteria). +ACTIVITY_ENDPOINT_P95_MAX_MS = 1500 # SC-AV-001 + + +def _save_chart_once(chart: Slice, suffix: str) -> None: + """One ORM-level save path, mimicking what ChartDAO.update does.""" + chart.slice_name = f"{chart.slice_name[:64]}_{suffix}" + db.session.commit() + + +def _timings_ms(seconds: list[float]) -> dict[str, float]: + ms = sorted(s * 1000.0 for s in seconds) + return { + "p50": statistics.median(ms), + "p95": ms[int(len(ms) * 0.95) - 1] if len(ms) >= 20 else max(ms), + "max": max(ms), + "n": len(ms), + } + + +@pytest.mark.skipif( + not os.environ.get("SUPERSET_PERF_VALIDATION"), + reason=SKIP_REASON, +) +class PerfValidationTests(SupersetTestCase): + """Runs only when SUPERSET_PERF_VALIDATION=1 is set.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices: Any) -> None: # noqa: F811, PT004 + pass + + def _seed_chart_with_n_versions(self, n: int) -> Slice: + """Save a chart N times to produce N version rows.""" + chart = db.session.query(Slice).first() + assert chart is not None, "birth_names fixture should provide charts" + + for i in range(n): + _save_chart_once(chart, f"v{i}") + db.session.commit() + return chart + + def test_sc002_list_endpoint_under_1s(self) -> None: + """SC-002: list endpoint responds in under 1 second.""" + self.login(ADMIN_USERNAME) + + # Generate enough versions to exercise the retention-capped state. + chart = self._seed_chart_with_n_versions(24) + chart_uuid = str(chart.uuid) + url = f"/api/v1/chart/{chart_uuid}/versions/" + + # Warm up the endpoint once (JIT caching, mapper configuration, etc.) + self.client.get(url) + + timings: list[float] = [] + for _ in range(10): + t0 = time.perf_counter() + response = self.client.get(url) + timings.append(time.perf_counter() - t0) + assert response.status_code == 200 + + stats = _timings_ms(timings) + print( + f"\n[SC-002] GET /versions/ (24 versions) " + f"p50={stats['p50']:.1f}ms p95={stats['p95']:.1f}ms " + f"max={stats['max']:.1f}ms n={stats['n']}" + ) + assert stats["p95"] < LIST_ENDPOINT_MAX_MS, ( + f"SC-002 failed: list endpoint p95 {stats['p95']:.1f}ms " + f">= {LIST_ENDPOINT_MAX_MS}ms" + ) + + def test_sc003_restore_endpoint_under_3s(self) -> None: + """SC-003: restore endpoint completes in under 3 seconds.""" + self.login(ADMIN_USERNAME) + + chart = self._seed_chart_with_n_versions(5) + chart_uuid = str(chart.uuid) + + list_response = self.client.get(f"/api/v1/chart/{chart_uuid}/versions/") + assert list_response.status_code == 200 + versions = list_response.get_json()["result"] + assert len(versions) >= 2, "need at least two versions to restore" + target_version_uuid = versions[-1]["version_uuid"] + + restore_url = ( + f"/api/v1/chart/{chart_uuid}/versions/{target_version_uuid}/restore" + ) + + # Warm up once + self.client.post(restore_url) + + timings: list[float] = [] + for _ in range(5): + t0 = time.perf_counter() + response = self.client.post(restore_url) + timings.append(time.perf_counter() - t0) + assert response.status_code == 200 + + stats = _timings_ms(timings) + print( + f"\n[SC-003] POST /restore chart " + f"p50={stats['p50']:.1f}ms max={stats['max']:.1f}ms n={stats['n']}" + ) + assert stats["max"] < RESTORE_ENDPOINT_MAX_MS, ( + f"SC-003 failed: restore max {stats['max']:.1f}ms " + f">= {RESTORE_ENDPOINT_MAX_MS}ms" + ) + + def test_sc004_save_overhead_under_50ms(self) -> None: + """SC-004: save path p95 overhead under 50ms (FR-014). + + Toggling Continuum on and off mid-process corrupts its internal + ``units_of_work`` state and is not a reliable measurement. Instead + this test directly measures the wall-clock time spent inside the + four session-level listeners Continuum attaches to + ``sa.orm.session.Session`` — ``before_flush``, ``after_flush``, + ``after_commit``, ``after_rollback`` — plus Superset's own + baseline / snapshot / retention-prune listeners (attached to + ``db.session``). The cumulative listener time per save is the + marginal overhead version capture adds over a save with + versioning removed entirely, because without these listeners + the ORM would not execute any of that code. + + The approach: + 1. Wrap each known listener with a timing proxy that adds its + wall-clock time to a per-save accumulator. + 2. Save the same chart N times, recording each save's + accumulator value. + 3. Compute p50 / p95 of the per-save overhead. + + This matches the measurement intent of SC-004 (how much does + versioning cost per save) without the fragility of toggling + Continuum mid-test. + """ + self.login(ADMIN_USERNAME) + + chart = db.session.query(Slice).first() + assert chart is not None + + # Per-save accumulator incremented by the wrapped listeners. + acc = [0.0] + + def wrap_listener(original: Any) -> Any: + def wrapper(*args: Any, **kwargs: Any) -> Any: + t0 = time.perf_counter() + try: + return original(*args, **kwargs) + finally: + acc[0] += time.perf_counter() - t0 + + wrapper.__wrapped__ = original # type: ignore[attr-defined] + return wrapper + + # Instrument Continuum's four session listeners by detaching the + # bound method, wrapping, and re-attaching under a single-use + # listener handle we can cleanly remove on teardown. + session_target = sa.orm.session.Session + attached: list[tuple[str, Any]] = [] + for event_name, listener in list(versioning_manager.session_listeners.items()): + sa.event.remove(session_target, event_name, listener) + wrapped = wrap_listener(listener) + sa.event.listen(session_target, event_name, wrapped) + attached.append((event_name, wrapped)) + + iterations = 100 + warmup = 5 + try: + # Warmup (first baseline INSERT, JIT, cache warming). + for i in range(warmup): + _save_chart_once(chart, f"warm_{i}") + acc[0] = 0.0 + + total_timings: list[float] = [] + overhead_timings: list[float] = [] + for i in range(iterations): + acc[0] = 0.0 + t0 = time.perf_counter() + _save_chart_once(chart, f"run_{i}") + total_timings.append(time.perf_counter() - t0) + overhead_timings.append(acc[0]) + finally: + for event_name, wrapped in attached: + sa.event.remove(session_target, event_name, wrapped) + sa.event.listen( + session_target, + event_name, + wrapped.__wrapped__, + ) + + total = _timings_ms(total_timings) + overhead = _timings_ms(overhead_timings) + + ver_cls = version_class(Slice) + produced = db.session.query(ver_cls).filter(ver_cls.id == chart.id).count() + print( + f"\n[SC-004] save iterations={iterations} chart_id={chart.id} " + f"version_rows_produced={produced}" + ) + print( + f"[SC-004] full save: " + f"p50={total['p50']:.2f}ms p95={total['p95']:.2f}ms " + f"max={total['max']:.2f}ms" + ) + print( + f"[SC-004] version-cap overhead: " + f"p50={overhead['p50']:.2f}ms p95={overhead['p95']:.2f}ms " + f"max={overhead['max']:.2f}ms" + ) + + assert overhead["p95"] < SAVE_OVERHEAD_P95_MAX_MS, ( + f"SC-004 failed: version-capture p95 overhead " + f"{overhead['p95']:.2f}ms >= {SAVE_OVERHEAD_P95_MAX_MS}ms" + ) + + # ---- T045: Activity-view perf validation ----------------------------- + + def _seed_activity_history(self) -> str: + """Generate dense history on the birth_names dashboard so the + activity endpoint has something realistic to read. + + T045's spec target is "25 charts × 3 dataset windows each". The + birth_names fixture has ~12 charts on a single dataset (no + multi-dataset support without a bespoke fixture). We approximate + the load by: (a) editing many charts on the dashboard, (b) + editing the dataset's description several times, (c) editing the + dashboard's own title once. That yields ~30+ change records + spanning all three entity kinds — enough to exercise the + decoration, visibility, and impact-batch paths without needing a + multi-dataset fixture builder. Returns the dashboard UUID. + + **Why this commits without rollback** (unlike the test bodies in + ``activity_view_tests.py``): the whole point of a perf seed is + that the rows it produces have actually been persisted, so the + endpoint hit that follows reads a realistic state of the + ``version_changes`` / shadow tables. T053's + ``try/finally``+``rollback`` convention is for tests that + assert on *which records were captured*; here the seed IS the + setup, not the unit under test. The fixture's session-scoped + ``_cleanup`` removes the dashboard / slices at session teardown, + which is when the shadow rows age out too. + """ + # pylint: disable=import-outside-toplevel + from superset.connectors.sqla.models import SqlaTable + from superset.models.dashboard import Dashboard + + dashboard = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title.like("USA Births%")) + .first() + ) + dataset = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert dashboard is not None + assert dataset is not None + dashboard_uuid = str(dashboard.uuid) + + # Many chart edits — most of the activity volume. + for chart in dashboard.slices[:12]: + chart.slice_name = f"{chart.slice_name[:48]}_perf" + db.session.commit() + + # A handful of dataset edits — exercises the impact-batch path + # (Dashboard path + SqlaTable related). + for i in range(5): + dataset.description = f"perf seed iteration {i}" + db.session.commit() + + # One dashboard self-edit. + dashboard.dashboard_title = f"{dashboard.dashboard_title}_perf" + db.session.commit() + + return dashboard_uuid + + def test_av_sc001_activity_endpoint_p95_under_1500ms(self) -> None: + """SC-AV-001: dashboard activity endpoint p95 < 1500ms across 50 + invocations against a realistic history.""" + self.login(ADMIN_USERNAME) + + dashboard_uuid = self._seed_activity_history() + url = f"/api/v1/dashboard/{dashboard_uuid}/activity/" + + # Warmup — JIT, mapper config, identity-map population. + for _ in range(3): + self.client.get(url) + + timings: list[float] = [] + for _ in range(50): + t0 = time.perf_counter() + response = self.client.get(url) + timings.append(time.perf_counter() - t0) + assert response.status_code == 200 + + stats = _timings_ms(timings) + body = response.get_json() + print( + f"\n[AV-SC-001] GET /dashboard//activity/ " + f"records_returned={len(body['result'])} count={body['count']}" + ) + print( + f"[AV-SC-001] p50={stats['p50']:.1f}ms " + f"p95={stats['p95']:.1f}ms max={stats['max']:.1f}ms " + f"n={stats['n']}" + ) + assert stats["p95"] < ACTIVITY_ENDPOINT_P95_MAX_MS, ( + f"AV-SC-001 failed: activity endpoint p95 {stats['p95']:.1f}ms " + f">= {ACTIVITY_ENDPOINT_P95_MAX_MS}ms — profile the query " + f"plan and consider the T046 index migration (see " + f"specs/sc-107283/data-model.md §Possible additive indexes)" + ) + + def test_av_sc003_save_path_p95_unaffected_by_activity_view(self) -> None: + """AV-SC-003: the activity-view feature is read-only. Save path + p95 must remain within sc-103156's SC-004 budget (50ms version- + capture overhead) even with the activity tables in place. + + We re-measure the same overhead SC-004 measures, with the + activity-view branch's code in scope, to catch any accidental + regression from a save-path coupling. + """ + self.login(ADMIN_USERNAME) + # Seed some history so the M2M shadow + version_changes have + # enough rows that any pathological save-time read against them + # would surface. + self._seed_activity_history() + + chart = db.session.query(Slice).first() + assert chart is not None + + acc = [0.0] + + def wrap_listener(original: Any) -> Any: + def wrapper(*args: Any, **kwargs: Any) -> Any: + t0 = time.perf_counter() + try: + return original(*args, **kwargs) + finally: + acc[0] += time.perf_counter() - t0 + + wrapper.__wrapped__ = original # type: ignore[attr-defined] + return wrapper + + session_target = sa.orm.session.Session + attached: list[tuple[str, Any]] = [] + for event_name, listener in list(versioning_manager.session_listeners.items()): + sa.event.remove(session_target, event_name, listener) + wrapped = wrap_listener(listener) + sa.event.listen(session_target, event_name, wrapped) + attached.append((event_name, wrapped)) + + iterations = 50 + warmup = 3 + try: + for i in range(warmup): + _save_chart_once(chart, f"av_warm_{i}") + acc[0] = 0.0 + + overhead_timings: list[float] = [] + for i in range(iterations): + acc[0] = 0.0 + _save_chart_once(chart, f"av_run_{i}") + overhead_timings.append(acc[0]) + finally: + for event_name, wrapped in attached: + sa.event.remove(session_target, event_name, wrapped) + sa.event.listen( + session_target, + event_name, + wrapped.__wrapped__, + ) + + overhead = _timings_ms(overhead_timings) + print( + f"\n[AV-SC-003] save-path overhead with activity-view in scope: " + f"p50={overhead['p50']:.2f}ms p95={overhead['p95']:.2f}ms " + f"max={overhead['max']:.2f}ms" + ) + assert overhead["p95"] < SAVE_OVERHEAD_P95_MAX_MS, ( + f"AV-SC-003 failed: save-path p95 overhead " + f"{overhead['p95']:.2f}ms >= {SAVE_OVERHEAD_P95_MAX_MS}ms — " + f"the activity-view branch has regressed sc-103156's SC-004 " + f"budget; check for a new save-path read coupling." + ) diff --git a/tests/integration_tests/versioning/skip_unmodified_tests.py b/tests/integration_tests/versioning/skip_unmodified_tests.py new file mode 100644 index 000000000000..cb6495f75f9e --- /dev/null +++ b/tests/integration_tests/versioning/skip_unmodified_tests.py @@ -0,0 +1,330 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""FR-026 — ``SkipUnmodifiedPlugin`` integration tests. + +Locks in the behavior that owners-only saves and content-equivalent +re-saves do *not* mint version rows. Exercises the plugin's +``_matches_previous_version`` comparator across the Dashboard's three +column-type families (String, Text, MediumText) so a future column-type +change can't silently regress to "always create version rows". +""" + +from __future__ import annotations + +from typing import Any + +import pytest +from sqlalchemy_continuum import version_class + +from superset.connectors.sqla.models import SqlaTable, TableColumn +from superset.extensions import db +from superset.models.dashboard import Dashboard +from superset.models.slice import Slice +from superset.utils import json as _json +from tests.integration_tests.base_tests import SupersetTestCase +from tests.integration_tests.constants import ADMIN_USERNAME +from tests.integration_tests.fixtures.birth_names_dashboard import ( # noqa: F401 + load_birth_names_dashboard_with_slices, + load_birth_names_data, +) + + +def _dashboard_version_count(dashboard_id: int) -> int: + ver_cls = version_class(Dashboard) + return db.session.query(ver_cls).filter(ver_cls.id == dashboard_id).count() + + +def _slice_version_count(slice_id: int) -> int: + ver_cls = version_class(Slice) + return db.session.query(ver_cls).filter(ver_cls.id == slice_id).count() + + +def _dataset_version_count(dataset_id: int) -> int: + ver_cls = version_class(SqlaTable) + return db.session.query(ver_cls).filter(ver_cls.id == dataset_id).count() + + +class TestSkipUnmodifiedPlugin(SupersetTestCase): + """FR-026 — version rows are not minted for content-equivalent updates.""" + + @pytest.fixture(autouse=True) + def _load_data(self, load_birth_names_dashboard_with_slices): # noqa: PT004, F811 + pass + + def _get_dashboard(self) -> Dashboard: + db.session.commit() + dash = ( + db.session.query(Dashboard) + .filter(Dashboard.dashboard_title == "USA Births Names") + .first() + ) + assert dash is not None + return dash + + def _put(self, pk: int, body: dict[str, Any]) -> None: + rv = self.client.put(f"/api/v1/dashboard/{pk}", json=body) + assert rv.status_code == 200, rv.data + + def test_owners_only_edit_does_not_create_version(self) -> None: + """Saving a dashboard with only owner changes is a no-op for + version-row creation.""" + dash = self._get_dashboard() + dash_id = dash.id + title = dash.dashboard_title + original_owner_ids = [o.id for o in dash.owners] + + self.login(ADMIN_USERNAME) + # Force a known baseline state with one save. + self._put(dash_id, {"dashboard_title": title}) + db.session.expire_all() + before = _dashboard_version_count(dash_id) + + try: + # Now save with only ``owners`` changed (toggle: drop one, + # then put it back). String / Text / MediumText columns are + # unchanged so the plugin should skip both saves. + new_owners = [oid for oid in original_owner_ids if oid != 1] or [] + self._put(dash_id, {"dashboard_title": title, "owners": new_owners}) + db.session.expire_all() + mid = _dashboard_version_count(dash_id) + assert mid == before, ( + f"owners-only edit minted a version row (before={before}, after={mid})" + ) + + self._put(dash_id, {"dashboard_title": title, "owners": original_owner_ids}) + db.session.expire_all() + after = _dashboard_version_count(dash_id) + assert after == before, ( + f"second owners-only edit minted a version row " + f"(before={before}, after={after})" + ) + finally: + # Always restore original ownership. + self._put(dash_id, {"dashboard_title": title, "owners": original_owner_ids}) + + def test_re_save_with_identical_values_does_not_create_version(self) -> None: + """Submitting the same scalar values back through PUT is a no-op + for version creation — exercises the json_metadata re-serialize + case (``set_dash_metadata`` rewrites the column with a different + byte sequence; plugin must compare against the prior shadow row + and skip).""" + dash = self._get_dashboard() + dash_id = dash.id + title = dash.dashboard_title + existing_metadata = dash.json_metadata or "{}" + + self.login(ADMIN_USERNAME) + # Prime: one real save to ensure the json_metadata is in canonical + # post-set_dash_metadata form. + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": existing_metadata}, + ) + db.session.expire_all() + before = _dashboard_version_count(dash_id) + + # Re-submit identical content. set_dash_metadata will round-trip + # the json — the resulting byte sequence might differ from the + # request body but must equal the previous stored value. + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": existing_metadata}, + ) + db.session.expire_all() + after = _dashboard_version_count(dash_id) + assert after == before, ( + f"identical re-save minted a version row (before={before}, after={after})" + ) + + def test_actual_change_creates_version(self) -> None: + """A real scalar change MUST mint a version row — the plugin + only suppresses no-ops, never legitimate edits.""" + dash = self._get_dashboard() + dash_id = dash.id + original_title = dash.dashboard_title + + self.login(ADMIN_USERNAME) + before = _dashboard_version_count(dash_id) + try: + self._put(dash_id, {"dashboard_title": "fr-026-modified-title"}) + db.session.expire_all() + after = _dashboard_version_count(dash_id) + assert after == before + 1, ( + f"real edit failed to mint a version row " + f"(before={before}, after={after})" + ) + finally: + self._put(dash_id, {"dashboard_title": original_title}) + + def test_chart_slice_name_change_creates_version(self) -> None: + """Same assertion for ``Slice`` (covers the ``String`` column path + on a different entity type).""" + db.session.commit() + chart = db.session.query(Slice).filter(Slice.slice_name == "Girls").first() + assert chart is not None + chart_id = chart.id + + self.login(ADMIN_USERNAME) + before = _slice_version_count(chart_id) + try: + rv = self.client.put( + f"/api/v1/chart/{chart_id}", + json={"slice_name": "fr-026-renamed"}, + ) + assert rv.status_code == 200 + db.session.expire_all() + after = _slice_version_count(chart_id) + assert after == before + 1 + finally: + self.client.put(f"/api/v1/chart/{chart_id}", json={"slice_name": "Girls"}) + + def test_dashboard_json_metadata_subkey_change_creates_version(self) -> None: + """Editing a non-audit key inside ``json_metadata`` MUST mint a + version row — exercises the MediumText column path past the + plugin's content-equality check.""" + dash = self._get_dashboard() + dash_id = dash.id + title = dash.dashboard_title + original_metadata = dash.json_metadata or "{}" + + self.login(ADMIN_USERNAME) + before = _dashboard_version_count(dash_id) + try: + md = _json.loads(original_metadata) + md["color_scheme"] = "fr026TestPalette" + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": _json.dumps(md)}, + ) + db.session.expire_all() + after = _dashboard_version_count(dash_id) + assert after == before + 1, ( + f"json_metadata edit failed to mint a version row " + f"(before={before}, after={after})" + ) + finally: + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": original_metadata}, + ) + + def test_map_label_colors_only_change_does_not_create_version(self) -> None: + """Re-stamped ``map_label_colors`` (and other frontend-derived + audit sub-keys) inside ``json_metadata`` MUST NOT mint a version + row. The frontend regenerates this map from the + ``LabelsColorMap`` singleton on every save, so two saves with no + user-authored change emit different bytes for the column. The + diff engine drops these sub-keys via + ``DASHBOARD_JSON_METADATA_AUDIT_KEYS``; the skip-plugin's + comparator must apply the same filter or every save mints an + empty-changes "Baseline" row in the UI. + """ + dash = self._get_dashboard() + dash_id = dash.id + title = dash.dashboard_title + original_metadata = dash.json_metadata or "{}" + + self.login(ADMIN_USERNAME) + # Prime with the existing metadata so the next save's only + # delta is the re-stamped ``map_label_colors``. + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": original_metadata}, + ) + db.session.expire_all() + before = _dashboard_version_count(dash_id) + try: + md = _json.loads(original_metadata) + md["map_label_colors"] = { + "test-label-fr026": "#abcdef", + "another-label": "#123456", + } + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": _json.dumps(md)}, + ) + db.session.expire_all() + after = _dashboard_version_count(dash_id) + assert after == before, ( + f"map_label_colors-only edit minted a version row " + f"(before={before}, after={after})" + ) + finally: + self._put( + dash_id, + {"dashboard_title": title, "json_metadata": original_metadata}, + ) + + def test_dataset_column_edit_creates_parent_version(self) -> None: + """Editing a ``TableColumn`` description MUST mint a parent + ``tables_version`` row even though the parent's own scalars are + unchanged. Without the force-touch in + ``baseline.force_parent_dirty_on_child_change``, child-only + edits leave the dataset's version-history dropdown empty. + """ + db.session.commit() + dataset = ( + db.session.query(SqlaTable) + .filter(SqlaTable.table_name == "birth_names") + .first() + ) + assert dataset is not None + dataset_id = dataset.id + column = ( + db.session.query(TableColumn) + .filter(TableColumn.table_id == dataset_id) + .order_by(TableColumn.id) + .first() + ) + assert column is not None + original_description = column.description + + self.login(ADMIN_USERNAME) + before = _dataset_version_count(dataset_id) + try: + rv = self.client.put( + f"/api/v1/dataset/{dataset_id}", + json={ + "columns": [ + { + "id": column.id, + "column_name": column.column_name, + "description": "fr-026 child-edit forces parent shadow", + }, + ], + }, + ) + assert rv.status_code == 200, rv.data + db.session.expire_all() + after = _dataset_version_count(dataset_id) + assert after == before + 1, ( + f"column edit did not force a parent dataset shadow row " + f"(before={before}, after={after})" + ) + finally: + self.client.put( + f"/api/v1/dataset/{dataset_id}", + json={ + "columns": [ + { + "id": column.id, + "column_name": column.column_name, + "description": original_description, + }, + ], + }, + ) diff --git a/tests/unit_tests/commands/importers/v1/assets_test.py b/tests/unit_tests/commands/importers/v1/assets_test.py index 56b372006da6..f5b44232a685 100644 --- a/tests/unit_tests/commands/importers/v1/assets_test.py +++ b/tests/unit_tests/commands/importers/v1/assets_test.py @@ -107,6 +107,12 @@ def test_import_adds_dashboard_charts(mocker: MockerFixture, session: Session) - expected_number_of_charts = len(charts_config_1) ImportAssetsCommand._import(base_configs) + # ``ImportAssetsCommand.run()`` is wrapped in ``@transaction``, + # so each production invocation gets its own DB (and Continuum) + # transaction. Calling ``_import`` directly twice in the same + # session would otherwise emit conflicting M2M shadow rows for + # ``dashboard_slices`` within a single Continuum tx. + db.session.commit() ImportAssetsCommand._import(new_configs) dashboard_ids = db.session.scalars( select(dashboard_slices.c.dashboard_id).distinct() @@ -574,6 +580,12 @@ def test_import_removes_dashboard_charts( expected_number_of_charts = len(charts_config_2) ImportAssetsCommand._import(base_configs) + # ``ImportAssetsCommand.run()`` is wrapped in ``@transaction``, + # so each production invocation gets its own DB (and Continuum) + # transaction. Calling ``_import`` directly twice in the same + # session would otherwise emit conflicting M2M shadow rows for + # ``dashboard_slices`` within a single Continuum tx. + db.session.commit() ImportAssetsCommand._import(new_configs) dashboard_ids = db.session.scalars( select(dashboard_slices.c.dashboard_id).distinct() diff --git a/tests/unit_tests/daos/test_version_dao.py b/tests/unit_tests/daos/test_version_dao.py new file mode 100644 index 000000000000..e6ded3dd2ad7 --- /dev/null +++ b/tests/unit_tests/daos/test_version_dao.py @@ -0,0 +1,97 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for ``VersionDAO``. + +Exercises the pure helpers (``derive_version_uuid``) and the +``restore_version`` control-flow branches that can be covered with mocks +alone. Full round-trip scalar restore / audit stamping / non-destructive +behaviour is covered by the integration tests in +``tests/integration_tests/{charts,dashboards,datasets}/version_history_tests.py`` +— those need a real Continuum stack and live DB, which unit tests here +deliberately avoid. +""" + +from __future__ import annotations + +from unittest.mock import MagicMock, patch +from uuid import UUID + +from superset.daos.version import ( + derive_version_uuid, + VERSION_UUID_NAMESPACE, + VersionDAO, +) + +# --------------------------------------------------------------------------- +# derive_version_uuid +# --------------------------------------------------------------------------- + + +def test_derive_version_uuid_is_deterministic(): + entity = UUID("14f48794-ebfa-4f60-a26a-582c49132f1b") + assert derive_version_uuid(entity, 42) == derive_version_uuid(entity, 42) + + +def test_derive_version_uuid_differs_across_tx(): + entity = UUID("14f48794-ebfa-4f60-a26a-582c49132f1b") + assert derive_version_uuid(entity, 1) != derive_version_uuid(entity, 2) + + +def test_derive_version_uuid_differs_across_entities(): + tx = 42 + a = UUID("14f48794-ebfa-4f60-a26a-582c49132f1b") + b = UUID("b388a396-cbca-4299-a443-3e41e870e2c2") + assert derive_version_uuid(a, tx) != derive_version_uuid(b, tx) + + +def test_derive_version_uuid_is_v5(): + """UUIDs must be version 5 — changing this is a breaking change.""" + entity = UUID("14f48794-ebfa-4f60-a26a-582c49132f1b") + result = derive_version_uuid(entity, 1) + assert result.version == 5 + + +def test_derive_version_uuid_uses_fixed_namespace(): + """Asserts the namespace constant hasn't drifted (changing it + invalidates every cached version_uuid — see the constant's comment).""" + assert VERSION_UUID_NAMESPACE == UUID("7a6f5d9b-4c3b-5d8e-9a1c-0e2b4c6d8f10") + + +# --------------------------------------------------------------------------- +# restore_version control-flow — unknown entity / out-of-range version +# --------------------------------------------------------------------------- + + +@patch("superset.versioning.restore.find_active_by_uuid", return_value=None) +def test_restore_version_returns_none_for_unknown_entity(mock_find): + """Unknown entity UUID → caller raises 404.""" + result = VersionDAO.restore_version( + MagicMock(__name__="Dashboard"), + UUID("00000000-0000-0000-0000-000000000000"), + 0, + ) + assert result is None + + +# Out-of-range version_num (the lookup query returns None) is verified +# end-to-end in the integration tests +# (``test_restore_returns_404_for_unknown_version_uuid`` in the three +# {charts,dashboards,datasets}/version_history_tests.py suites). A pure +# unit-level version of that test would require mocking the full +# SQLAlchemy expression tree — including ``ver_cls.operation_type != 0`` +# — which is fragile and doesn't add coverage beyond what the +# integration path already provides. diff --git a/tests/unit_tests/dashboards/commands/importers/v1/import_command_test.py b/tests/unit_tests/dashboards/commands/importers/v1/import_command_test.py index 8b56d86e81be..5278078cbb6d 100644 --- a/tests/unit_tests/dashboards/commands/importers/v1/import_command_test.py +++ b/tests/unit_tests/dashboards/commands/importers/v1/import_command_test.py @@ -72,6 +72,13 @@ def test_dashboard_import_with_overwrite_replaces_charts( initial_chart_ids = db.session.scalars(select(dashboard_slices.c.slice_id)).all() assert len(initial_chart_ids) == 2 + # ``ImportDashboardsCommand.run()`` is wrapped in ``@transaction``, + # so each production invocation gets its own DB (and Continuum) + # transaction. Calling ``_import`` directly twice in the same + # session would otherwise emit conflicting M2M shadow rows for + # ``dashboard_slices`` within a single Continuum tx. + db.session.commit() + # Second import: same dashboard with only 1 chart (charts_config_2 has 1 chart) updated_configs = { **copy.deepcopy(databases_config), diff --git a/tests/unit_tests/databases/commands/importers/v1/import_test.py b/tests/unit_tests/databases/commands/importers/v1/import_test.py index 1385d8a2bbb6..187f9ca54305 100644 --- a/tests/unit_tests/databases/commands/importers/v1/import_test.py +++ b/tests/unit_tests/databases/commands/importers/v1/import_test.py @@ -17,6 +17,7 @@ # pylint: disable=unused-argument, import-outside-toplevel, invalid-name import copy +import uuid import pytest from flask import current_app @@ -56,7 +57,7 @@ def test_import_database(mocker: MockerFixture, session: Session) -> None: assert database.allow_dml is True assert database.allow_file_upload is True assert database.extra == "{}" - assert database.uuid == "b8a1ccd3-779d-4ab7-8ad8-9ab119d7fe89" + assert database.uuid == uuid.UUID("b8a1ccd3-779d-4ab7-8ad8-9ab119d7fe89") assert database.is_managed_externally is False assert database.external_url is None @@ -89,7 +90,7 @@ def test_import_database_no_creds(mocker: MockerFixture, session: Session) -> No assert database.database_name == "imported_database_no_creds" assert database.sqlalchemy_uri == "bigquery://test-db/" assert database.extra == "{}" - assert database.uuid == "2ff17edc-f3fa-4609-a5ac-b484281225bc" + assert database.uuid == uuid.UUID("2ff17edc-f3fa-4609-a5ac-b484281225bc") def test_import_database_sqlite_invalid( diff --git a/tests/unit_tests/examples/generic_loader_test.py b/tests/unit_tests/examples/generic_loader_test.py index e921d748f3dc..96d9e4d826f4 100644 --- a/tests/unit_tests/examples/generic_loader_test.py +++ b/tests/unit_tests/examples/generic_loader_test.py @@ -16,6 +16,7 @@ # under the License. """Tests for generic_loader.py UUID threading functionality.""" +import uuid from unittest.mock import MagicMock, patch @@ -54,7 +55,7 @@ def test_load_parquet_table_sets_uuid_on_new_table(mock_db, mock_get_db): uuid=test_uuid, ) - assert tbl.uuid == test_uuid + assert tbl.uuid == uuid.UUID(test_uuid) @patch("superset.examples.generic_loader.get_example_database") diff --git a/tests/unit_tests/initialization_test.py b/tests/unit_tests/initialization_test.py index 65d2ea4c96d0..b6d3626d8978 100644 --- a/tests/unit_tests/initialization_test.py +++ b/tests/unit_tests/initialization_test.py @@ -257,3 +257,152 @@ def test_param_takes_precedence_over_env_var(self, mock_init_app): assert isinstance(app.wsgi_app, AppRootMiddleware) assert app.wsgi_app.app_root == "/from-param" + + +class TestInitVersioning: + """Cover the operational instrumentation added to ``init_versioning`` + in the v4→v5 cycle: the ``ENABLE_VERSIONING_CAPTURE`` kill-switch + and the ``_warn_if_retention_beat_missing`` startup check. + + The happy path (capture on, listeners attach, retention beat entry + present) is exercised by the integration tests; this file pins the + behavioural contract on the misconfiguration / kill-switch branches + that the v5 continuous-delivery review surfaced as load-bearing for + operator alerting and recovery.""" + + def _initializer(self, config: dict) -> SupersetAppInitializer: + """Build a ``SupersetAppInitializer`` against a minimal mock app + whose only meaningful attribute is the config dict. The methods + under test (`_warn_if_retention_beat_missing` and the kill-switch + branch of `init_versioning`) only read from ``self.config``; + nothing about the full Flask app lifecycle is needed.""" + app = MagicMock() + app.config = config + return SupersetAppInitializer(app) + + @patch("superset.initialization.logger") + @patch("superset.versioning.changes.register_change_record_listener") + @patch("superset.versioning.baseline.register_baseline_listener") + def test_kill_switch_off_skips_listener_registration( + self, mock_baseline, mock_change, mock_logger + ): + """``ENABLE_VERSIONING_CAPTURE=False`` MUST short-circuit + ``init_versioning`` before either listener registers. The + operator's 30-second recovery story relies on this.""" + initializer = self._initializer( + { + "ENABLE_VERSIONING_CAPTURE": False, + "CELERY_CONFIG": None, # avoid the warn-log noise + } + ) + + initializer.init_versioning() + + mock_baseline.assert_not_called() + mock_change.assert_not_called() + # One WARNING explaining the skip — operator-visible in deploy log. + assert any( + "ENABLE_VERSIONING_CAPTURE is False" in str(call) + for call in mock_logger.warning.call_args_list + ), ( + "Expected a WARNING log when ENABLE_VERSIONING_CAPTURE=False; " + f"got {mock_logger.warning.call_args_list}" + ) + + @patch("superset.initialization.logger") + def test_warn_when_celery_beat_schedule_missing_retention_entry( + self, mock_logger + ): + """When ``CELERY_CONFIG.beat_schedule`` is present but lacks the + ``version_history.prune_old_versions`` entry, the helper emits + a WARNING. This is the silent-failure mode the v4 CD review + called out: capture writes rows; the prune never fires.""" + + class _PartialCeleryConfig: + beat_schedule = {"reports.scheduler": {"task": "reports.scheduler"}} + + initializer = self._initializer({"CELERY_CONFIG": _PartialCeleryConfig}) + initializer._warn_if_retention_beat_missing() + + assert any( + "version_history.prune_old_versions" in str(call) + for call in mock_logger.warning.call_args_list + ), ( + "Expected a WARNING naming the missing retention entry; " + f"got {mock_logger.warning.call_args_list}" + ) + + @patch("superset.initialization.logger") + def test_no_warn_when_celery_beat_schedule_includes_retention_entry( + self, mock_logger + ): + """When the default ``CeleryConfig`` (or any class with the + entry) is in play, no warning fires. The happy path.""" + + class _CompleteCeleryConfig: + beat_schedule = { + "version_history.prune_old_versions": { + "task": "version_history.prune_old_versions", + }, + } + + initializer = self._initializer({"CELERY_CONFIG": _CompleteCeleryConfig}) + initializer._warn_if_retention_beat_missing() + + mock_logger.warning.assert_not_called() + + @patch("superset.initialization.logger") + def test_no_warn_when_celery_config_is_none(self, mock_logger): + """``CELERY_CONFIG = None`` is the documented "disable Celery + entirely" path. The warn-log MUST NOT fire — the operator made + a deliberate choice; complaining about a missing retention entry + on a Celery-disabled deployment trains operators to ignore the + warning.""" + initializer = self._initializer({"CELERY_CONFIG": None}) + initializer._warn_if_retention_beat_missing() + mock_logger.warning.assert_not_called() + + @patch("superset.initialization.logger") + def test_dict_form_celery_config_with_entry_does_not_warn(self, mock_logger): + """Celery accepts a dict-shaped config via + ``config_from_object``. The warn-log MUST discriminate by + ``isinstance(dict)`` so an operator who supplies a dict with the + entry doesn't see a false-positive warning.""" + initializer = self._initializer( + { + "CELERY_CONFIG": { + "broker_url": "redis://localhost", + "beat_schedule": { + "version_history.prune_old_versions": { + "task": "version_history.prune_old_versions", + }, + }, + }, + } + ) + initializer._warn_if_retention_beat_missing() + mock_logger.warning.assert_not_called() + + @patch("superset.initialization.logger") + def test_dict_form_celery_config_without_entry_warns(self, mock_logger): + """The dict-shape symmetry of the previous test: a dict without + the entry MUST emit the warning, same as a class without it.""" + initializer = self._initializer( + { + "CELERY_CONFIG": { + "broker_url": "redis://localhost", + "beat_schedule": { + "reports.scheduler": {"task": "reports.scheduler"}, + }, + }, + } + ) + initializer._warn_if_retention_beat_missing() + + assert any( + "version_history.prune_old_versions" in str(call) + for call in mock_logger.warning.call_args_list + ), ( + "Expected a WARNING for dict-form CELERY_CONFIG missing the " + f"entry; got {mock_logger.warning.call_args_list}" + ) diff --git a/tests/unit_tests/migrations/composite_pk_association_tables_test.py b/tests/unit_tests/migrations/composite_pk_association_tables_test.py new file mode 100644 index 000000000000..05a69293a23b --- /dev/null +++ b/tests/unit_tests/migrations/composite_pk_association_tables_test.py @@ -0,0 +1,144 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for the composite-PK association-tables migration (revision +2bee73611e32). Verifies the post-migration constraint enforcement: duplicate +``(fk1, fk2)`` insertions fail with IntegrityError, distinct pairs succeed. + +Schema is built from the live ORM ``Table`` definitions via +``metadata.create_all(engine)`` against in-memory SQLite. This reflects the +post-T015–T018 ORM model state (composite-PK), independent of whether the +Alembic migration has run against the test DB. The two should agree. +""" + +import pytest +import sqlalchemy as sa +from sqlalchemy.exc import IntegrityError + +# (table_name, fk1_col, fk2_col, fk1_parent_table, fk2_parent_table) +# Parent-table names are needed to build the FK targets in the in-memory schema. +AFFECTED_TABLES = [ + ("dashboard_roles", "dashboard_id", "role_id", "dashboards", "ab_role"), + ("dashboard_slices", "dashboard_id", "slice_id", "dashboards", "slices"), + ("dashboard_user", "user_id", "dashboard_id", "ab_user", "dashboards"), + ( + "report_schedule_user", + "user_id", + "report_schedule_id", + "ab_user", + "report_schedule", + ), + ( + "rls_filter_roles", + "role_id", + "rls_filter_id", + "ab_role", + "row_level_security_filters", + ), + ( + "rls_filter_tables", + "table_id", + "rls_filter_id", + "tables", + "row_level_security_filters", + ), + ("slice_user", "user_id", "slice_id", "ab_user", "slices"), + ("sqlatable_user", "user_id", "table_id", "ab_user", "tables"), +] + + +def _build_in_memory_schema( + table_name: str, fk1: str, fk2: str, fk1_parent: str, fk2_parent: str +) -> tuple[sa.engine.Engine, sa.Table]: + """Build an in-memory SQLite schema with two minimal parent tables and + the junction table under test (composite-PK shape). Returns the engine + and the junction-table object for inserts.""" + metadata = sa.MetaData() + sa.Table( + fk1_parent, + metadata, + sa.Column("id", sa.Integer, primary_key=True), + ) + if fk2_parent != fk1_parent: + sa.Table( + fk2_parent, + metadata, + sa.Column("id", sa.Integer, primary_key=True), + ) + junction = sa.Table( + table_name, + metadata, + sa.Column( + fk1, + sa.Integer, + sa.ForeignKey(f"{fk1_parent}.id"), + primary_key=True, + ), + sa.Column( + fk2, + sa.Integer, + sa.ForeignKey(f"{fk2_parent}.id"), + primary_key=True, + ), + ) + engine = sa.create_engine("sqlite:///:memory:") + metadata.create_all(engine) + # Seed parent rows so the FK constraints can be satisfied. + # Identifiers come from the AFFECTED_TABLES test parameter list, not user input. + with engine.begin() as conn: + conn.execute( + sa.text(f"INSERT INTO {fk1_parent} (id) VALUES (1), (2)") # noqa: S608 + ) + if fk2_parent != fk1_parent: + conn.execute( + sa.text(f"INSERT INTO {fk2_parent} (id) VALUES (1), (2)") # noqa: S608 + ) + return engine, junction + + +@pytest.mark.parametrize("table,fk1,fk2,fk1_parent,fk2_parent", AFFECTED_TABLES) +def test_duplicate_insert_rejected( + table: str, fk1: str, fk2: str, fk1_parent: str, fk2_parent: str +) -> None: + """Inserting the same ``(fk1, fk2)`` pair twice raises ``IntegrityError``. + + Verifies SC-004 / FR-007 — the composite primary key enforces uniqueness + at the database level on every affected table. + """ + engine, junction = _build_in_memory_schema(table, fk1, fk2, fk1_parent, fk2_parent) + with engine.begin() as conn: + conn.execute(junction.insert().values({fk1: 1, fk2: 1})) + with pytest.raises(IntegrityError): + conn.execute(junction.insert().values({fk1: 1, fk2: 1})) + + +@pytest.mark.parametrize("table,fk1,fk2,fk1_parent,fk2_parent", AFFECTED_TABLES) +def test_distinct_pairs_accepted( + table: str, fk1: str, fk2: str, fk1_parent: str, fk2_parent: str +) -> None: + """Two distinct ``(fk1, fk2)`` pairs both succeed. + + Sanity check that the PK isn't accidentally a single-column constraint + (which would reject ``(1, 1)`` and ``(1, 2)`` as a duplicate on column 1). + """ + engine, junction = _build_in_memory_schema(table, fk1, fk2, fk1_parent, fk2_parent) + with engine.begin() as conn: + conn.execute(junction.insert().values({fk1: 1, fk2: 1})) + conn.execute(junction.insert().values({fk1: 1, fk2: 2})) + result = conn.execute( + sa.text(f"SELECT COUNT(*) FROM {table}") # noqa: S608 + ).scalar_one() + assert result == 2 diff --git a/tests/unit_tests/models/helpers_test.py b/tests/unit_tests/models/helpers_test.py index c93f6121dbd5..49496c078d61 100644 --- a/tests/unit_tests/models/helpers_test.py +++ b/tests/unit_tests/models/helpers_test.py @@ -2847,3 +2847,57 @@ def test_process_sql_expression_no_gate_when_denylists_empty( template_processor=None, ) assert result is not None + + +# ---- UUIDMixin._coerce_uuid ----------------------------------------------- + + +def test_coerce_uuid_converts_valid_uuid_string() -> None: + """The validator coerces a well-formed UUID string to a ``uuid.UUID`` + instance — that's the primary contract that makes downstream callers + (importers, test fixtures, ad-hoc construction) see a consistent + ``UUID`` regardless of what they assigned.""" + import uuid + + from superset.connectors.sqla.models import SqlMetric + + metric = SqlMetric(uuid="00000000-0000-0000-0000-000000000001") + assert isinstance(metric.uuid, uuid.UUID) + assert metric.uuid == uuid.UUID("00000000-0000-0000-0000-000000000001") + + +def test_coerce_uuid_preserves_uuid_instance_unchanged() -> None: + """Already-UUID values must not be re-wrapped or copied.""" + import uuid + + from superset.connectors.sqla.models import SqlMetric + + u = uuid.uuid4() + metric = SqlMetric(uuid=u) + # ``is`` check: the validator MUST return the exact instance for UUIDs, + # not a copy. Round-tripping through ``uuid.UUID(str(u))`` would also + # equal-compare, but defeats this performance contract. + assert metric.uuid is u + + +def test_coerce_uuid_passes_non_uuid_strings_through() -> None: + """Non-UUID-shaped strings pass through unchanged. This keeps test + mocks that use placeholder strings (e.g. + ``test_dashboard_schemas.py``'s ``"dashboard-uuid-7"``) working. + The SQL bind layer surfaces a clearer error if such a value ever + reaches the database. If this contract is ever tightened to raise, + the placeholder-using tests need to migrate to ``uuid.uuid4()``.""" + from superset.connectors.sqla.models import SqlMetric + + metric = SqlMetric(uuid="dashboard-uuid-7") + assert metric.uuid == "dashboard-uuid-7" + + +def test_coerce_uuid_passes_none_through() -> None: + """``None`` (the unset case before the column default fires) must + pass through. ``isinstance(None, str)`` returning False already + covers this, but the test pins the contract.""" + from superset.connectors.sqla.models import SqlMetric + + metric = SqlMetric(uuid=None) + assert metric.uuid is None diff --git a/tests/unit_tests/tasks/test_version_history_retention.py b/tests/unit_tests/tasks/test_version_history_retention.py new file mode 100644 index 000000000000..b1b92e641afc --- /dev/null +++ b/tests/unit_tests/tasks/test_version_history_retention.py @@ -0,0 +1,152 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for the operational instrumentation added to +``superset.tasks.version_history_retention`` in the v4 → v5 cycle. + +Covers the three branches that emit statsd counters but didn't previously +have direct test coverage: the ``retention_days <= 0`` short-circuit, the +``no versioned classes resolved`` short-circuit, and the +``OperationalError`` retry path. The "happy path" / SERIALIZABLE retry +behaviour itself is exercised by +``tests/integration_tests/dashboards/version_history_tests.py`` against +a real database; this file pins the metric-emission contract that the +v5 continuous-delivery review surfaced as load-bearing for operator +alerting. +""" + +from __future__ import annotations + +from unittest.mock import MagicMock, patch + +import pytest +from sqlalchemy.exc import OperationalError + +from superset.tasks import version_history_retention + + +@pytest.fixture(name="stats") +def _stats_fixture() -> MagicMock: + """Patch the shared stats logger so every test can assert on + emissions without standing up the real statsd backend.""" + with patch.object( + version_history_retention, "stats_logger_manager" + ) as mock_manager: + mock_manager.instance = MagicMock() + yield mock_manager.instance + + +def test_retention_disabled_emits_skipped_metric(stats: MagicMock) -> None: + """``retention_days <= 0`` is the documented "disable retention" + config. The early-return must emit ``superset.versioning.retention.skipped`` + so a dashboard can tell "operator disabled it" apart from "scheduler + isn't running".""" + result = version_history_retention._prune_old_versions_impl(retention_days=0) + assert result == {"skipped": 