[auto-review #1430] fix CI failures in DB backend benchmark

[Chrisclone]

Auto-review fixes for #1430 ([Ignore] WIP Benches). Targets bench so it flows into your PR.

Issues found

All CI failures on pixeltable#1430 trace to the new tests/test_db_backend_bench.py; pixeltable/env.py (external-Postgres pgvector bootstrap) is sound and left untouched.

1. CI failure — test jobs (minimal / standard-ubuntu / standard-windows / validate gate). The benchmark was missing the @pytest.mark.benchmark marker, so it ran in the default CI matrix instead of being excluded. Every other benchmark in the repo (tests/benchmarks/*, test_query.py) carries this marker, and pyproject.toml addopts deselects benchmark. Running it spun a docker pgvector container per config and ate ~30-60 min per job before failing.

2. CI failure — "FATAL: sorry, too many clients already" at [local-1cpu-8-16]. The busiest sweep point opens max(_N_PROCESSES) * max(_N_THREADS) = 8x16 = 128 pooled connections at once (each process has pool_size=n_threads), which exceeds Postgres's default max_connections=100. Real bug in the benchmark — that data point could never complete.

3. CI failure — static-checks (ruff PLW0603). global _PROC_ENGINE in the ProcessPoolExecutor initializer tripped the lint rule. static-checks runs ruff on all files regardless of markers, so this needed fixing independently.

Changes made

• Added @pytest.mark.benchmark to test_db_backend_direct_bench — right-layer fix that keeps this docker-spinning benchmark out of the default suite (fixes [CI Fix] PXT-1063 GC dispatch — fix premature gc_targets for cross-context slots #1 and the validate gate).
• Raised the local container's max_connections to max(_N_PROCESSES) * max(_N_THREADS) + 16 (derived from the sweep, so it stays correct if the sweep changes) via -c max_connections=... on docker run (fixes [CI Fix] [PXT-1065 + PXT-1066] Add export_csv and export_json #2).
• Added # noqa: PLW0603 on the global statement, matching repo convention (6+ existing uses) (fixes [Quality Review] PXT-1047 string to_sql — fix replace_re backreference divergence #3).

Validated locally

• python3 -m py_compile on both changed files — OK.
• ruff check and ruff format --check on both files — pass ("All checks passed" / "already formatted").
• Could not run pytest/collection or spin the docker backend here — the available interpreter has no pixeltable install, and I didn't run docker. Marker-based deselection is mechanically identical to the repo's other benchmarks, so I'm confident CI's default matrix will now skip it.

False-positive risk

• max_connections fix only covers the local-1cpu docker backend. The planetscale backend (gated behind PXT_BENCH_PLANETSCALE_URL) opens the same 128 connections and could hit PlanetScale's own connection ceiling — not something I can set server-side. If you run that backend at 8x16, watch for the same error.
• Intent check: if you wanted this benchmark to run as a gating test in CI (rather than on-demand), the @pytest.mark.benchmark marker is the wrong call — but given it's a timing benchmark with only an assert n > 0, excluding it matches how the repo treats benchmarks.
• I did not run the benchmark end-to-end, so I haven't empirically confirmed 128 connections succeed at max_connections=144; the arithmetic is straightforward but worth a sanity check on your first real run.

🤖 Generated with Claude Code