Performance

Argon's performance numbers live in exactly one place: the public, reproducible benchmark suite.

github.com/argon-lab/benchmarks — pinned engine refs, recorded environment, reproducible with docker compose up; results are committed to RESULTS.md there.

This is policy, not an accident. Earlier versions of these docs quoted numbers ("1ms branching", "86x faster", "37,905+ ops/sec", "220,000+ queries/sec") that could not be traced to a reproducible run, so we removed them all. A performance claim you cannot reproduce is marketing, and it does not belong in documentation.

What the benchmarks measure today:

• Branch creation latency — a branch is a metadata write; the suite measures p50/p99 on projects with substantial history, and the storage cost per branch.
• Snapshot effectiveness — cold-read latency with and without snapshots, i.e. what bounding replay depth actually buys.
• Write capture overhead — ingest lag between a driver write landing in the physical database and its WAL entry existing.

If you publish an Argon number anywhere — a README, a blog post, a talk — it must come from a linked RESULTS.md run. Regressions the local regression canaries can't catch show up there; treat the suite as the source of truth.

In-repo performance tests (tests/wal/*_performance_test.go) are deliberately regression canaries with loose thresholds — they catch order-of-magnitude regressions in CI, they are not benchmarks, and their numbers must never be quoted.