1} + stats.incr.assert_called_once_with("superset.versioning.retention.skipped") + stats.gauge.assert_not_called() + + +def test_no_versioned_classes_resolved_emits_skipped_metric( + stats: MagicMock, +) -> None: + """When ``_resolve_shadow_tables`` returns an empty parent list (the + init-order regression case), the task emits ``.skipped`` and returns + without raising. Same metric name as the operator-disabled branch on + purpose — the dashboard alert is "task is running and has nothing to + do", not "task discovered a misconfiguration"; the WARNING log + carries the diagnostic detail.""" + empty_tables = version_history_retention.ShadowTables( + parent=[], child=[], m2m=None, transaction=MagicMock() + ) + with patch.object( + version_history_retention, + "_resolve_shadow_tables", + return_value=empty_tables, + ): + result = version_history_retention._prune_old_versions_impl(retention_days=30) + assert result == {"skipped": 1} + stats.incr.assert_called_once_with("superset.versioning.retention.skipped") + + +def test_serialization_failure_then_success_increments_retried_once( + stats: MagicMock, +) -> None: + """A single ``OperationalError`` on attempt 1 should: + * fire ``.retried`` once (one retry happened), + * sleep for ``_RETRY_BACKOFF_BASE_SECONDS`` (patched away in tests), + * succeed on attempt 2 with ``stats["retried"] == 1``, + * fire ``.pruned_transactions`` gauge with the success count. + + The contract on ``.retried`` is "fires per retry attempt observed" + — the v5 review noted the commit message implied "per session" but + the code is per-attempt. This test pins the per-attempt shape so a + future refactor doesn't silently change the metric semantics.""" + pass_fn = MagicMock( + side_effect=[ + OperationalError("SELECT 1", {}, Exception("could not serialize access")), + {"pruned_transactions": 7, "cutoff": "2026-01-01T00:00:00"}, + ] + ) + tables = version_history_retention.ShadowTables( + parent=[MagicMock()], child=[MagicMock()], m2m=None, transaction=MagicMock() + ) + with ( + patch.object( + version_history_retention, "_resolve_shadow_tables", return_value=tables + ), + patch.object(version_history_retention, "_run_prune_pass", pass_fn), + patch.object(version_history_retention.time, "sleep"), + ): + result = version_history_retention._prune_old_versions_impl(retention_days=30) + + assert result["retried"] == 1 + assert result["pruned_transactions"] == 7 + incr_calls = [call.args[0] for call in stats.incr.call_args_list] + assert incr_calls == ["superset.versioning.retention.retried"], ( + f"Expected exactly one .retried emission; got {incr_calls}" + ) + stats.gauge.assert_called_once_with( + "superset.versioning.retention.pruned_transactions", 7 + ) + + +def test_all_attempts_fail_reraises_after_max_retries(stats: MagicMock) -> None: + """When every attempt raises ``OperationalError``, the task re-raises + after ``_MAX_RETRY_ATTEMPTS`` so the outer Celery wrapper logs it. + The retry counter fires once per attempt that hit the exception.""" + exc = OperationalError("SELECT 1", {}, Exception("conflict")) + tables = version_history_retention.ShadowTables( + parent=[MagicMock()], child=[MagicMock()], m2m=None, transaction=MagicMock() + ) + with ( + patch.object( + version_history_retention, "_resolve_shadow_tables", return_value=tables + ), + patch.object(version_history_retention, "_run_prune_pass", side_effect=exc), + patch.object(version_history_retention.time, "sleep"), + pytest.raises(OperationalError), + ): + version_history_retention._prune_old_versions_impl(retention_days=30) + + incr_calls = [call.args[0] for call in stats.incr.call_args_list] + assert ( + incr_calls.count("superset.versioning.retention.retried") + == version_history_retention._MAX_RETRY_ATTEMPTS + ), ( + f"Expected {version_history_retention._MAX_RETRY_ATTEMPTS} " + f".retried emissions (one per attempt); got {incr_calls}" + ) diff --git a/tests/unit_tests/versioning/__init__.py b/tests/unit_tests/versioning/__init__.py new file mode 100644 index 000000000000..13a83393a912 --- /dev/null +++ b/tests/unit_tests/versioning/__init__.py @@ -0,0 +1,16 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. diff --git a/tests/unit_tests/versioning/test_activity.py b/tests/unit_tests/versioning/test_activity.py new file mode 100644 index 000000000000..573924d4d5ac --- /dev/null +++ b/tests/unit_tests/versioning/test_activity.py @@ -0,0 +1,602 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for ``superset.versioning.activity`` pure helpers (sc-107283). + +No app context, no DB, no Flask. Covers the helpers that can be exercised +in isolation: window intersection, scope resolution branching, entity- +window merging, AV-012 summary headlines, ``changed_by`` projection, +read-predicate fall-through, and the no-impact paths of +``_compute_impact``. The DB-touching helpers +(``charts_attached_to_dashboard``, ``datasets_used_by_chart``, +``fetch_change_records``, ``apply_entity_name_denormalization``, +``check_entity_tombstones``, ``_lookup_entity_uuids``) are exercised +by the integration suite in +``tests/integration_tests/versioning/activity_view_tests.py``. +""" + +from __future__ import annotations + +from datetime import datetime +from unittest.mock import patch + +import pytest + +from superset.versioning.activity import ( + ActivityParamsError, + EntityWindows, + parse_activity_query_params, + Window, +) +from superset.versioning.activity.impact import ( + collect_impact_pairs, + impact_for_record, +) +from superset.versioning.activity.kinds import API_KIND_TO_TABLE, TABLE_KIND_TO_API +from superset.versioning.activity.orchestrator import ( + _DEFAULT_PAGE_SIZE, + _emit_request_shape_attributes, + _MAX_PAGE_SIZE, +) +from superset.versioning.activity.render import _build_summary, _changed_by_dict +from superset.versioning.activity.scope import resolve_scope +from superset.versioning.activity.windows import ( + intersect_windows, + merge_entity_windows, + row_within_any_window, + union_windows, +) + +# ---- intersect_windows --------------------------------------------------- + + +@pytest.mark.parametrize( + "outer, inner, expected", + [ + # Inner fully inside outer + (Window(10, 20), Window(15, 18), Window(15, 18)), + # Left overlap — clipped on the left + (Window(10, 20), Window(5, 15), Window(10, 15)), + # Right overlap — clipped on the right + (Window(10, 20), Window(15, 25), Window(15, 20)), + # Outer fully inside inner + (Window(10, 20), Window(5, 25), Window(10, 20)), + # Touching at end → half-open semantics yield disjoint + (Window(10, 20), Window(20, 30), None), + # Disjoint to the right + (Window(10, 20), Window(25, 30), None), + # Disjoint to the left + (Window(10, 20), Window(0, 5), None), + # Open-ended outer (end_tx=None means +∞) + (Window(10, None), Window(5, 25), Window(10, 25)), + # Open-ended inner + (Window(10, 20), Window(5, None), Window(10, 20)), + # Both open-ended + (Window(10, None), Window(5, None), Window(10, None)), + # Identical + (Window(10, 20), Window(10, 20), Window(10, 20)), + ], +) +def test_intersect_windows( + outer: Window, inner: Window, expected: Window | None +) -> None: + assert intersect_windows(outer, inner) == expected + + +# ---- resolve_scope ------------------------------------------------------- + + +def test_resolve_scope_self_only_for_dashboard() -> None: + """``include='self'`` yields exactly one tuple covering all transactions.""" + assert resolve_scope("Dashboard", 42, "self") == [ + ("Dashboard", 42, [Window(0, None)]), + ] + + +def test_resolve_scope_self_only_for_chart() -> None: + assert resolve_scope("Slice", 7, "self") == [("Slice", 7, [Window(0, None)])] + + +def test_resolve_scope_self_only_for_dataset() -> None: + assert resolve_scope("SqlaTable", 9, "self") == [ + ("SqlaTable", 9, [Window(0, None)]), + ] + + +def test_dataset_has_no_related_scope() -> None: + """AV-004: datasets are not transitive recipients of activity in V2.""" + assert resolve_scope("SqlaTable", 9, "related") == [] + + +def test_dataset_all_returns_only_self() -> None: + """For datasets, ``include='all'`` == ``include='self'`` (AV-004).""" + assert resolve_scope("SqlaTable", 9, "all") == [ + ("SqlaTable", 9, [Window(0, None)]), + ] + + +# ---- merge_entity_windows ----------------------------------------------- + + +def test_merge_entity_windows_collapses_repeated_keys() -> None: + """Repeated ``(api_kind, entity_id)`` entries union their window lists + so the fetch query's OR-clause stays compact.""" + merged = merge_entity_windows( + [ + ("Slice", 1, [Window(0, 100)]), + ("Slice", 1, [Window(200, 300)]), + ("SqlaTable", 5, [Window(0, None)]), + ] + ) + by_key = {(kind, eid): windows for kind, eid, windows in merged} + assert by_key[("Slice", 1)] == [Window(0, 100), Window(200, 300)] + assert by_key[("SqlaTable", 5)] == [Window(0, None)] + + +def test_merge_entity_windows_preserves_singletons() -> None: + """Non-duplicated entries pass through unchanged.""" + inputs: list[EntityWindows] = [ + ("Slice", 1, [Window(0, 100)]), + ("Dashboard", 2, [Window(10, 20)]), + ] + merged = merge_entity_windows(inputs) + assert sorted(merged) == sorted(inputs) + + +def test_merge_entity_windows_unions_overlapping_windows_for_one_entity() -> None: + """Same entity, many redundant attachment windows → collapsed to one. + + This guards the SQLite expression-tree limit: a fixture that + re-creates a chart-on-dashboard association across many transactions + used to produce N separate OR branches in the fetch query (one per + redundant window). merge_entity_windows must coalesce them. + """ + scope: list[EntityWindows] = [ + ("Slice", 1, [Window(10, 20)]), + ("Slice", 1, [Window(15, 25)]), # overlaps + ("Slice", 1, [Window(25, 30)]), # touches + ("Slice", 1, [Window(40, 50)]), # disjoint + ] + merged = merge_entity_windows(scope) + assert merged == [("Slice", 1, [Window(10, 30), Window(40, 50)])] + + +# ---- union_windows ------------------------------------------------------- + + +@pytest.mark.parametrize( + "windows, expected", + [ + # Disjoint windows pass through + ( + [Window(10, 20), Window(30, 40)], + [Window(10, 20), Window(30, 40)], + ), + # Overlapping windows merge + ([Window(10, 20), Window(15, 25)], [Window(10, 25)]), + # Touching windows merge (half-open: [10,20) + [20,30) = [10,30)) + ([Window(10, 20), Window(20, 30)], [Window(10, 30)]), + # Many overlapping windows collapse to one + ( + [Window(10, 20), Window(15, 25), Window(20, 30), Window(25, 35)], + [Window(10, 35)], + ), + # Input order doesn't matter + ( + [Window(30, 40), Window(10, 20), Window(15, 25)], + [Window(10, 25), Window(30, 40)], + ), + # Open-ended absorbs everything to the right + ([Window(10, None), Window(50, 60)], [Window(10, None)]), + # Open-ended at the right merges into open-ended + ([Window(10, 20), Window(15, None)], [Window(10, None)]), + # Empty input + ([], []), + # Single window pass-through + ([Window(5, 10)], [Window(5, 10)]), + ], +) +def test_union_windows(windows: list[Window], expected: list[Window]) -> None: + assert union_windows(windows) == expected + + +# ---- row_within_any_window (Python post-filter for the fetch query) ------ + + +def test_row_in_window_inside() -> None: + assert row_within_any_window({"transaction_id": 15}, [Window(10, 20)]) + + +def test_row_in_window_at_start_boundary_inclusive() -> None: + """Half-open: ``[10, 20)`` includes 10.""" + assert row_within_any_window({"transaction_id": 10}, [Window(10, 20)]) + + +def test_row_in_window_at_end_boundary_exclusive() -> None: + """Half-open: ``[10, 20)`` excludes 20.""" + assert not row_within_any_window({"transaction_id": 20}, [Window(10, 20)]) + + +def test_row_in_open_ended_window() -> None: + """``end=None`` means +∞.""" + assert row_within_any_window({"transaction_id": 999}, [Window(10, None)]) + + +def test_row_in_any_of_several_windows() -> None: + assert row_within_any_window( + {"transaction_id": 50}, + [Window(10, 20), Window(40, 60), Window(90, 100)], + ) + + +def test_row_in_no_windows_returns_false() -> None: + assert not row_within_any_window({"transaction_id": 50}, []) + assert not row_within_any_window( + {"transaction_id": 25}, [Window(10, 20), Window(30, 40)] + ) + + +# ---- Kind translation round-trip ----------------------------------------- + + +def test_kind_translation_is_bijective_for_supported_kinds() -> None: + """Every API kind maps to a table kind and back to the same value. + Locks in the contract that the two maps don't drift.""" + for api_kind, table_kind in API_KIND_TO_TABLE.items(): + assert TABLE_KIND_TO_API[table_kind] == api_kind + + +# ---- _build_summary (AV-012) --------------------------------------------- + + +def test_summary_for_dataset_column_change() -> None: + rec = {"kind": "column", "entity_name": "Sales Transactions"} + assert _build_summary("SqlaTable", rec) == ( + "Dataset column changed: Sales Transactions" + ) + + +def test_summary_for_chart_filter_change() -> None: + rec = {"kind": "filter", "entity_name": "Top Charts"} + assert _build_summary("Slice", rec) == "Chart filter changed: Top Charts" + + +def test_summary_for_restore_event() -> None: + rec = {"kind": "restore", "entity_name": "Q4 Dashboard"} + assert _build_summary("Dashboard", rec) == "Dashboard restored: Q4 Dashboard" + + +def test_summary_unknown_kind_falls_back_to_updated() -> None: + """Unmapped change kinds collapse to a generic 'updated' verb.""" + rec = {"kind": "mystery_kind", "entity_name": "X"} + assert _build_summary("Dashboard", rec) == "Dashboard updated: X" + + +def test_summary_without_entity_name_drops_colon() -> None: + """Tombstoned entities have no name; the headline reads naturally + without a trailing colon and empty value.""" + rec = {"kind": "filter", "entity_name": ""} + assert _build_summary("Slice", rec) == "Chart filter changed" + + +# ---- _changed_by_dict ---------------------------------------------------- + + +def test_changed_by_returns_none_when_no_user_attached() -> None: + """Saves from CLI/Celery/import have no Flask user (sc-103156 §Session + 2026-05-18 clarification).""" + assert _changed_by_dict({"changed_by_id": None}) is None + + +def test_changed_by_projects_only_display_fields() -> None: + """Per the ActivityChangedBy contract: id + first_name + last_name only. + Username is intentionally omitted (data-model.md).""" + record = { + "changed_by_id": 5, + "first_name": "Mike", + "last_name": "Bridge", + "user_id": 5, # internal column, must not leak + } + result = _changed_by_dict(record) + assert result == {"id": 5, "first_name": "Mike", "last_name": "Bridge"} + assert result is not None + assert "username" not in result + + +# ---- impact_for_record (pure, post-batch) ------------------------------- + + +def test_impact_for_record_dashboard_path_dataset_related_uses_count() -> None: + """The only path/related shape that carries impact: ``Dashboard`` → + ``SqlaTable``. The count comes from the pre-batched lookup.""" + record = {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100} + counts = {(5, 100): 3} + assert impact_for_record(record, "Dashboard", counts) == {"charts": 3} + + +def test_impact_for_record_missing_count_yields_none() -> None: + """A pair the batch query didn't return (no matching siblings) + collapses to ``None`` rather than ``{"charts": 0}``.""" + record = {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100} + assert impact_for_record(record, "Dashboard", {}) is None + + +def