Commit graph

1 commit

Author SHA1 Message Date
ZdenekSrotyr
f35b53dba4
fix(db): get_analytics_db() singleton — close #163 (release 0.54.5) (#286)
* fix(db): get_analytics_db() singleton mirroring get_system_db pattern

Closes #163. Pre-fix `get_analytics_db()` opened a fresh
`duckdb.connect()` on every call; most callers don't `.close()` the
returned handle, so each leaked connection held a WAL ref + FD until
GC kicked in. Under load this manifested as "too many open files" or
DuckDB lock contention on the analytics DB.

Singleton + cursor-per-call mirrors `get_system_db()` (lines 882-904
of src/db.py) — one underlying connection persists; callers that
.close() the returned handle only close the cursor. Re-opens
transparently when DATA_DIR changes (test fixtures that swap data
dirs across cases).

`get_analytics_db_readonly()` deliberately stays per-call —
each invocation re-ATTACHes extract.duckdb files into a fresh
read-only context, so caching the connection would require careful
re-ATTACH bookkeeping the read-only path doesn't currently do.

New `close_analytics_db()` mirrors `close_system_db()` (best-effort
CHECKPOINT then close, swallow exceptions). Wired into the FastAPI
shutdown hook in `app/main.py` alongside `close_system_db()`.

5 tests in `tests/test_analytics_db_singleton.py` pin the contract:
- caches connection (two calls → same _analytics_db_conn)
- closing a cursor doesn't close the underlying connection
- DATA_DIR change → singleton drops + reopens
- thread-safe (16 concurrent calls share the singleton)
- close_analytics_db() clears state + reopen works

Out of scope (per #163): auditing all caller sites to drop their now-
redundant `.close()` calls. Closing a cursor is harmless; the
production benefit is the connection cap.

Verified: 4525 passed locally (4520 + 5 new), 1 pre-existing fail in
test_readers_in_pre_init_dir (subprocess timeout, on origin/main, no
relation to this diff).

* release: 0.54.5 — close #163 analytics_db singleton

Last commit on the PR per CLAUDE.md hard rule. Patch bump (0.54.4 →
0.54.5) for the get_analytics_db() singleton refactor.

No DB migration; no operator-facing config change. Internal-only
cleanup of a known FD-leak under load. Closes #163.
2026-05-13 16:11:13 +00:00