BACK TO DOCS

BENCHMARK STREAM

Benchmark hygiene

A credible repo-CI run needs immutable images, clean compose output, and retained artifacts for every repetition. Treat GitHub-hosted runs as evidence with provenance, not as final official hardware.

Build and Release publishes exact tag families on non-PR runs: latest on the default branch, semantic v<major>.<minor>.<patch> release tags on main, immutable ci-<full-sha>, and Docker metadata sha-* tags. Docs-only changes intentionally skip that workflow unless manually dispatched.

Keep the 1 CPU / 350 MB envelope visible and report false positives, false negatives, and HTTP errors beside p99. The public scoring model heavily penalizes HTTP errors and high detection error rates.

CHECKS AND SOURCE CUES

  • Immutable image pin when benchmarking prebuilt code
  • docker compose config --quiet
  • 0 FP, 0 FN, 0 HTTP errors beside latency
  • GHCR tags: latest, vX.Y.Z, ci-<sha>, sha-*
  • Build and Release paths-ignore docs/**