本文へスキップ
Kitbay
スキル

perf-comparison

Apache-2.02026年11月11日 更新50,877GitHubスタースクリプトあり

A maintainer skill in the ClickHouse repository for deciding whether a change in a performance test result, between a pull request build and a reference build, is real, noise, a flaky test or already known.

スキルの内容

A maintainer skill in the ClickHouse repository for deciding whether a change in a performance test result, between a pull request build and a reference build, is real, noise, a flaky test or already known. Evidence comes first from ClickHouse's performance dashboard and its API, then from master status counts, CI artifacts and finally local runs of the perf.py script against two explicit binaries. The agent works up a ladder of checks: every changed row, repeats across runs, confidence tier, master history, related metrics, code coverage and flame graphs, stopping once the verdict is clear. It never calls a regression real from one row, never mixes architectures, and does not file issues or comment unasked.

できること

  • Inventory changed rows in a performance run
  • Compare results against master history
  • Rerun performance queries locally on two builds
  • Give a verdict with a confidence level