Honest benchmarks

These rows are the 2026-08-05 committed artifacts. They are not a claim of full Postgres or Supabase parity. A lower ratio means Librebase p95 was under the Postgres 16 or PostgREST p95 in that harness.

Core path vs Postgres 16

Manual 3-rep streak, mode embed_execjson, long-lived session IPC, 10,000 rows. Postgres 16 was lb-pg-bench:5433 (postgres:16). Host note in the artifact: local win32 plus that container — not a published same-bare-metal pair.

• point_lookup_with_index p95 ratios: 0.1856×, 0.1955×, 0.2512× (median 0.1955×). Gate ≤ 1.2×. Index implementation recorded as hash_map.
• Evidence: benchmarks/oltp-compare/results/nightly-streak.json, lidb pin d7f5cb5.

Range prefix vs Postgres 16

Release build, sorted_tree (in-memory ordered secondary, not a Postgres page B-tree). Same container Postgres. Recorded p95 ratios 0.2863×, 0.3657×, 0.3577× (median 0.3577×; the artifact note rounds that to 0.36×). Gate ≤ 1.2×.

• Evidence: benchmarks/oltp-compare/results/range-scan-streak.json, lidb pin 041b9a2.

Memory

• Linux VmRSS, empty project database plus a small warm pool: 3.797 MB (proc_vmrss). Gate was 64 MB. GitLab pipeline 27003, job 99197, sha 2bfbd3c, 2026-08-05.
• Registry-min profile on that runner: 4.859 MB against a 256 MB gate.
• Evidence: benchmarks/oltp-compare/results/rss-linux-librebase-lean.json.

HTTP vs PostgREST

Soft pass only. Per-rep max p95 ratios are 0.5879×, 0.7522×, and 0.5979× (median 0.5979×) against PostgREST on the same bench containers. This is a Python MVP surface, not a hard gate.

• Evidence: benchmarks/oltp-compare/results/http-streak.json.

What this page does not say

Connect an agent · Sign in