mirror of
https://github.com/block/buzz.git
synced 2026-10-27 04:30:52 +00:00
2270b66e14
## Summary
The partition Postgres tests dropped their scratch schema with a
`drop_schema` call at the end of each test body. When a test panicked,
that call never ran, so the schema leaked into the shared database,
along with any `{schema}_*` sibling schemas or reader role the test
created. A reviewer on #7993 hit this locally: the leaked schema plus a
database-wide catalog query made later runs fail. CI hides it because
the test wrapper gives each process its own database.
`scratch_pool` / `scratch_pool_with_max_connections` now return `(pool,
ScratchSchema)`. The guard owns the admin pool and schema name and
cleans up on `Drop`:
- **Fresh thread and connection.** `Drop` runs on the test's runtime
thread, which can't drive the test's pools while it blocks, so cleanup
runs on its own thread with its own runtime and connection.
- **Terminates the test's backends first.** Both pools record each
connection's backend pid. A test that panics mid-transaction keeps its
locks, because the rollback would only run on the blocked runtime.
Without this step `DROP SCHEMA` waits on those locks.
- **Drops `{name}` plus every schema and role named `{name}_*`.** This
replaces the per-test sibling-schema and `DROP ROLE` cleanup. The reader
role is now named `{schema}_reader`.
- **Bounded.** `lock_timeout = '10s'`. Failures print a warning and
never panic inside `Drop`.
The 52 call sites are migrated mechanically. Test-only change.
### Related issue
Follow-up from review on #7993. No existing issue found.
### Testing
- New `scratch_schema_is_dropped_when_a_test_panics_holding_a_lock`: the
test body creates a sibling schema, takes `ACCESS EXCLUSIVE` on a
scratch table in an open transaction, and panics. The test then asserts
no `{name}` / `{name}_*` schema remains.
- Proved it can fail: with the backend-termination step disabled,
cleanup hits the lock timeout and the test fails, reporting both leaked
schemas.
- Partition Postgres lane: `scripts/postgres-test-run.sh -p buzz-db
--lib -E 'test(partition)'`, 53/53 passed.
- The same filter run directly against the shared `buzz` database: 53/53
passed, then 0 `partition_audit_test_%` schemas and 0
`partition_audit_%` roles remained.
- `cargo clippy -p buzz-db --tests` clean. Pre-push hooks passed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: Ian Oberst <ioberst@block.xyz>
Co-authored-by: Ian Oberst <ioberst@block.xyz>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>