[Flaky test] TestBackgroundWorktreeIsolationAndMerge

未关闭
#4,922 0 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

@TDnorthgarden 已经在做这个了。

开始于 2026年11月13日。

  • #4923 来自 @TDnorthgarden —— 未关闭

评估

难度
3/5
预计耗时
半天
新手友好度
20/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
停滞
技术栈
git, go
领域
backend, testing

调研方向

Start with the concurrent launch in pkg/runtime/background_worktree_test.go (lines 134-139) and the completed assertion at line 156, then find where the background agent path runs git worktree add. Run the repro loop from the issue to see the commondir failure. Done means the flaky subtest passes reliably, either by serializing per-repo creation or retrying transient failures. PR #4923 is already open against this, so check it before starting.

由索引模型根据 Issue 内容生成。

描述

area/runtime area/testing
Description

Failed once on PR #4907's test-linux (run 43133474124, first attempt, Nov 12). The linked=false subtest launches two background agents that create their git worktrees concurrently in the same repo (pkg/runtime/background_worktree_test.go:134-139); task first failed with

creating background worktree: creating git worktree: exit status 128:
Preparing worktree (new branch 'worktree-agent-bb8bc299-...')
fatal: failed to read .git/worktrees/agent-788eb22e-.../commondir: Success

while task second completed normally — its worktree is exactly the admin entry first's git worktree add stumbled on. A short read of a half-written commondir (errno 0 ⇒ "Success") means one git worktree add read another's in-flight administrative entry and exited 128, so the task was marked failed and the test's completed assertion (background_worktree_test.go:156) failed.

I reproduced the git-level race locally (linux/amd64, git 2.51.0), independently of the test harness: 16 concurrent git worktree add in a fresh repo, 400 rounds — 51 rounds hit exactly this error (fatal: failed to read .git/worktrees/<name>/commondir: Success, plus an ENOENT variant), with several concurrent adds stumbling over the same half-written entry within one round. The test itself only races 2 concurrent adds, which matches its low flake rate; CI runner I/O widens the window.

Expected Behavior

Concurrent background worktree creation in one repo is safe; a transient git failure of this kind doesn't fail a task the test expects to complete.

Actual Behavior

Intermittently, one concurrent creation fails on a half-written admin entry and the task is recorded as failed.

Steps to Reproduce
  1. Run the repro script (pasted under Error output) on Linux
  2. It races 16 concurrent git worktree add in a fresh repo, 400 rounds
  3. Measured here: 51/400 rounds hit the error below
Docker Agent version

No response

OS & terminal

No response

Model used

No response

Error output
hits=0
for r in $(seq 1 400); do
base=$(mktemp -d)
git init -q $base/repo
git -C $base/repo commit -q --allow-empty -m i
for i in $(seq 1 16); do
  ( git -C $base/repo worktree add $base/w$i HEAD >/dev/null 2>$base/err$i ) &
done
wait 
grep -q commondir $base/err* 2>/dev/null && hits=$((hits+1)) 
rm -rf $base 
done 
echo "rounds with a commondir failure: $hits/400"   # measured: 51/400
Screenshots

No response

Additional context

report-main-failures only auto-files flakes seen on main, so this PR-run failure isn't in the tracker yet. Suggested direction:
serialize per-repo worktree creation in the background-agent path, or retry a creation that fails with a transient commondir read error.
Happy to work on the fix.

主要语言
Go
星标
4.7k
派生
565
平均合并
18 小时 52 分钟
30 天内合并 PR
221

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

  • 提供 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 没有贡献指南

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。

docker/docker-agent 的其他 Issue

查看 docker/docker-agent 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。