test_impact_for_record_zero_count_yields_none() -> None: + """Explicit zero in the counts map is treated the same as missing — + no impact field on the wire.""" + record = {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100} + assert impact_for_record(record, "Dashboard", {(5, 100): 0}) is None + + +def test_impact_for_record_dashboard_path_chart_related_yields_none() -> None: + """Dashboard → chart is a direct dependency; no further sibling + layer to count.""" + record = {"entity_kind": "chart", "entity_id": 5, "transaction_id": 100} + assert impact_for_record(record, "Dashboard", {(5, 100): 999}) is None + + +def test_impact_for_record_chart_path_with_dataset_related_yields_none() -> None: + """Chart → dataset: the chart is itself the only dependent of the + dataset edit.""" + record = {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100} + assert impact_for_record(record, "Slice", {(5, 100): 999}) is None + + +def test_impact_for_record_dataset_path_yields_none() -> None: + """Datasets have no transitive layer (AV-004).""" + record = {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100} + assert impact_for_record(record, "SqlaTable", {(5, 100): 999}) is None + + +# ---- collect_impact_pairs ----------------------------------------------- + + +def test_collect_impact_pairs_dashboard_path_collects_only_datasets() -> None: + """The batched pre-query only needs ``(dataset_id, tx)`` pairs. + Chart-related and self records aren't relevant.""" + records = [ + {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100}, + {"entity_kind": "dataset", "entity_id": 7, "transaction_id": 200}, + {"entity_kind": "chart", "entity_id": 9, "transaction_id": 300}, + {"entity_kind": "dashboard", "entity_id": 1, "transaction_id": 400}, + ] + assert collect_impact_pairs(records, "Dashboard") == {(5, 100), (7, 200)} + + +def test_collect_impact_pairs_dedupes_repeated_pairs() -> None: + """Multiple change records for the same (dataset, tx) collapse to + one pair — the batch query computes the count once.""" + records = [ + {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100}, + {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100}, + {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100}, + ] + pairs = collect_impact_pairs(records, "Dashboard") + assert pairs == {(5, 100)} + + +def test_collect_impact_pairs_chart_path_returns_empty() -> None: + """Chart paths have no dashboard layer to count siblings on, so the + batch never needs to fire.""" + records = [ + {"entity_kind": "dataset", "entity_id": 5, "transaction_id": 100}, + ] + assert collect_impact_pairs(records, "Slice") == set() + + +def test_collect_impact_pairs_dataset_path_returns_empty() -> None: + records = [ + {"entity_kind": "chart", "entity_id": 5, "transaction_id": 100}, + ] + assert collect_impact_pairs(records, "SqlaTable") == set() + + +def test_collect_impact_pairs_empty_records_returns_empty() -> None: + assert collect_impact_pairs([], "Dashboard") == set() + + +# ---- parse_activity_query_params (shared API helper) --------------------- + + +def test_parser_defaults_when_empty() -> None: + """No params → ``include='all'``, ``page=0``, ``page_size=DEFAULT``.""" + assert parse_activity_query_params({}) == { + "include": "all", + "page": 0, + "page_size": _DEFAULT_PAGE_SIZE, + } + + +def test_parser_clamps_page_size_to_max() -> None: + """A request for more than the contract maximum is clamped, not 400'd + (silent clamp matches AV-019's bounded-payload guarantee).""" + params = parse_activity_query_params({"page_size": str(_MAX_PAGE_SIZE * 5)}) + assert params["page_size"] == _MAX_PAGE_SIZE + + +def test_parser_accepts_iso_datetime_with_z_suffix() -> None: + """Python <3.11 fromisoformat rejects 'Z'; the parser tolerates it.""" + params = parse_activity_query_params({"since": "2026-01-01T00:00:00Z"}) + assert params["since"].year == 2026 + + +def test_parser_normalises_z_suffix_to_naive_utc() -> None: + """The 'Z' result must be tz-NAIVE: ``issued_at`` is a naive column, so a + tz-aware bind shifts the comparison by the session offset (or raises) on + PostgreSQL. The 'Z' instant is already UTC, so the value is unchanged.""" + since = parse_activity_query_params({"since": "2026-01-01T00:00:00Z"})["since"] + assert since.tzinfo is None + assert since == datetime(2026, 1, 1, 0, 0, 0) + + +def test_parser_normalises_offset_to_naive_utc() -> None: + """A non-UTC offset is converted to UTC and stripped to naive, so the + comparison against the naive ``issued_at`` column is in the same frame.""" + since = parse_activity_query_params({"since": "2026-01-01T05:00:00+02:00"})["since"] + assert since.tzinfo is None + assert since == datetime(2026, 1, 1, 3, 0, 0) # 05:00 +02:00 -> 03:00 UTC + + +def test_parser_rejects_invalid_include() -> None: + with pytest.raises(ActivityParamsError, match="include"): + parse_activity_query_params({"include": "sibling"}) + + +def test_parser_rejects_malformed_datetime() -> None: + with pytest.raises(ActivityParamsError, match="since"): + parse_activity_query_params({"since": "yesterday"}) + + +def test_parser_rejects_negative_page() -> None: + with pytest.raises(ActivityParamsError, match="page"): + parse_activity_query_params({"page": "-1"}) + + +def test_parser_rejects_zero_page_size() -> None: + with pytest.raises(ActivityParamsError, match="page_size"): + parse_activity_query_params({"page_size": "0"}) + + +def test_parser_error_is_a_value_error() -> None: + """``ActivityParamsError`` subclasses ``ValueError`` so callers that + only know about the standard library exception hierarchy still catch + it correctly.""" + with pytest.raises(ValueError, match="include"): + parse_activity_query_params({"include": "nope"}) + + +# ---- Observability metric-key convention (T050 cross-coupling) ---------- + + +def test_metric_prefix_matches_versioning_namespace_convention() -> None: + """T050: cross-coupling sanity. The activity-view's instrumentation + prefix (``superset.activity_view.*``) must be a sibling of sc-103156's + eventual ``superset.versioning.*`` namespace, not nested under + a different root. Both endpoint families belong to the versioning + feature; their metrics should be discoverable from one Grafana + filter (``superset.activity_view.*`` OR ``superset.versioning.*``). + + Locking the prefix in a test catches accidental drift in a code + review — a future PR renaming the prefix would fail this assertion + and require explicit acknowledgement. + """ + from superset.versioning.activity.orchestrator import _METRIC_PREFIX + + assert _METRIC_PREFIX == "superset.activity_view", ( + f"Activity-view metrics prefix changed from " + f"'superset.activity_view' to {_METRIC_PREFIX!r}. If this was " + "intentional, update sc-103156's FR-027 instrumentation to " + "match the new convention OR document the new naming in plan §D-17." + ) + # Sibling-namespace check: starts with the versioning-feature root. + assert _METRIC_PREFIX.startswith("superset."), ( + "All Superset metrics live under 'superset.*'; activity_view must too." + ) + + +# ---- _emit_request_shape_attributes: related-entity counts --------------- +# +# The ``related_entity_count.*`` gauges report how many *other* entities an +# activity request fanned out to. ``resolve_scope`` always prepends the path +# entity itself (the "self" window) to the scope list, so the metric loop +# must exclude that self entry — otherwise a chart/dataset request reports +# one phantom related entity of its own kind. + + +def _gauge_value(mock_sl: object, suffix: str) -> float: + """Return the value of the single ``gauge`` call whose metric name ends + with *suffix*. Fails loudly if absent so a renamed metric surfaces here.""" + for call in mock_sl.gauge.call_args_list: # type: ignore[attr-defined] + name = call.args[0] + if name.endswith(suffix): + return call.args[1] + emitted = [c.args[0] for c in mock_sl.gauge.call_args_list] # type: ignore[attr-defined] + raise AssertionError(f"no gauge ending {suffix!r}; emitted {emitted}") + + +@patch("superset.extensions.stats_logger_manager") +def test_related_entity_count_excludes_self_for_chart(mock_mgr) -> None: + """A chart request scopes to itself + the datasets it used. The charts + gauge must read 0 (no *related* charts) even though the self Slice is in + the scope list; the datasets gauge counts only the two related datasets.""" + sl = mock_mgr.instance + entity_windows: list[EntityWindows] = [ + ("Slice", 7, [Window(0, None)]), # self — must not be counted + ("SqlaTable", 5, [Window(0, None)]), # related dataset + ("SqlaTable", 9, [Window(0, None)]), # related dataset + ] + + _emit_request_shape_attributes( + "slice", + include="all", + has_since_filter=False, + page_size=25, + record_count=3, + entity_windows=entity_windows, + path_kind="Slice", + path_id=7, + ) + + assert _gauge_value(sl, "related_entity_count.charts") == 0.0 + assert _gauge_value(sl, "related_entity_count.datasets") == 2.0 + + +@patch("superset.extensions.stats_logger_manager") +def test_related_entity_count_excludes_self_for_dataset(mock_mgr) -> None: + """Datasets have no related scope, so an ``include=all`` dataset request + scopes to itself only. The datasets gauge must read 0, not 1.""" + sl = mock_mgr.instance + entity_windows: list[EntityWindows] = [ + ("SqlaTable", 9, [Window(0, None)]), # self only + ] + + _emit_request_shape_attributes( + "sqlatable", + include="all", + has_since_filter=False, + page_size=25, + record_count=1, + entity_windows=entity_windows, + path_kind="SqlaTable", + path_id=9, + ) + + assert _gauge_value(sl, "related_entity_count.datasets") == 0.0 + + +@patch("superset.extensions.stats_logger_manager") +def test_related_entity_count_counts_genuine_related_of_same_kind(mock_mgr) -> None: + """Self-exclusion keys on (kind, id), not kind alone: a dashboard whose + scope happened to include another dashboard would still count it.""" + sl = mock_mgr.instance + entity_windows: list[EntityWindows] = [ + ("Dashboard", 1, [Window(0, None)]), # self + ("Slice", 5, [Window(0, None)]), # related chart + ("Slice", 6, [Window(0, None)]), # related chart + ] + + _emit_request_shape_attributes( + "dashboard", + include="all", + has_since_filter=False, + page_size=25, + record_count=2, + entity_windows=entity_windows, + path_kind="Dashboard", + path_id=1, + ) + + assert _gauge_value(sl, "related_entity_count.charts") == 2.0 diff --git a/tests/unit_tests/versioning/test_diff.py b/tests/unit_tests/versioning/test_diff.py new file mode 100644 index 000000000000..6170c034e1e2 --- /dev/null +++ b/tests/unit_tests/versioning/test_diff.py @@ -0,0 +1,1408 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for ``superset.versioning.diff`` (T051). + +Pure-function tests — no app context, no DB. Covers: + +- (a) scalar field change +- (b) filter added / removed / modified (Slice params) +- (c) metric added / removed (Slice params + dataset SqlMetric) +- (d) column added / removed / type-changed (dataset TableColumn) +- (e) ``dashboard_slices`` added / removed +- (f) replay round-trip — applying records in order reconstructs post-state (SC-008) +""" + +from __future__ import annotations + +from copy import deepcopy +from typing import Any + +from superset.utils import json as _json +from superset.versioning.diff import ( + _diff_layout_node, + _LAYOUT_META_DIFF_DEPTH, + _recursive_leaf_diff, + ChangeRecord, + diff_dashboard, + diff_dashboard_layout, + diff_dashboard_slices, + diff_dataset, + diff_dataset_columns, + diff_dataset_metrics, + diff_json_field, + diff_scalar_fields, + diff_slice, + diff_slice_params, + scalar_fields_for, +) + +# Field universes used by tests. In production the listener passes the +# result of ``scalar_fields_for(ModelClass, special=...)``; in tests we +# pass explicit sets so assertions remain stable even if a contributor +# later adds or renames a column on the real model. + +_SLICE_TEST_FIELDS: frozenset[str] = frozenset( + { + "slice_name", + "datasource_type", + "datasource_id", + "viz_type", + "description", + "cache_timeout", + "external_url", + "is_managed_externally", + "certified_by", + "certification_details", + } +) + +_DASHBOARD_TEST_FIELDS: frozenset[str] = frozenset( + { + "dashboard_title", + "position_json", + "json_metadata", + "slug", + "css", + "external_url", + "is_managed_externally", + "certified_by", + "certification_details", + "published", + } +) + +_DATASET_TEST_FIELDS: frozenset[str] = frozenset( + { + "table_name", + "sql", + "description", + "cache_timeout", + "template_params", + "extra", + "main_dttm_col", + "default_endpoint", + "offset", + "schema", + "catalog", + "filter_select_enabled", + "fetch_values_predicate", + "is_sqllab_view", + "is_managed_externally", + "external_url", + "normalize_columns", + "always_filter_main_dttm", + } +) + +# --------------------------------------------------------------------------- +# (a) Scalar field change +# --------------------------------------------------------------------------- + + +def test_slice_scalar_rename() -> None: + pre = {"slice_name": "Sales Report"} + post = {"slice_name": "Sales Report Q1"} + records = diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) + assert records == [ + ChangeRecord( + kind="field", + operation="edit", + path=["slice_name"], + from_value="Sales Report", + to_value="Sales Report Q1", + ) + ] + + +def test_slice_scalar_unchanged_emits_nothing() -> None: + pre = {"slice_name": "Sales Report", "description": "x"} + post = {"slice_name": "Sales Report", "description": "x"} + assert diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) == [] + + +def test_dashboard_scalar_change_falls_through_to_field() -> None: + pre = {"dashboard_title": "Old", "position_json": '{"a":1}'} + post = {"dashboard_title": "New", "position_json": '{"a":2}'} + records = diff_dashboard(pre, post, fields=_DASHBOARD_TEST_FIELDS) + assert len(records) == 2 + kinds = {r.kind for r in records} + assert kinds == {"field"} + paths = {tuple(r.path) for r in records} + assert paths == {("dashboard_title",), ("position_json",)} + + +def test_dataset_scalar_change_falls_through_to_field() -> None: + pre = {"sql": "SELECT 1", "description": "old"} + post = {"sql": "SELECT 2", "description": "new"} + records = diff_dataset(pre, post, fields=_DATASET_TEST_FIELDS) + kinds = {r.kind for r in records} + paths = {tuple(r.path) for r in records} + assert kinds == {"field"} + assert paths == {("sql",), ("description",)} + + +def test_unknown_fields_are_ignored() -> None: + # Fields outside the known scalar set are silently skipped — we + # don't emit spurious ``field`` records for ORM-internal columns. + pre = {"__unmapped__": "x"} + post = {"__unmapped__": "y"} + assert diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) == [] + assert diff_dashboard(pre, post, fields=_DASHBOARD_TEST_FIELDS) == [] + assert diff_dataset(pre, post, fields=_DATASET_TEST_FIELDS) == [] + + +# --------------------------------------------------------------------------- +# scalar_fields_for — model reflection +# --------------------------------------------------------------------------- + + +class _FakeColumn: + """Stand-in for a SQLAlchemy ``Column`` that exposes just ``.name``.""" + + def __init__(self, name: str) -> None: + self.name = name + + +class _FakeTable: + """Stand-in for ``Model.