Skip to content
MarginaliaDocsGitHub
Contents
handovers/2026-08-18-session-a7-test-postgres.md

Rendered from the repository — the file stays the source of truth.

Session A7 — chore/test-postgres (D-9: real Postgres for DB-backed tests)

Date: 2026-08-18 · PR: #33 · Spec: product/feasibility.md D-9

Goal and outcome

Implement D-9: retire PGlite (exit-99 under Bun, cold-WASM-init CI flakes, found by B2) and run database-backed tests against a real Postgres with real pgvector, locally and in CI, removing both INTERIM CI workarounds.

Done. 143/143 tests green (18 files), exit code 0, on both backends; zero test files edited; PGlite dependencies removed; CI runs plain bun test against a pgvector/pgvector:pg17 service container.

The choices D-9 left open, and how they were resolved

  1. Plain pgvector image vs supabase start stack → both, each where it’s strongest. Locally tests default to the Supabase stack’s Postgres on :54322 (supabase start is already the required local-dev setup — zero new commands, closest to hosted; verified the non-superuser postgres role can CREATE DATABASE and install pgvector 0.8.2). CI provisions pgvector/pgvector:pg17 as a service container: measured 12.1 s pull (~150 MB compressed) + 3.8 s to pg_isready, vs ~8 GB / 13 containers for the Supabase stack on an uncached runner. Same factory code either way; TEST_DATABASE_URL overrides the server.
  2. Isolation → one throwaway database per test file (marginalia_test_<pid>_<n>; bun runs files sequentially in one process, pid separates parallel invocations). Leftovers are swept at the start of the next run with plain DROP DATABASE (no FORCE — an in-use database from a concurrent run is skipped, not killed). Verified: 11 leftovers from a full run were all swept by the next run.
  3. Fidelity improved, not just preserved: the factory now applies the hand-written enable_pgvector migration verbatim (recreating the extensions schema + Supabase’s search_path first), then the Drizzle journal — PGlite had installed the extension into public. RLS/storage migrations stay intentionally unapplied (app-layer authz under test), unchanged.

Verification evidence

  • bun test vs Supabase stack: 143 pass, 0 fail, exit 0, 6.4 s.
  • TEST_DATABASE_URL=…54329… bun test vs the CI image: 143 pass, 0 fail, exit 0, 5.8 s.
  • bun run build green, tsc --noEmit green.
  • Real-engine payoff already visible: pgvector itself rejects the wrong-dimensionality write in ingestion-service.test.ts (expected 2000 dimensions, not 3 — that error in test output is an intentional error-path test, not a failure).

⚠️ Open item for the foreman: pre-existing lint failure gates green CI

Main’s CI has been red at the Lint step since the #28/#29 merges (before this session): 2 × react-hooks/set-state-in-effect in src/components/studio/audio-overview-dialog.tsx:53 and src/components/studio/studio-panel.tsx:178, plus one unused eslint-disable warning. Reproduced on a clean main checkout. Components are hard out-of-bounds for A7, so PR #33’s CI stops at Lint and the Test step never executes there — the proof of the restored exit-code contract is the two local full-suite runs above. Once lint is fixed (owning lane, or a two-line foreman-authorized fix), re-run CI on #33 to see the Test step green with the service container. No workaround was reintroduced.

Boundary deviations (all flagged in the PR)

  • apps/webapp/package.json + bun.lock: removal of the two PGlite dev-deps (no additions).
  • Correct-the-record edits outside the listed files: product/architecture/development.md (living doc, updated in place), product/history/{supabase,infrastructure}.md and handovers A1/B2 (dated annotation blockquotes) — all described the PGlite era as current.
  • TEST_DATABASE_URL is read with a safe default but not declared in .env.schema (file out of boundary); consider adding it as optional.

Gotchas for future sessions

  • Test databases (marginalia_test_*) linger on the local Supabase Postgres between runs by design; any bun test sweeps them, and they’re safe to drop manually. supabase db reset doesn’t remove them.
  • The factory needs a role with CREATEDB on the target server; both default backends satisfy this.
  • If Docker/supabase start isn’t running, DB-backed tests fail fast with an actionable error naming the one-command fix; pure-logic tests are unaffected.