Skip to content

release-readiness.md

Latest commit

 

History

History
146 lines (123 loc) · 7.81 KB

release-readiness.md

File metadata and controls

146 lines (123 loc) · 7.81 KB
Raw

AgentFlow Release Readiness

Release line: v2.0.0

Package status: published to PyPI (agentflow-runtime, agentflow-client) and npm (@yuliaedomskikh/agentflow-client) via OIDC Trusted Publishers with SLSA provenance attestations — see dv2-multi-branch/RELEASE_STATUS.md for registry links and upload evidence.

Golden-topology status (2026-07-30): production candidate, not production accepted. Repository implementation and local contract gates are complete; clean-checkout OCI build + real Flink job submission smoke is PASS (perf/golden-flink-submission-2026-07-30.md); clean kind + Flink Kubernetes Operator + Helm deployment of the verified image on exact HEAD 36ed1ec is also PASS (perf/golden-operator-acceptance-2026-07-30.md). That acceptance used live Kafka runtime fixes; the kind acceptance scaffold is now tracked at k8s/acceptance/kafka-kraft.yaml with a unit contract (not a production Kafka claim). Lake-to-serving E2E, recovery, soak, rollback, and external security evidence remain pending as listed below.

Summary

AgentFlow ships an event-native metrics layer: business metrics that move when events happen, served over typed contracts to people, dashboards, services, and AI agents. The repository-level blockers from the 2026-07-23 audit are implemented: fail-closed CDC attribution, one operator-compatible PyFlink runtime, tenant-scoped replay identity, explicit lake and serving materializers, role-aware lifecycle cleanup, deterministic partial search, one canonical session job, narrow control-plane capabilities, and a verified common SDK surface with explicit language-specific differences. Runtime, evidence, Python-version, quality-gate, and SDK capability statements are tracked in the machine-readable project claims. Architecture decisions are recorded as ADRs in docs/decisions/.

Performance baseline

Source: docs/benchmark-baseline.json.

Endpoint p50 (ms) p99 (ms) RPS Gate Status
GET /v1/entity/order/{id} 55 300 4.24 p50 < 100, p99 < 500 ✅
GET /v1/entity/product/{id} 49 320 2.39 p50 < 100, p99 < 500 ✅
GET /v1/entity/user/{id} 38 290 3.07 p50 < 100, p99 < 500 ✅
GET /v1/metrics/{name} 53 220 7.27 informational ✅
POST /v1/query 74 370 5.22 informational ✅
POST /v1/batch 62 340 5.56 informational ✅

Aggregate run: 569 requests, 0 failures, 27.76 RPS, p50 56 ms, p95 260 ms, p99 330 ms. CI-runner thresholds are intentionally divergent from local hardware and documented in docs/perf/ci-hardware-gap-2026-05-24.md.

Code health

  • God-class split completed for auth, alerts, and query modules with compatibility imports preserved.
  • SQL injection exposure closed via parameterized queries and sqlglot AST validation. Tenant scoping is applied at every read surface, but the mechanism behind it was rebuilt after this release: what shipped here was a schema qualification that nothing provisioned, so it isolated nothing. The boundary is now a tenant_id column in each table's write key (ADR-004); see STATUS.md for what is proven on which store.
  • Flink critical paths covered by unit tests (session_aggregator, stream_processor).

CI gates

main is protected with 15 required status checks — lint, test-unit, test-integration, perf-check, helm-schema-live, schema-check, terraform-validate, bandit, safety, npm-audit, trivy, contract, build-smoke, sdk-ts, lock-check. Branch protection requires every one of them; force-pushes and deletions are disabled. Dependabot security updates (vulnerability alerts + automated security fixes) are enabled alongside the weekly version-update schedule in .github/dependabot.yml.

The required lint job now contains strict MkDocs and claims validation. The required test-unit job contains the 80% changed-code coverage gate and depends on the 3.11/3.12/3.13 python-compat matrix, so a failed compatibility lane cannot be bypassed merely because its job name is not a separate protected context. The same job enforces a 60% full-suite line/branch floor and separate 90% floors for security-critical modules. Codecov upload is non-blocking reporting until the repository is enabled in that external service; see operations/codecov-setup.md.

The 2026-07-30 closing remediation additionally covers two clean-install boundaries: the core-only API imports and serves health/entity/query without the optional pyiceberg package, and the final API image removes pip/setuptools/wheel after its hash-locked install and pip check. The cloud profile separately includes pyiceberg-core==0.7.0, whose stable-ABI build supports the full Python matrix and restores PyIceberg 0.11 write transforms; the optional MCP integration is capped below the breaking 2.0 API. Clean Mac Python 3.13 verification passed with MCP 1.29 and all 39 selected contract/MCP/Iceberg tests. Independent Mac validation with Trivy 0.70.0 reported zero HIGH/CRITICAL findings; details and evidence limits are in security-runtime-image-trivy-2026-07-30.md and dependency-compatibility-2026-07-30.md. All required GitHub checks must still be green on the exact final pushed SHA.

Scope

Component, contract, Helm, and replay tests validate the checked-in streaming, lake, serving, and deployment artifacts. They do not substitute for full live production-acceptance evidence.

Completed (limited scope, 2026-07-30):

  1. clean-checkout PyFlink OCI build and real job submission smoke on deproject-mac — see perf/golden-flink-submission-2026-07-30.md;
  2. clean kind + Flink Kubernetes Operator + Helm golden-topology deployment of the verified OCI image on exact HEAD 36ed1ec (stable hold, checkpoints growing, zero leadership flaps) — see perf/golden-operator-acceptance-2026-07-30.md.

These close submission smoke and Operator/Helm deploy only. They are not lake-to-serving E2E or production acceptance. Kafka on the acceptance stand required evidence-backed scaffold fixes (enableServiceLinks: false and controller quorum voters at 127.0.0.1:29093); that is recorded as acceptance-scaffold reproducibility debt, not a product source of truth from untracked prompts.

Still required for production acceptance:

  1. one tenant-scoped event observed through Kafka → PyFlink → Iceberg → ClickHouse → API, then replayed after checkpoint restore without duplicate (tenant_id, event_id) rows;
  2. a fresh four-hour soak, backup/restore, and rollback rehearsal on that same artifact and topology (the existing 2026-07-19 soak predates the Iceberg materializer);
  3. an external penetration-test report and remediation/retest evidence;
  4. GitHub Environment npm created with approval protection. The workflow now requires environment: npm and a matching release tag, but the repository settings API returned 404 for that environment on 2026-07-23, so approval protection is not yet evidenced.

Wiring AgentFlow to a live production source also needs inputs that live outside the repo: CDC source onboarding (runbook in docs/operations/cdc-production-onboarding.md), a public benchmark on production-grade hardware, and operator-owned credentials. None of the claims those would unlock are made here.