feat(bench): track benchmark and memory results over time instead of a single-run gate

未关闭
#269 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
12/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
冷清
技术栈
github, github-actions, go

调研方向

Start with the regression job in .github/workflows/bench.yml and the baseline in tools/bench/baseline.json, then find CheckRegression to see why it stops at the first failure. The issue asks for an OpenSpec change to PERF-02 and INDX-06 before any code, and the tool choice (github-action-benchmark or Bencher) is left to the spec. Done means a spec is agreed, the tool is picked, and a per-phase metrics output is designed, which is design work rather than a first contribution.

由索引模型根据 Issue 内容生成。

描述

area:ci enhancement priority:low spec-needed
What exists today?

The regression job in .github/workflows/bench.yml runs each night. It indexes the synthetic corpus three times, takes the median, and compares it with one stored baseline in tools/bench/baseline.json. It fails at a 10% throughput drop, at 15% more peak RSS, or above 4 GiB. It never blocks a merge. The publish job reports absolute numbers each week.

How does it behave now?

The gate measures the host as much as the code, and it keeps no history.

  • Two runs on the same commit differed by 21% in throughput (#252).
  • After #250, the three trials of one night spread by up to 12% in peak RSS.
  • The job failed every night from 3 to 7 November. On three of those nights, its message named only throughput, because CheckRegression returns at its first failure. The 40% RSS growth from #250 stayed hidden.
  • No one can see when a number started to move, or which phase of an index moved it. The #265 investigation rebuilt that by hand from profiles.
How should it behave?
  1. Each nightly run and each dispatch records its results as a time series. The series holds throughput, peak RSS, query latency, node and edge counts, and numbers per phase of an index (extract, resolve, write): bytes allocated, live heap at the end of the phase, GC cycles and batch bytes.
  2. The binary reports the per-phase numbers from runtime/metrics, so the runner reads them and no profile is needed.
  3. A page shows the series as charts. An alert fires on a change that lasts over several runs, with a statistical threshold, and not on one run.
  4. Optional: a pull request that changes Go code gets a report of a paired comparison. The job builds the base and the head on one runner, alternates their trials, and compares them with benchstat. The report does not block the merge.
  5. The tool comes from OSS. The candidates are github-action-benchmark and Bencher. The spec picks one, and the implementation checks its current version.

This issue replaces #252.

What kind of improvement?

Performance

Would this change existing behavior?

Yes. The nightly job stops being a pass or fail result from one run against one baseline. That changes PERF-02 and INDX-06, so it needs an OpenSpec change first. If the binary reports per-phase numbers through a new flag or output, that flag is a change to the CLI surface, and the spec names it.

🤖 Generated with Claude Code

主要语言
Go
星标
0
派生
0
平均合并
7 小时 42 分钟
30 天内合并 PR
111

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

seanb4t/codegraph-go 的其他 Issue

查看 seanb4t/codegraph-go 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。