__table__`` that exposes an iterable ``columns``.""" + + def __init__(self, column_names: list[str]) -> None: + self.columns = [_FakeColumn(n) for n in column_names] + + +def test_scalar_fields_for_strips_audit_and_excludes() -> None: + """Reflection excludes __versioned__.exclude + audit fields + special.""" + + class _Model: + __table__ = _FakeTable( + [ + "id", + "uuid", + "name", + "description", + "secret_field", + "created_on", + "changed_on", + "created_by_fk", + "changed_by_fk", + "params", + ] + ) + __versioned__ = {"exclude": ["secret_field"]} + + result = scalar_fields_for(_Model, special=frozenset({"params"})) + assert result == frozenset({"name", "description"}) + + +def test_scalar_fields_for_no_versioned_attr() -> None: + """Models without ``__versioned__`` work — exclude defaults to empty.""" + + class _Model: + __table__ = _FakeTable(["id", "name", "created_on"]) + + result = scalar_fields_for(_Model) + assert result == frozenset({"name"}) + + +def test_scalar_fields_for_empty_versioned_dict() -> None: + """``__versioned__ = {}`` is treated as no additional exclusions.""" + + class _Model: + __table__ = _FakeTable(["id", "name"]) + __versioned__: dict[str, Any] = {} + + result = scalar_fields_for(_Model) + assert result == frozenset({"name"}) + + +def test_scalar_fields_for_no_table_returns_empty() -> None: + """Objects without ``__table__`` produce an empty set, not an error.""" + + class _NotAModel: + pass + + assert scalar_fields_for(_NotAModel) == frozenset() + + +def test_scalar_fields_for_custom_field_in_derivative() -> None: + """Derivatives get custom scalar fields without editing ``diff.py``.""" + + class _DerivedSlice: + """Simulates a downstream fork that added ``preset_embedded_config``.""" + + __table__ = _FakeTable( + [ + "id", + "uuid", + "slice_name", + "params", + "preset_embedded_config", # downstream addition + "created_on", + "changed_on", + "created_by_fk", + "changed_by_fk", + ] + ) + __versioned__ = {"exclude": ["query_context"]} + + result = scalar_fields_for(_DerivedSlice, special=frozenset({"params"})) + # Core and downstream fields both appear — zero maintenance in diff.py. + assert "slice_name" in result + assert "preset_embedded_config" in result + assert "params" not in result # handled specially + assert "id" not in result # audit + + +# --------------------------------------------------------------------------- +# diff_scalar_fields — generic primitive used by all entity types +# --------------------------------------------------------------------------- + + +def test_diff_scalar_fields_only_emits_changed_fields() -> None: + pre = {"a": 1, "b": "x", "c": True} + post = {"a": 2, "b": "x", "c": False} + records = diff_scalar_fields(pre, post, fields={"a", "b", "c"}) + paths = {tuple(r.path) for r in records} + assert paths == {("a",), ("c",)} + assert {r.kind for r in records} == {"field"} + + +def test_diff_scalar_fields_ignores_fields_outside_universe() -> None: + # ``extra`` differs, but isn't in the fields set → no record. + pre = {"a": 1, "extra": 100} + post = {"a": 2, "extra": 200} + records = diff_scalar_fields(pre, post, fields={"a"}) + assert len(records) == 1 + assert records[0].path == ["a"] + + +def test_null_to_empty_string_is_not_a_change() -> None: + """Superset's save path normalises nullable strings (``css``, + ``certified_by``, ``certification_details``) to ``""`` on first + write. The transition ``null → ""`` carries no user-authored + signal and must not produce a record. Same for the reverse. + """ + # Both directions silently pass. + pre = {"css": None, "certified_by": "", "title": "Old"} + post = {"css": "", "certified_by": None, "title": "New"} + records = diff_scalar_fields(pre, post, fields={"css", "certified_by", "title"}) + paths = [r.path for r in records] + assert ["css"] not in paths + assert ["certified_by"] not in paths + # Real change still emits. + assert ["title"] in paths + + +def test_real_string_change_still_emits() -> None: + """Sanity: the null/"" filter must not swallow genuine edits.""" + pre = {"description": ""} + post = {"description": "non-empty"} + records = diff_scalar_fields(pre, post, fields={"description"}) + assert len(records) == 1 + assert records[0].from_value == "" + assert records[0].to_value == "non-empty" + + +# --------------------------------------------------------------------------- +# (b) Chart params — filters +# --------------------------------------------------------------------------- + + +FILTER_COUNTRY = { + "subject": "country", + "operator": "==", + "comparator": "Canada", + "expressionType": "SIMPLE", +} +FILTER_COUNTRY_REGION = { + "subject": "country", + "operator": "==", + "comparator": "Canada/Quebec", + "expressionType": "SIMPLE", +} +FILTER_DATE = { + "subject": "order_date", + "operator": ">", + "comparator": "2020-01-01", + "expressionType": "SIMPLE", +} + + +def _params_json(**kwargs: Any) -> str: + return _json.dumps(kwargs) + + +def test_filter_added() -> None: + records = diff_slice_params( + _params_json(adhoc_filters=[]), + _params_json(adhoc_filters=[FILTER_COUNTRY]), + ) + assert len(records) == 1 + r = records[0] + assert r.kind == "filter" + assert r.path == ["params", "adhoc_filters", "country"] + assert r.from_value is None + assert r.to_value == FILTER_COUNTRY + + +def test_filter_removed() -> None: + records = diff_slice_params( + _params_json(adhoc_filters=[FILTER_COUNTRY, FILTER_DATE]), + _params_json(adhoc_filters=[FILTER_DATE]), + ) + assert len(records) == 1 + r = records[0] + assert r.kind == "filter" + assert r.path == ["params", "adhoc_filters", "country"] + assert r.from_value == FILTER_COUNTRY + assert r.to_value is None + + +def test_filter_modified_same_subject() -> None: + records = diff_slice_params( + _params_json(adhoc_filters=[FILTER_COUNTRY]), + _params_json(adhoc_filters=[FILTER_COUNTRY_REGION]), + ) + assert len(records) == 1 + r = records[0] + assert r.kind == "filter" + assert r.path == ["params", "adhoc_filters", "country"] + assert r.from_value == FILTER_COUNTRY + assert r.to_value == FILTER_COUNTRY_REGION + + +def test_filter_insert_in_middle_is_still_one_record() -> None: + # Position-based diffing would emit three records for this case. + # Natural-key diffing emits exactly one. + records = diff_slice_params( + _params_json(adhoc_filters=[FILTER_COUNTRY, FILTER_DATE]), + _params_json( + adhoc_filters=[ + FILTER_COUNTRY, + {"subject": "city", "operator": "in", "comparator": ["Montreal"]}, + FILTER_DATE, + ] + ), + ) + assert len(records) == 1 + assert records[0].path == ["params", "adhoc_filters", "city"] + assert records[0].from_value is None + assert records[0].to_value["subject"] == "city" + + +# --------------------------------------------------------------------------- +# (b-continued) Chart params — scalar first-class kinds +# --------------------------------------------------------------------------- + + +def test_time_range_change() -> None: + records = diff_slice_params( + _params_json(time_range="Last week"), + _params_json(time_range="Last month"), + ) + assert records == [ + ChangeRecord( + kind="time_range", + operation="edit", + path=["params", "time_range"], + from_value="Last week", + to_value="Last month", + ) + ] + + +def test_time_range_added_from_null() -> None: + records = diff_slice_params( + _params_json(), + _params_json(time_range="Last week"), + ) + assert records == [ + ChangeRecord( + kind="time_range", + operation="add", + path=["params", "time_range"], + from_value=None, + to_value="Last week", + ) + ] + + +def test_color_palette_change() -> None: + records = diff_slice_params( + _params_json(color_scheme="supersetColors"), + _params_json(color_scheme="presetColors"), + ) + assert records[0].kind == "color_palette" + assert records[0].path == ["params", "color_scheme"] + + +def test_unknown_params_sub_key_falls_through_to_field() -> None: + records = diff_slice_params( + _params_json(something_custom="x"), + _params_json(something_custom="y"), + ) + assert records == [ + ChangeRecord( + kind="field", + operation="edit", + path=["params", "something_custom"], + from_value="x", + to_value="y", + ) + ] + + +def test_params_audit_keys_are_excluded() -> None: + """``params.slice_id`` is a machine-stamped self-reference and must + not produce a record. Superset's save paths add or refresh it on + every save (see ``superset/views/core.py``), so without this filter + every chart save would emit a spurious ``["params", "slice_id"]`` + record on the first save after the key was missing. + """ + # slice_id added (null → 104): no record. + assert diff_slice_params(_params_json(), _params_json(slice_id=104)) == [] + # slice_id changed (101 → 104): no record. + assert ( + diff_slice_params(_params_json(slice_id=101), _params_json(slice_id=104)) == [] + ) + # slice_id alongside a real edit: only the real edit is emitted. + records = diff_slice_params( + _params_json(slice_id=104, time_range="Last week"), + _params_json(slice_id=104, time_range="Last month"), + ) + assert records == [ + ChangeRecord( + kind="time_range", + operation="edit", + path=["params", "time_range"], + from_value="Last week", + to_value="Last month", + ) + ] + + +# --------------------------------------------------------------------------- +# (c) Chart params — metrics +# --------------------------------------------------------------------------- + + +METRIC_SUM_SALES = { + "label": "SUM(sales)", + "aggregate": "SUM", + "column": {"column_name": "sales"}, + "expressionType": "SIMPLE", +} +METRIC_COUNT_ORDERS = { + "label": "COUNT(orders)", + "aggregate": "COUNT", + "column": {"column_name": "orders"}, + "expressionType": "SIMPLE", +} + + +def test_chart_metric_added() -> None: + records = diff_slice_params( + _params_json(metrics=[]), + _params_json(metrics=[METRIC_SUM_SALES]), + ) + assert records == [ + ChangeRecord( + kind="metric", + operation="add", + path=["params", "metrics", "SUM(sales)"], + from_value=None, + to_value=METRIC_SUM_SALES, + ) + ] + + +def test_chart_metric_removed() -> None: + records = diff_slice_params( + _params_json(metrics=[METRIC_SUM_SALES, METRIC_COUNT_ORDERS]), + _params_json(metrics=[METRIC_COUNT_ORDERS]), + ) + assert records == [ + ChangeRecord( + kind="metric", + operation="remove", + path=["params", "metrics", "SUM(sales)"], + from_value=METRIC_SUM_SALES, + to_value=None, + ) + ] + + +# --------------------------------------------------------------------------- +# (c-continued) Chart params — dimensions +# --------------------------------------------------------------------------- + + +def test_dimension_added() -> None: + records = diff_slice_params( + _params_json(groupby=["country"]), + _params_json(groupby=["country", "city"]), + ) + assert records == [ + ChangeRecord( + kind="dimension", + operation="add", + path=["params", "groupby", "city"], + from_value=None, + to_value="city", + ) + ] + + +def test_dimension_removed() -> None: + records = diff_slice_params( + _params_json(groupby=["country", "city"]), + _params_json(groupby=["country"]), + ) + assert records == [ + ChangeRecord( + kind="dimension", + operation="remove", + path=["params", "groupby", "city"], + from_value="city", + to_value=None, + ) + ] + + +# --------------------------------------------------------------------------- +# (d) Dataset columns +# --------------------------------------------------------------------------- + + +COLUMN_COUNTRY = {"column_name": "country", "type": "VARCHAR(255)", "is_dttm": False} +COLUMN_COUNTRY_TEXT = {"column_name": "country", "type": "TEXT", "is_dttm": False} +COLUMN_DATE = {"column_name": "order_date", "type": "DATE", "is_dttm": True} + + +def test_column_added() -> None: + records = diff_dataset_columns([], [COLUMN_COUNTRY]) + assert records == [ + ChangeRecord( + kind="column", + operation="add", + path=["columns", "country"], + from_value=None, + to_value=COLUMN_COUNTRY, + ) + ] + + +def test_column_removed() -> None: + records = diff_dataset_columns([COLUMN_COUNTRY, COLUMN_DATE], [COLUMN_DATE]) + assert records == [ + ChangeRecord( + kind="column", + operation="remove", + path=["columns", "country"], + from_value=COLUMN_COUNTRY, + to_value=None, + ) + ] + + +def test_column_type_changed() -> None: + records = diff_dataset_columns([COLUMN_COUNTRY], [COLUMN_COUNTRY_TEXT]) + assert records == [ + ChangeRecord( + kind="column", + operation="edit", + path=["columns", "country"], + from_value=COLUMN_COUNTRY, + to_value=COLUMN_COUNTRY_TEXT, + ) + ] + + +def test_column_unchanged_emits_nothing() -> None: + assert diff_dataset_columns([COLUMN_COUNTRY], [COLUMN_COUNTRY]) == [] + + +def test_column_audit_only_change_is_ignored() -> None: + """Refreshed ``changed_on`` alone must not produce a record. + + Reproduces the dataset-editor scenario where adding one calculated + column refreshes ``changed_on`` on every other column as a + side-effect of the save. Before the audit-field strip, each + untouched column produced a spurious 'changed' record. + """ + pre = { + "column_name": "country", + "type": "VARCHAR", + "id": 1226, + "table_id": 17, + "changed_on": "2026-04-24T18:49:07.368009", + "created_on": "2026-04-24T18:49:07.368008", + "changed_by_fk": 1, + "created_by_fk": 1, + } + post = dict(pre, changed_on="2026-04-24T18:49:07.502720") + assert diff_dataset_columns([pre], [post]) == [] + + +def test_column_id_change_with_same_content_is_ignored() -> None: + """``override_columns`` re-insert gives new ids; don't fire a record. + + Under DatasetDAO.update_columns' override_columns pattern a + column's row can be deleted and re-inserted with the same natural + key (``column_name``) and content but a new auto-increment id. + The natural key matches, so we don't emit add+remove; the id-only + difference must be filtered so we don't emit a spurious 'changed'. + """ + pre = {"column_name": "country", "type": "VARCHAR", "id": 1226, "table_id": 17} + post = dict(pre, id=1234) + assert diff_dataset_columns([pre], [post]) == [] + + +def test_column_real_content_change_still_emits() -> None: + """After stripping audit fields, a genuine content change still fires.""" + pre = { + "column_name": "country", + "type": "VARCHAR", + "id": 1226, + "changed_on": "2026-04-24T18:49:07.368009", + } + post = dict(pre, type="TEXT", changed_on="2026-04-24T18:49:07.502720") + records = diff_dataset_columns([pre], [post]) + assert len(records) == 1 + # Stripped values reach the renderer — no audit noise in the record. + assert "changed_on" not in records[0].from_value + assert "changed_on" not in records[0].to_value + assert records[0].from_value["type"] == "VARCHAR" + assert records[0].to_value["type"] == "TEXT" + + +# --------------------------------------------------------------------------- +# (d-continued) Dataset metrics +# --------------------------------------------------------------------------- + + +DATASET_METRIC_SUM = {"metric_name": "sum_sales", "expression": "SUM(sales)"} +DATASET_METRIC_AVG = {"metric_name": "avg_sales", "expression": "AVG(sales)"} + + +def test_dataset_metric_added() -> None: + records = diff_dataset_metrics([], [DATASET_METRIC_SUM]) + assert records == [ + ChangeRecord( + kind="metric", + operation="add", + path=["metrics", "sum_sales"], + from_value=None, + to_value=DATASET_METRIC_SUM, + ) + ] + + +def test_dataset_metric_removed() -> None: + records = diff_dataset_metrics( + [DATASET_METRIC_SUM, DATASET_METRIC_AVG], [DATASET_METRIC_AVG] + ) + assert records == [ + ChangeRecord( + kind="metric", + operation="remove", + path=["metrics", "sum_sales"], + from_value=DATASET_METRIC_SUM, + to_value=None, + ) + ] + + +# --------------------------------------------------------------------------- +# (e) Dashboard slices (chart membership) +# --------------------------------------------------------------------------- + + +def test_dashboard_chart_added() -> None: + records = diff_dashboard_slices(["u-1"], ["u-1", "u-2"]) + assert records == [ + ChangeRecord( + kind="chart", + operation="add", + path=["slices", "u-2"], + from_value=None, + to_value="u-2", + ) + ] + + +def test_dashboard_chart_removed() -> None: + records = diff_dashboard_slices(["u-1", "u-2"], ["u-1"]) + assert records == [ + ChangeRecord( + kind="chart", + operation="remove", + path=["slices", "u-2"], + from_value="u-2", + to_value=None, + ) + ] + + +def test_dashboard_chart_no_change() -> None: + assert diff_dashboard_slices(["u-1"], ["u-1"]) == [] + + +def test_dashboard_chart_swap_emits_add_plus_remove() -> None: + records = diff_dashboard_slices(["u-1"], ["u-2"]) + kinds = {r.kind for r in records} + tos = {r.to_value for r in records} + froms = {r.from_value for r in records} + assert kinds == {"chart"} + assert tos == {"u-2", None} + assert froms == {"u-1", None} + + +# --------------------------------------------------------------------------- +# (e2) Dashboard JSON-blob fields (json_metadata, position_json) +# --------------------------------------------------------------------------- + + +def test_diff_json_field_emits_per_changed_top_level_key() -> None: + """Each changed top-level key produces a separate record. + + Mirrors the behaviour of ``diff_slice_params`` for chart params: + walking the parsed JSON dict means a save that only adds + ``map_label_colors`` doesn't also re-emit the entire blob — only + one record for that key. + """ + pre = _json.dumps({"color_scheme": "", "label_colors": {}, "refresh_frequency": 0}) + post = _json.dumps( + { + "color_scheme": "", # unchanged + "label_colors": {}, # unchanged + "refresh_frequency": 30, # changed + "map_label_colors": {"x": "#fff"}, # added + } + ) + records = diff_json_field("json_metadata", pre, post) + paths = {tuple(r.path) for r in records} + assert paths == { + ("json_metadata", "refresh_frequency"), + ("json_metadata", "map_label_colors"), + } + assert {r.kind for r in records} == {"field"} + + +def test_diff_json_field_treats_null_and_empty_string_as_equivalent() -> None: + """A key that flips from missing/null/"" to "" produces no record.""" + pre = _json.dumps({"color_scheme": None, "label_colors": {}}) + post = _json.dumps({"color_scheme": "", "label_colors": {}}) + assert diff_json_field("json_metadata", pre, post) == [] + + +def test_diff_json_field_handles_invalid_or_null_input() -> None: + """Malformed JSON / None / non-string values must not crash — + both sides degrade to the empty dict, so no records are emitted. + """ + assert diff_json_field("json_metadata", None, None) == [] + assert diff_json_field("json_metadata", "not-json", "{}") == [] + assert diff_json_field("position_json", "{}", None) == [] + + +def test_diff_dashboard_walks_json_blobs_structurally() -> None: + """Full dashboard diff: scalar edit + json_metadata edit produce + one record each, keyed by sub-path. The json_metadata blob is + NOT emitted as a single opaque ``["json_metadata"]`` record. + """ + pre = { + "dashboard_title": "Old", + "json_metadata": _json.dumps({"refresh_frequency": 0}), + "position_json": _json.dumps({"GRID_ID": {"type": "GRID"}}), + } + post = { + "dashboard_title": "New", + "json_metadata": _json.dumps({"refresh_frequency": 30}), + "position_json": _json.dumps({"GRID_ID": {"type": "GRID"}}), + } + records = diff_dashboard(pre, post, fields={"dashboard_title"}) + paths = {tuple(r.path) for r in records} + assert paths == { + ("dashboard_title",), + ("json_metadata", "refresh_frequency"), + } + # Confirm the full json_metadata string is NOT in any record's + # from/to_value — the structural walk replaced opaque-blob storage. + for r in records: + assert "refresh_frequency" not in str(r.from_value or "") or ( + r.path == ["json_metadata", "refresh_frequency"] + ) + + +# --------------------------------------------------------------------------- +# (f) Replay round-trip — SC-008 +# --------------------------------------------------------------------------- + + +def _apply_field(state: dict[str, Any], path: list[Any], value: Any) -> None: + """Generic set-by-path for ``kind="field"`` records.""" + cursor = state + for seg in path[:-1]: + cursor = cursor.setdefault(seg, {}) + cursor[path[-1]] = value + + +def _replay(pre: dict[str, Any], records: list[ChangeRecord]) -> dict[str, Any]: + """Apply change records to the pre-state. + + Dispatches on ``kind`` because named kinds use natural-key paths + (e.g. ``["columns", "country"]``) that are not valid JSON Pointer + locations — the replay function has to understand the semantics + of each kind. + """ + state = deepcopy(pre) + for r in records: + if r.kind == "field": + _apply_field(state, r.path, r.to_value) + elif r.kind == "filter": + _apply_list_by_key(state, r, list_key="adhoc_filters", id_key="subject") + elif r.kind == "metric" and r.path[:2] == ["params", "metrics"]: + _apply_list_by_key(state, r, list_key="metrics", id_key="label") + elif r.kind == "metric" and r.path[:1] == ["metrics"]: + _apply_dataset_list_by_key( + state, r, list_key="metrics", id_key="metric_name" + ) + elif r.kind == "column": + _apply_dataset_list_by_key( + state, r, list_key="columns", id_key="column_name" + ) + elif r.kind == "dimension": + _apply_scalar_list_by_key(state, r) + elif r.kind in ("time_range", "color_palette"): + params = _coerce_params_in_state(state) + params[r.path[-1]] = r.to_value + state["params"] = _json.dumps(params) + elif r.kind == "chart": + _apply_chart_membership(state, r) + else: + raise AssertionError(f"replay: unknown kind {r.kind!r}") + return state + + +def _coerce_params_in_state(state: dict[str, Any]) -> dict[str, Any]: + raw = state.get("params") + if raw is None: + return {} + if isinstance(raw, str): + return _json.loads(raw) if raw else {} + return raw + + +def _apply_list_by_key( + state: dict[str, Any], r: ChangeRecord, list_key: str, id_key: str +) -> None: + """Apply a record to a ``params.`` natural-keyed list.""" + params = _coerce_params_in_state(state) + items = list(params.get(list_key, [])) + natural_key = r.path[-1] + idx = next( + (i for i, item in enumerate(items) if item.get(id_key) == natural_key), None + ) + if r.to_value is None: + # removal + if idx is not None: + items.pop(idx) + elif idx is not None: + # modify in place + items[idx] = r.to_value + else: + items.append(r.to_value) + params[list_key] = items + state["params"] = _json.dumps(params) + + +def _apply_scalar_list_by_key(state: dict[str, Any], r: ChangeRecord) -> None: + """Dimension-style: groupby/columns are lists of strings.""" + params = _coerce_params_in_state(state) + list_key = r.path[1] # "groupby" or "columns" + items = list(params.get(list_key, [])) + natural_key = r.path[-1] + if r.to_value is None: + items = [x for x in items if x != natural_key] + elif natural_key not in items: + items.append(r.to_value) + params[list_key] = items + state["params"] = _json.dumps(params) + + +def _apply_dataset_list_by_key( + state: dict[str, Any], r: ChangeRecord, list_key: str, id_key: str +) -> None: + """Dataset children live at top level, not inside ``params``.""" + items = list(state.get(list_key, [])) + natural_key = r.path[-1] + idx = next( + (i for i, item in enumerate(items) if item.get(id_key) == natural_key), None + ) + if r.to_value is None: + if idx is not None: + items.pop(idx) + elif idx is not None: + items[idx] = r.to_value + else: + items.append(r.to_value) + state[list_key] = items + + +def _apply_chart_membership(state: dict[str, Any], r: ChangeRecord) -> None: + items = list(state.get("slice_uuids", [])) + target = r.path[-1] + if r.to_value is None: + items = [u for u in items if u != target] + elif target not in items: + items.append(r.to_value) + state["slice_uuids"] = items + + +def test_replay_slice_scalar_roundtrip() -> None: + pre = {"slice_name": "Old", "description": None, "params": _params_json()} + post = { + "slice_name": "New", + "description": "added", + "params": _params_json(), + } + records = diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) + assert _replay(pre, records)["slice_name"] == post["slice_name"] + assert _replay(pre, records)["description"] == post["description"] + + +def test_replay_slice_params_roundtrip_filter_added() -> None: + pre = {"slice_name": "x", "params": _params_json(adhoc_filters=[])} + post = { + "slice_name": "x", + "params": _params_json(adhoc_filters=[FILTER_COUNTRY]), + } + records = diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) + result = _replay(pre, records) + assert _json.loads(result["params"]) == _json.loads(post["params"]) + + +def test_replay_slice_params_roundtrip_filter_removed() -> None: + pre = { + "slice_name": "x", + "params": _params_json(adhoc_filters=[FILTER_COUNTRY, FILTER_DATE]), + } + post = { + "slice_name": "x", + "params": _params_json(adhoc_filters=[FILTER_DATE]), + } + records = diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) + result = _replay(pre, records) + assert _json.loads(result["params"]) == _json.loads(post["params"]) + + +def test_replay_time_range_and_color_palette() -> None: + pre = { + "slice_name": "x", + "params": _params_json(time_range="Last week", color_scheme="supersetColors"), + } + post = { + "slice_name": "x", + "params": _params_json(time_range="Last month", color_scheme="presetColors"), + } + records = diff_slice(pre, post, fields=_SLICE_TEST_FIELDS) + result = _replay(pre, records) + assert _json.loads(result["params"]) == _json.loads(post["params"]) + + +def test_replay_dataset_columns_roundtrip() -> None: + pre = {"columns": [COLUMN_COUNTRY, COLUMN_DATE]} + post = {"columns": [COLUMN_COUNTRY_TEXT, COLUMN_DATE]} # type-changed + records = diff_dataset_columns(pre["columns"], post["columns"]) + assert _replay(pre, records)["columns"] == post["columns"] + + +def test_replay_dataset_metrics_roundtrip() -> None: + pre = {"metrics": [DATASET_METRIC_SUM]} + post = {"metrics": [DATASET_METRIC_AVG]} # add avg, remove sum + records = diff_dataset_metrics(pre["metrics"], post["metrics"]) + result_metrics = _replay(pre, records)["metrics"] + # order-insensitive comparison + assert sorted(result_metrics, key=lambda m: m["metric_name"]) == sorted( + post["metrics"], key=lambda m: m["metric_name"] + ) + + +def test_replay_dashboard_slices_roundtrip() -> None: + pre = {"slice_uuids": ["u-1", "u-2"]} + post = {"slice_uuids": ["u-2", "u-3"]} # remove u-1, add u-3 + records = diff_dashboard_slices(pre["slice_uuids"], post["slice_uuids"]) + result = _replay(pre, records) + assert sorted(result["slice_uuids"]) == sorted(post["slice_uuids"]) + + +def test_replay_dashboard_scalar_roundtrip() -> None: + pre = {"dashboard_title": "Old", "position_json": '{"a":1}'} + post = {"dashboard_title": "New", "position_json": '{"a":2}'} + records = diff_dashboard(pre, post, fields=_DASHBOARD_TEST_FIELDS) + assert _replay(pre, records) == { + "dashboard_title": "New", + "position_json": '{"a":2}', + } + + +# --------------------------------------------------------------------------- +# Edge cases +# --------------------------------------------------------------------------- + + +def test_malformed_params_string_is_treated_as_empty() -> None: + # If ``params`` is not valid JSON, ``diff_slice_params`` degrades + # to "no params recorded" rather than crashing the save path. + records = diff_slice_params("not json", _params_json(time_range="Last week")) + assert records == [ + ChangeRecord( + kind="time_range", + operation="add", + path=["params", "time_range"], + from_value=None, + to_value="Last week", + ) + ] + + +def test_none_params_on_both_sides() -> None: + assert diff_slice_params(None, None) == [] + + +def test_filter_without_subject_falls_back_to_position() -> None: + # Keyless filters should not crash; they fall back to the list index. + filter_no_subject = {"operator": "==", "comparator": "x"} + records = diff_slice_params( + _params_json(adhoc_filters=[]), + _params_json(adhoc_filters=[filter_no_subject]), + ) + assert len(records) == 1 + assert records[0].kind == "filter" + assert records[0].to_value == filter_no_subject + + +def test_empty_state_emits_nothing() -> None: + assert diff_slice({}, {}, fields=_SLICE_TEST_FIELDS) == [] + assert diff_dashboard({}, {}, fields=_DASHBOARD_TEST_FIELDS) == [] + assert diff_dataset({}, {}, fields=_DATASET_TEST_FIELDS) == [] + assert diff_dataset_columns([], []) == [] + assert diff_dataset_metrics([], []) == [] + assert diff_dashboard_slices([], []) == [] + + +# --------------------------------------------------------------------------- +# (g) Shape B — leaf-level recursion into nested JSON values +# --------------------------------------------------------------------------- + + +def test_recursive_leaf_diff_emits_one_record_per_changed_leaf() -> None: + """Two leaves change inside the same dict → two records, each + carrying just the changed leaf value (not the whole sub-tree).""" + pre = {"a": 1, "b": {"c": "old", "d": "same"}} + post = {"a": 2, "b": {"c": "new", "d": "same"}} + records = _recursive_leaf_diff( + kind="field", path_prefix=["root"], pre=pre, post=post, max_depth=10 + ) + paths = {tuple(r.path): (r.from_value, r.to_value) for r in records} + assert paths == { + ("root", "a"): (1, 2), + ("root", "b", "c"): ("old", "new"), + } + + +def test_recursive_leaf_diff_equivalent_inputs_emit_nothing() -> None: + """No record when sides are equal — including the None-vs-empty + equivalence carved out by ``_values_equivalent``.""" + assert _recursive_leaf_diff("field", [], {"x": 1}, {"x": 1}, max_depth=5) == [] + assert _recursive_leaf_diff("field", [], None, "", max_depth=5) == [] + + +def test_recursive_leaf_diff_treats_list_as_opaque_leaf() -> None: + """A list on either side is emitted as a single leaf — positional + paths would break under reorder, so we don't recurse into lists.""" + pre = {"items": [1, 2, 3]} + post = {"items": [1, 2, 3, 4]} + records = _recursive_leaf_diff( + kind="field", path_prefix=["x"], pre=pre, post=post, max_depth=10 + ) + assert len(records) == 1 + assert records[0].path == ["x", "items"] + assert records[0].from_value == [1, 2, 3] + assert records[0].to_value == [1, 2, 3, 4] + + +def test_recursive_leaf_diff_emits_leaf_on_type_mismatch() -> None: + """Dict on one side, scalar/None on the other → leaf record carrying + both raw values. No recursion possible across the mismatch.""" + records = _recursive_leaf_diff( + kind="field", path_prefix=["x"], pre=None, post={"a": 1}, max_depth=5 + ) + assert records == [ + ChangeRecord( + kind="field", + operation="add", + path=["x"], + from_value=None, + to_value={"a": 1}, + ) + ] + + +def test_recursive_leaf_diff_depth_cap_emits_opaque_subtree() -> None: + """When recursion hits the depth cap with dicts on both sides, the + sub-tree is emitted as a single leaf rather than walked deeper.""" + pre = {"a": {"b": {"c": "old"}}} + post = {"a": {"b": {"c": "new"}}} + # max_depth=1 means: recurse into the top dict (a), then stop. + # The sub-tree under "a" is emitted as one opaque leaf. + records = _recursive_leaf_diff( + kind="field", path_prefix=[], pre=pre, post=post, max_depth=1 + ) + assert len(records) == 1 + assert records[0].path == ["a"] + assert records[0].from_value == {"b": {"c": "old"}} + assert records[0].to_value == {"b": {"c": "new"}} + + +def test_diff_json_field_recurses_into_nested_dict() -> None: + """Single nested leaf change inside ``json_metadata`` produces one + record at the leaf path — NOT one record carrying the whole + top-level sub-tree on both sides.""" + pre = _json.dumps( + { + "native_filter_configuration": { + "NATIVE_FILTER-abc": { + "defaultDataMask": { + "filterState": {"value": ["US"]}, + } + } + } + } + ) + post = _json.dumps( + { + "native_filter_configuration": { + "NATIVE_FILTER-abc": { + "defaultDataMask": { + "filterState": {"value": ["CA"]}, + } + } + } + } + ) + records = diff_json_field("json_metadata", pre, post) + assert len(records) == 1 + assert records[0].path == [ + "json_metadata", + "native_filter_configuration", + "NATIVE_FILTER-abc", + "defaultDataMask", + "filterState", + "value", + ] + assert records[0].from_value == ["US"] + assert records[0].to_value == ["CA"] + + +def test_diff_json_field_emits_one_record_per_leaf_when_multiple_change() -> None: + """Two leaves change inside the same nested sub-tree → two records, + NOT one record carrying both leaves' diff.""" + pre = _json.dumps({"settings": {"theme": "light", "density": "compact"}}) + post = _json.dumps({"settings": {"theme": "dark", "density": "comfortable"}}) + records = diff_json_field("json_metadata", pre, post) + paths = {tuple(r.path): (r.from_value, r.to_value) for r in records} + assert paths == { + ("json_metadata", "settings", "theme"): ("light", "dark"), + ("json_metadata", "settings", "density"): ("compact", "comfortable"), + } + + +def test_diff_layout_node_edit_emits_leaf_record_not_whole_node() -> None: + """The dashboard-header-text case that motivated Shape B: editing + one meta field emits a record at the leaf path with from/to carrying + only the changed string — not the surrounding layout-node object.""" + pre_node: dict[str, Any] = { + "id": "HEADER-id-1", + "type": "HEADER", + "meta": {"text": "VERSION 2!", "background": "WHITE", "headerSize": "MEDIUM"}, + "parents": ["ROOT_ID", "GRID_ID"], + "children": [], + } + post_node: dict[str, Any] = deepcopy(pre_node) + post_node["meta"]["text"] = "HEADER!" + records = _diff_layout_node("HEADER-id-1", pre_node, post_node) + assert len(records) == 1 + # Path is pure navigation: [node_id, ...leaf]. No verb prefix. + assert records[0].path == ["HEADER-id-1", "text"] + assert records[0].kind == "header" + assert records[0].operation == "edit" + assert records[0].from_value == "VERSION 2!" + assert records[0].to_value == "HEADER!" + + +def test_diff_layout_node_edit_emits_one_record_per_changed_leaf() -> None: + """Multiple meta fields change in one save → multiple records.""" + pre_node: dict[str, Any] = { + "id": "HEADER-id-1", + "type": "HEADER", + "meta": {"text": "Old", "headerSize": "MEDIUM"}, + "parents": ["ROOT_ID", "GRID_ID"], + "children": [], + } + post_node: dict[str, Any] = deepcopy(pre_node) + post_node["meta"]["text"] = "New" + post_node["meta"]["headerSize"] = "LARGE" + records = _diff_layout_node("HEADER-id-1", pre_node, post_node) + paths = {tuple(r.path) for r in records} + assert paths == { + ("HEADER-id-1", "text"), + ("HEADER-id-1", "headerSize"), + } + # Every record's operation is "edit" — both leaves changed from one + # populated value to another. kind stays "header" for both. + assert {r.operation for r in records} == {"edit"} + assert {r.kind for r in records} == {"header"} + + +def test_diff_layout_node_add_remove_move_unchanged_shape() -> None: + """Layout add/remove/move emit a single record carrying the minimal + node payload. Path is pure navigation (``[node_id]``); the verb lives + in ``operation``, the element type in ``kind``.""" + chart_node = { + "id": "CHART-x", + "type": "CHART", + "meta": {"chartId": 42, "sliceName": "Sales", "uuid": "u-1"}, + "parents": ["ROOT_ID"], + "children": [], + } + # Add + added = _diff_layout_node("CHART-x", None, chart_node) + assert len(added) == 1 + assert added[0].path == ["CHART-x"] + assert added[0].kind == "chart" + assert added[0].operation == "add" + assert added[0].from_value is None + assert added[0].to_value == { + "id": "CHART-x", + "type": "CHART", + "name": "Sales", + "chartId": 42, + "uuid": "u-1", + } + + # Remove + removed = _diff_layout_node("CHART-x", chart_node, None) + assert len(removed) == 1 + assert removed[0].path == ["CHART-x"] + assert removed[0].kind == "chart" + assert removed[0].operation == "remove" + assert removed[0].to_value is None + + # Move + moved_node = deepcopy(chart_node) + moved_node["parents"] = ["GRID_ID"] + moved = _diff_layout_node("CHART-x", chart_node, moved_node) + assert len(moved) == 1 + assert moved[0].path == ["CHART-x"] + assert moved[0].kind == "chart" + assert moved[0].operation == "move" + + +def test_diff_dashboard_layout_aggregates_records_across_nodes() -> None: + """End-to-end: multiple node edits in one save produce one set of + leaf records, ordered by node_id.""" + pre = _json.dumps( + { + "HEADER-a": {"id": "HEADER-a", "type": "HEADER", "meta": {"text": "A"}}, + "HEADER-b": {"id": "HEADER-b", "type": "HEADER", "meta": {"text": "B"}}, + } + ) + post = _json.dumps( + { + "HEADER-a": {"id": "HEADER-a", "type": "HEADER", "meta": {"text": "A2"}}, + "HEADER-b": {"id": "HEADER-b", "type": "HEADER", "meta": {"text": "B2"}}, + } + ) + records = diff_dashboard_layout(pre, post) + paths = {tuple(r.path) for r in records} + assert paths == { + ("HEADER-a", "text"), + ("HEADER-b", "text"), + } + assert {r.kind for r in records} == {"header"} + assert {r.operation for r in records} == {"edit"} + + +def test_diff_layout_node_edit_respects_depth_cap() -> None: + """A meta value deeper than ``_LAYOUT_META_DIFF_DEPTH`` is emitted + as an opaque sub-tree rather than walked further. Confirms the + cap is wired through from the constant — not a property of the + helper alone.""" + deep_subtree_pre: dict[str, Any] = {"l1": {"l2": {"l3": {"l4": "old"}}}} + deep_subtree_post: dict[str, Any] = {"l1": {"l2": {"l3": {"l4": "new"}}}} + pre_node: dict[str, Any] = { + "id": "X", + "type": "HEADER", + "meta": {"deep": deep_subtree_pre}, + "parents": ["ROOT_ID"], + "children": [], + } + post_node: dict[str, Any] = deepcopy(pre_node) + post_node["meta"]["deep"] = deep_subtree_post + records = _diff_layout_node("X", pre_node, post_node) + # Path is [node_id, ]. _LAYOUT_META_DIFF_DEPTH=3 bounds + # the recursion starting from the meta dict, so the deepest path + # captures at most 3 levels below ``X``. The leaf is emitted as an + # opaque sub-tree at the cap. + assert len(records) == 1 + assert len(records[0].path) <= 1 + _LAYOUT_META_DIFF_DEPTH + # The opaque leaf must still carry the entire change as from/to so + # restoration is lossless even when the cap fires. + assert "old" in str(records[0].from_value) + assert "new" in str(records[0].to_value) + + +def test_diff_slice_params_unknown_key_recurses_to_leaf() -> None: + """An unknown ``params`` sub-key carrying a nested dict no longer + emits the whole sub-tree on both sides — only the changed leaf.""" + pre = _json.dumps( + { + "custom_viz_options": { + "axis": {"y": {"format": "%d"}, "x": {"format": ".2f"}} + } + } + ) + post = _json.dumps( + { + "custom_viz_options": { + "axis": {"y": {"format": "%.2f"}, "x": {"format": ".2f"}} + } + } + ) + records = diff_slice_params(pre, post) + assert len(records) == 1 + assert records[0].path == [ + "params", + "custom_viz_options", + "axis", + "y", + "format", + ] + assert records[0].from_value == "%d" + assert records[0].to_value == "%.2f" + assert records[0].kind == "field" diff --git a/tests/unit_tests/versioning/test_pin_audit_columns.py b/tests/unit_tests/versioning/test_pin_audit_columns.py new file mode 100644 index 000000000000..36bf8cb04ef7 --- /dev/null +++ b/tests/unit_tests/versioning/test_pin_audit_columns.py @@ -0,0 +1,203 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Unit tests for ``pin_audit_columns`` in ``superset.versioning.baseline``. + +Locks in the SA-version-dependent semantic the helper relies on: calling +``attributes.flag_modified(parent, "changed_by_fk")`` causes SQLAlchemy +to include the in-memory value in the next UPDATE statement instead of +invoking the column's ``onupdate=callable`` default. This is the +mechanism that prevents a stale ``g.user.id`` from being written into +the parent's ``changed_by_fk`` when the synthetic flag-flush triggers +an UPDATE during an autoflush at a time when the test user has already +been deleted from ``ab_user`` (the original failure mode that motivated +``pin_audit_columns``; see ``baseline.py`` docstring). + +If a future SQLAlchemy version changes this behavior — i.e. ``onupdate`` +fires even when the column is in dirty attribute history — this test +fails and the cascade returns. That's the invariant we're guarding. +""" + +from __future__ import annotations + +from typing import Any + +import pytest +import sqlalchemy as sa +from sqlalchemy.orm import declarative_base, Session + + +def _make_dummy_mapped_class() -> tuple[Any, sa.engine.Engine]: + """Build a minimal mapped class with an ``onupdate=callable`` column, + backed by an in-memory SQLite engine. Returns ``(cls, engine)``.""" + + Base = declarative_base() # noqa: N806 — SA convention + + # Mutable counter so we can assert how many times onupdate fired. + onupdate_calls = {"count": 0} + + def _bump_counter() -> int: + onupdate_calls["count"] += 1 + return 9999 # the value onupdate would write if it fires + + class Parent(Base): + __tablename__ = "parent" + id = sa.Column(sa.Integer, primary_key=True) + description = sa.Column(sa.Text) + changed_by_fk = sa.Column(sa.Integer, onupdate=_bump_counter) + + Parent._onupdate_calls = onupdate_calls # type: ignore[attr-defined] + engine = sa.create_engine("sqlite://") + Base.metadata.create_all(engine) + return Parent, engine + + +def test_flag_modified_suppresses_onupdate_callable() -> None: + """The contract ``pin_audit_columns`` depends on: when an attribute + is marked dirty via ``flag_modified``, SQLAlchemy uses the in-memory + value rather than invoking the column's ``onupdate=callable``. + + The cascade fixed in sc-103156 T062 (and in PR #40451's discussion) + relied on this exact behavior — without it, the synthetic UPDATE that + ``force_parent_dirty_on_child_change`` triggers would stamp + ``changed_by_fk`` with whatever ``get_user_id()`` resolves to at flush + time, including stale user ids from a teardown autoflush. + + Setup note: uses ``expire_on_commit=False`` so the column stays + loaded in instance state after the initial commit. This mirrors the + listener's real-world flow, where the parent's attributes are + already loaded (the listener reads them via ``getattr`` before + calling ``flag_modified``, which forces a load). In the + ``expire_on_commit=True`` path the attribute would be expired and + ``flag_modified`` would raise ``InvalidRequestError`` — that case + is the production path ``pin_audit_columns`` catches and skips + (covered in ``test_pin_audit_columns_tolerates_invalid_request_error``). + """ + from sqlalchemy.orm import attributes, sessionmaker + + parent_cls, engine = _make_dummy_mapped_class() + Parent = parent_cls # noqa: N806 — declarative class, capitalized intentionally + session_factory = sessionmaker(engine, expire_on_commit=False) + with session_factory() as session: + # Seed with a valid value (mimics a row that was committed earlier + # with a real ``g.user.id``). + parent = Parent(id=1, description="initial", changed_by_fk=42) + session.add(parent) + session.commit() + + # Now: edit ``description`` (the column the listener actually + # flags) and pin ``changed_by_fk`` via ``flag_modified``. + parent.description = "edited" + attributes.flag_modified(parent, "changed_by_fk") + + baseline_count = Parent._onupdate_calls["count"] + session.commit() + + # Re-read from a fresh session (no shared identity map) to check + # what was actually written to the database. + with Session(engine) as fresh: + row = fresh.get(Parent, 1) + assert row is not None + # The invariant: ``changed_by_fk`` carries the in-memory + # value (``42``), not the onupdate-callable's return (``9999``). + assert row.changed_by_fk == 42, ( + f"Expected in-memory value 42, got {row.changed_by_fk} — " + "SA may have changed flag_modified semantics; " + "pin_audit_columns would no longer suppress get_user_id()" + ) + + # And the onupdate callable was NOT invoked. + assert Parent._onupdate_calls["count"] == baseline_count, ( + "onupdate fired despite the column being flag_modified — " + "SA version regression" + ) + + +def test_onupdate_does_fire_without_flag_modified() -> None: + """Sanity check / negative case: without ``flag_modified``, the + ``onupdate`` callable DOES fire on a regular update. Pins the half + of the contract we DON'T want for ``pin_audit_columns``.""" + from sqlalchemy.orm import sessionmaker + + parent_cls, engine = _make_dummy_mapped_class() + Parent = parent_cls # noqa: N806 — declarative class, capitalized intentionally + session_factory = sessionmaker(engine, expire_on_commit=False) + with session_factory() as session: + parent = Parent(id=1, description="initial", changed_by_fk=42) + session.add(parent) + session.commit() + + # Edit ``description``; do NOT touch ``changed_by_fk``. + parent.description = "edited" + baseline_count = Parent._onupdate_calls["count"] + session.commit() + + # Confirm onupdate fired exactly once. + assert Parent._onupdate_calls["count"] == baseline_count + 1 + + with Session(engine) as fresh: + row = fresh.get(Parent, 1) + assert row is not None + # And the value was overwritten by the onupdate callable. + assert row.changed_by_fk == 9999 + + +def test_pin_audit_columns_skips_missing_attribute() -> None: + """``pin_audit_columns`` must tolerate parents that don't carry the + audit attributes (e.g., a model variant without ``AuditMixin``). + Uses a bare object so ``hasattr`` returns False.""" + # pylint: disable=import-outside-toplevel + from superset.versioning.baseline import pin_audit_columns + + class NoAuditMixin: + pass + + parent = NoAuditMixin() + # Must not raise. + pin_audit_columns(parent) + + +def test_pin_audit_columns_tolerates_invalid_request_error() -> None: + """``pin_audit_columns`` catches ``InvalidRequestError`` raised when + an attribute is unloaded in instance state — e.g., on a freshly + constructed ``session.new`` instance whose attribute defaults haven't + fired yet. Without this guard, the listener would crash mid-flush + on dataset INSERTs.""" + # pylint: disable=import-outside-toplevel + from unittest.mock import patch + + from sqlalchemy.exc import InvalidRequestError + + from superset.versioning.baseline import pin_audit_columns + + class _HasAuditCols: + changed_by_fk = 1 + changed_on = None + + parent = _HasAuditCols() + + with patch( + "superset.versioning.baseline.dirty.attributes.flag_modified", + side_effect=InvalidRequestError("not loaded"), + ) as mock_flag: + # Must not raise — must swallow the InvalidRequestError per + # attribute and keep going. + pin_audit_columns(parent) + assert mock_flag.call_count == 2 + + +if __name__ == "__main__": + pytest.main([__file__, "-v"])