#[pg_bench] — in-process benchmark harness (pgrx 0.18) for the perf tickets

[styk-tv]

pgrx 0.18.0 shipped #[pg_bench] — the benchmarking counterpart to #[pg_test].
pgRDF uses it zero times. All our benchmarking is external to the extension.

What it gives us

Criterion-style benchmarks that run inside a live Postgres backend, with
direct SPI access and no round trip through psql:

• persistent history tables in a pgrx_bench schema that survive extension
  reinstalls, so runs compare across versions;
• git metadata + pg_settings snapshots recorded per run;
• per-benchmark setup via #[pg_bench(setup = ...)];
• transaction modes (shared, subtransaction_per_batch,
  subtransaction_per_iteration);
• Criterion's mixed-bootstrap T-test comparison model.

Run with cargo pgrx bench.

Why this matters to the roadmap

Several open tickets are performance bets with no in-repo instrument to settle
them:

• C5 · right-sizing — auto-fit the slice + settings to operator RAM/cores #22 (C5 right-sizing — auto-fit slice + settings to operator RAM/cores)
  needs exactly a measurement record tied to a pg_settings snapshot. That is
  what #[pg_bench]'s history tables are.
• Port catalog-build / slicer-estimator analysis into Rust pgRDF API (index-only, resolve-once) #34 (slicer-estimator) needs cost calibration.
• Staged loader: set-based cross-load dedup (keep parallel fast path on a populated dict) #18 ("keep parallel fast path on a populated dict") is an unfalsifiable
  claim without a benchmark regression guard.
• C3 · carve index reorganization — full 6-way hexastore + pg_trgm string index #20 (6-way hexastore vs 3) is a pure perf trade-off.
• Closure-expanded type catalog precompute pass #15 (closure-expanded type catalog precompute) is a perf bet.

Scope sketch

A benches module behind a feature gate, starting with the paths we already
argue about: dictionary intern rate, property-path expansion, staged-loader
phase timings.

Why P2, not P1

This is new capability, not foundation. It does not block the run and nothing
regresses without it.

Note on results discipline

Micro-benchmark numbers produced by this harness are pgRDF behaviour and are
disclosure-safe. Full-scale run results stay where they already go and do not
belong in this repo.