Migrate to Strawberry / FastAPI / Mangum (final sibling: container + PynamoDB)

[chrisbc]

Summary

Migrate solvis-graphql-api from Graphene 3 / Flask / serverless-wsgi to Strawberry / FastAPI / Mangum — the 4th and final sibling, run last because it adds the two biggest deltas: a container (ECR) Lambda and a PynamoDB ORM.

Epic: GNS-Science/nshm-toshi-api#359
Runbook: nshm-toshi-api/docs/MIGRATION_RUNBOOK.md
Worked log: docs/MIGRATION_LOG.md (branch migrate/strawberry)
Owner: @chrisbc

Phase 0 inventory — done

Captured in docs/MIGRATION_LOG.md. Key findings (several easier than the runbook's §A4 framing):

• PynamoDB is a single model — BinaryLargeObjectModel (5 attrs) already wrapped by a hand-written BinaryLargeObject class. Convert to pydantic + boto3 behind the unchanged wrapper API.
• Container entry is the serverless-wsgi handler.py workaround → Mangum replaces it; update Dockerfile CMD + serverless image.command.
• Repo already ships a cross-stage differential validator (ab_test/ + cli_ab_test, 9 query checks) → reuse as the Phase 5 validation tool and corpus seed.
• moto-based tests already exist → no testcontainers.
• Still on poetry; .yarnrc.yml lacks the age gate; both serverless-s3-local + serverless-dynamodb load at boot (runbook §4.7 risk); heavy container image (matplotlib/solvis/shapely).
• Secrets posture = repo-level static AWS keys + single TEST env (the Model §4.2 caveat, not AWS_TEST/AWS_PROD).

Plan (tailored — full detail in the log)

• Phase 0 — inventory, branch migrate/strawberry
• poetry → uv via the dockerbash poetry2uv skill (running externally on the branch)
• Phase 0 artifacts — legacy SDL baseline (guard stdout) + vendor the A/B query set as corpus
• Phase 1 — FastAPI/Mangum + container entry swap; .yarnrc.yml age gate; drop serverless-wsgi
• Phase 2 — PynamoDB→pydantic+boto3 (keep wrapper API); port ~25 graphene types to Strawberry at parity
• Phase 3 — tests to Strawberry (keep moto); CI → -uv workflows; remove branches: filter
• Phase 4 — container build/dep-install path; yarn resolutions hygiene; vuln bumps; image-size pass
• Phase 5 — deploy to test; cli_ab_test prod-vs-new differential validation; promote + prod watch; file runbook feedback PR

Status

Phase 0 inventory complete; branch pushed and awaiting the external poetry2uv run before Phase 1 code begins.

[chrisbc]

Migration engineering complete — stacked PRs #88–#94

All phases implemented + verified; only the prod cutover (AWS creds + go-ahead) remains.

Phase	PR	State
0 — pre-flight (poetry→uv, SDL baseline, corpus incl. kororaa)	#88	✅
1 — FastAPI/Mangum + container entry swap	#89	✅
2a — PynamoDB → pydantic + boto3 (wrapper API kept)	#90	✅
2b — schema port, SDL byte-identical	#91	✅
3 — runtime resolvers, 11 differential-parity checks	#92	✅
4 — deploy/CI/deps hardening (360MB image, pip-audit clean)	#93	✅
5 — cutover plan (pre-staged)	#94	🟡 deploy held

Verified: SDL byte-identical to legacy; full surface runtime-equivalent (in-process differential harness caught real bugs — relay global-id, float defaults, coercion); amd64 container builds + handler imports in-image; no runtime vulns.

Remaining (needs you): deploy to test → cli_ab_test prod-vs-test → promote → post-healthy legacy cleanup. See docs/PHASE5_CUTOVER.md.

[chrisbc]

✅ Merged to deploy-test + deployed to the TEST stage

Combined PR #95 merged (one merge → one clean deploy, per runbook G7; #88–#94 landed via it). The deploy-aws-lambda run succeeded:

• tests + ruff + mypy ✓
• ECR container build + test-stage deploy ✓
• deploy smoke query QueryRoot{about} ✓

The Strawberry app is live on test — container packaging, Mangum entry, and the full schema verified in the real Lambda env.

Folded in at merge: a CI fix (drop the follow_untyped_imports regression poetry2uv added + a SetOperationEnum type stub → whole package mypy-clean), and a rebase over 3 concurrent deploy-test commits (#81/#82/#84).

Remaining (needs you — AWS creds + prod go-ahead)

1. cli_ab_test WORK/ab.toml -A prod -B test -v — differential prod-vs-test (the built-in harness).
2. Promote deploy-test → main (prod) with the pre-staged revert PR + ~30-min watch.
3. Post-healthy legacy cleanup (delete graphene schema.py/Flask/handler.py, rename strawberry_schema.py → schema.py, drop legacy deps). See docs/PHASE5_CUTOVER.md.