ci: Rust (Windows) job intermittently lost the runner after 50 min (8 times since 2026-11-15)

未关闭
#520 6 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
4/5
预计耗时
3-5 天
新手友好度
55/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
github-actions, rust
领域
ci-cd, testing-qa

调研方向

Start by reading atlas_core::sys::kill_tree and its Windows descendants and force_kill logic, then trace the callers crew::proc::run_bounded and crew::engine::TurnProc. The issue identifies the suspected process ID reuse bug and describes a proposed fix; done means applying that fix and checking whether the Windows runner losses continue. If they do, add step timeouts and log upload as the issue suggests.

由索引模型根据 Issue 内容生成。

描述

area: ci os: windows type: bug

What

The job Rust (Windows, full workspace) (ci.yml → rust-gate, cargo nextest on windows-latest) sometimes never finishes: it ends exactly 50:00 min after it starts with the annotation "The hosted runner lost communication with the server. Anything in your workflow that terminates the runner process, starves it for CPU/Memory, or blocks its network access can cause this error.". No log is kept for the job (BlobNotFound), the Run ./.github/actions/rust-gate step is in_progress forever. A normal run takes 5–9 min; a rerun of the same commit passes.

Evidence (observed)

Scan of every ci.yml run since 2026-11-01 (631 Windows jobs that completed, via gh api .../actions/runs/<id>/jobs):

started (UTC) run head what
11-15 01:00 43660731836 af11513b PR
11-15 04:09 43675462339 08a6ccd2 PR
11-15 22:53 43771698009 0e292a3c PR #460 (diff: one taskcache test file)
11-16 11:17 43854338812 4303661f main dispatch (#424 merge)
11-16 14:22 43885042374 e041082a PR perf/lazy-hidden-webviews (desktop-only diff)
11-16 21:34 43961249893 d26c1780 PR #514

plus the first attempts of #507 (6a92ff64, 20:21) and #512 (962e4e78, 21:31) that were overwritten by their reruns: 8 incidents.

  • None before 2026-11-15 01:00 UTC in 316 earlier jobs; 6 of the 319 after (about 2 %). Every hung head contains 76eee2ce (#335) and the batch merged with it (21:50–21:57 -03 on 11-14: #332, #335, #338, #341, #342, #344, #345, #334, #327); the hung heads' diffs against main are unrelated to each other, so the cause is in main's code, not in a PR.
  • The runner image (windows-2025-vs2026), runner (2.338.2) and Hosted Compute Agent (20261103.629) are the same in good and bad runs (logs of successful runs on 11-14, 11-15, 11-16): not an image rollout.
  • #424 is not the cause: its first hang (0e292a3c, af11513b, 08a6ccd2) predates it, and almost all of its tests are cfg(unix).

Suspect (code defect proven by reading; link to the runner loss is inferred)

atlas_core::sys::kill_tree(root) on Windows (used by crew::proc::run_bounded, crew::engine::TurnProc, the pty control, the language servers):

  • it called descendants(root, rows) which did born(p) = start_time_secs(p).unwrap_or(0). For a root that has already exited (callers end the tree after wait_with_output) born(root) was 0, so every process whose recorded parent pid equals that number passed the born(child) >= born(parent) guard. Windows keeps ppid after the parent ends and reuses pids: any older process whose original parent had the same number (the CI runner's own children are such candidates) was killed with TerminateProcess;
  • force_kill(root) was also called on a pid whose process had ended: after run_bounded consumed the child (wait_with_output) its handle was closed and the number belongs to the next process.

#335 (TurnProc::kill_tree) and #352 (run_bounded → Tree::kill, git/gh calls of the crew) added these callers on Windows around the first incident. Precedent: #223 (kill_uid ending every process of the runner's user).

Not proven: no log of a hung job survives, so it is not known which process was ended, and which test ran. The Windows-visible tests that reach these callers are crew::project, crew::runner::github, gitrun, the pty/hostd Windows tests and the lsp server unit tests (most crew tests are cfg(unix)).

Other candidates checked (not the cause)

Every other kill in the workspace that a Windows test can reach acts on a pid its test spawned or one verified by name/start time (kill_orphans, kill_prior, retire_hostd, hygiene::reap start-time recheck, tunnel adopt is_ssh_name, cleanup_lock, desktop_agent::terminate is unix-only); port_conflict's fake pids only reach classify, never a kill; hostd_windows.rs waits are bounded (ping -n 30/60, timeouts).

Fix

PR: root without a creation time has no tree; caller and its ancestors never in a tree; run_bounded keeps the child's handle until it ended the tree. If the 50-minute losses continue after this, the next step is a step-level timeout-minutes and nextest --no-capture log upload on the Windows job so a loss leaves a log.

Separate failure, same day

runtime::port_conflict::tests::a_holder_in_our_pid_tree_is_own_and_anything_else_is_foreign on PR #514: classify treats pid == std::process::id() as own, the test calls 900 "somebody else's"; Windows pids are small multiples of 4, so a test process numbered 900 fails line 128 (the 900 assertion, which only fails when pid == own pid). Fixed in the same PR (classify_as(me, …)).

主要语言
Rust
星标
3
派生
1
平均合并
9 小时 51 分钟
30 天内合并 PR
310

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。

codeatlasdev/atlas 的其他 Issue

查看 codeatlasdev/atlas 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。