Benchmarks

An executed, reproducible HTTP-level benchmark comparing a Cratebase release build against PocketBase v0.40.2, run back to back on the same machine by one script (benchmarks/run.sh). Both servers are driven identically by a plain-fetch Bun script — no SDKs — across six categories (create, auth, search, search-wide, search-auth, delete) at four concurrency levels (1, 20, 50, 100), median of 3 runs after a discarded warm-up.

22 of 24 cells were faster on Cratebase in the latest idle-host run (2026-09-04). The standout: search at concurrency 50 — 49,986 vs 4,774 req/s, a 10.47x ratio, with p99 latency at 1.60ms vs 42.05ms. create at concurrency 100 was 11,960 vs 6,606 req/s (1.81x), p99 8.87ms vs 57.88ms. The two cells that weren’t faster — delete at concurrency 1 and 20 (0.95x, 0.97x) — are within noise of parity, not a regression: the same category is 2.23x/2.25x at c50/c100 in the same run.

Read honestly: the auth win (up to 3.87x) is a security-parameter difference, not engineering work — Argon2id at OWASP parameters is simply cheaper than PocketBase’s bcrypt cost-12, and shouldn’t be read as throughput work. This is a shared, virtualized sandbox, not a dedicated benchmark rig — relative comparisons within one run are more trustworthy than absolute numbers across runs.

Reproduce it yourself:

benchmarks/run.sh                       # builds, downloads PocketBase, runs everything
benchmarks/run.sh --skip-build          # reuse target/release/cratebase
benchmarks/run.sh --concurrency=1,20    # any bench.ts flag is forwarded

Full methodology, the complete table, and the WAL-checkpoint story behind an earlier delete regression are in benchmarks/README.md; raw numbers in benchmarks/results/.