Repository navigation
Conversation
(ai: vllm-qwen/qwen3.8-27b) 进度汇报(此前会话被强制重启一次,现继续): 查重: 不重复。 同家族相关项: #3923 是原始 split-brain 报告; #3952(已合并)修的是"日志点名双方"; #4989(已合并)用 标签: bug / scope/storage / p2(缺陷类;writer marker 属 state 目录持久化面;纯观测误报、可用但有噪音,非 p1)。 审核进展(已完成方向与回归审查):
下一步: 按"小问题直接修"把上述修复作为标注 commit 推到 PR 分支(若 fork 不允许 maintainer 编辑则改为贴出补丁),随后给出八问结论。定向测试已在 PR head 全绿(12/12),全量交给 CI。 |
…k via shared processStartMs (ranxianglei#5028)
… younger-twin negative row (ranxianglei#5028)
(ai: vllm-qwen/qwen3.8-27b) 审核结论(完整八问)。查重: 不重复 —— #3923 是原始 split-brain 报告,#3952(已合)修日志点名,#4989(已合)用 修复已推送到 PR 分支(maintainer edit,两个标注 commit,API 核实 head 已更新为
另:master 在审核期间前进了 4 个 commit(全部 test/CI-only:#5017 ci-fixes、#4982 e2e pin),我已把当前 master 合入 PR 分支( 八问:
验证证据(本地定向,全量交 CI):
判定:可合(待 CI 在新 head 上绿;mergeable=True,base 仅落后 test/CI-only 提交,后台重试 merge commit 推送中)。合并动作留给你。 |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 重启后进度汇报:审查工作本身已完成(结论见上一条八问回复),卡在最后一步推送上——本环境到 GitHub 的 git 传输间歇性断连,之前的后台重试循环随进程重启被杀。现已直接推成功并验证:
等新 run 结果,全绿后在此补终局一行。 |
Sync current master ahead of re-review of ranxianglei#5028 (ranxianglei#3923 same-PID container generation gate in served-markers). Clean 3-way merge: no conflicts; hot-file src/server.ts auto-merged keeping both master's new imports and the PR's processStartMs wiring.
|
(ai: vllm-qwen/qwen3.8-27b) Synced origin/master into this branch as part of the merge-guard: head is now 41f748e and fully contains the current master tip (a8c907d). Both sync merges were clean auto-merges with no conflicts — none of this PR's files (served-markers.ts, server.ts, port-owner.js) were touched by the intervening master commits (#5169/#5159/#5167/#4625 etc.). Local typecheck is green and the served-markers suite passes 13/13 (incl. the two same-PID-generation regression tests). The branch is up-to-date with master and mergeable; CI is re-running on the new head. |
… — union of parallel guard syncs (full master 44067a5 + ci-fixes)
|
(ai: vllm-qwen/qwen3.8-27b) Merge-guard re-sync: master advanced again to 44067a5 — merged it in cleanly (zero conflicts; none of this PR's files were touched by the new master commits). A parallel guard worker had separately merged origin/ci-fixes onto this same fork branch (8242ab3); instead of clobbering it I union-merged both lineages into head 30d794f, which now carries full latest master + the ci-fixes convergence + this PR's served-markers fix (the 3 PR files were byte-identical across both workers, so no semantic conflict). Local typecheck green, served-markers suite 13/13. Branch is up-to-date with master and mergeable; CI re-running on the new head. |
… on restart In containers processes virtually always run as PID 1; after a restart the old writer marker (pid: 1) still passes isPidAlive() against the new process, so findContestedWriters false-positives [split-brain] on every boot. Fix: record startedAt (epoch ms) on WriterIdentity/WriterMarkerFile; a marker sharing our PID whose startedAt (or mtime for older builds without startedAt) predates this process boot by more than SAME_PID_GENERATION_MARGIN_MS is skipped as a prior generation. Integrated onto master's contestedServedAt helper (#4626); both mechanisms keep working. Original commit c115e35 of PR #5028 (external contribution by @fathonix; later review-fix commits of the same PR condensed into this one — full credit retained via git author).
…glei#5028 conflict 3 conflicts resolved by taking master's side (master is a verified-CI-green superset of the PR's ranxianglei#3923 fix): the same-PID generation guard is already in master (subsumed), plus master's contestedServedAt() refactor + disposedAt alias (ranxianglei#4626). Local verification: typecheck clean, tests/served-markers.test.ts 15/15.
(ai: vllm-qwen/qwen3.8-27b) Conflict resolved for PR #5028 — the PR is now mergeable (no conflict). What I did this run:
CI: re-running (triggered by the push). The mirror cache hasn't captured the new run yet (last poll 2026-10-30), and I did not gh-poll per policy. Local targeted verification (typecheck + the touched module's test) is green; the full-suite verdict is CI's. 八问(中文)
收工判定:冲突已解决,PR 现可合并( |
[ework] CI fix verified green ✓ |
…ranxianglei#5028 to current master Master advanced 90 commits (c26c31a -> 259cd87) after the prior conflict-resolution merge; none touch the PR's files (served-markers.ts / server.ts / served-markers.test.ts), so this is a clean update that keeps the PR validated against the master it will merge into.
(ai: vllm-qwen/qwen3.8-27b) This run: confirmed the conflict was already resolved, then synced the PR to current master.
Net result: conflict resolved, PR in sync with master, mergeable, CI re-running (expected green). 八问(中文)
收工判定:冲突已解决,PR 已对齐最新 master( |
(ai: vllm-qwen/qwen3.8-27b) Status after restart — the work is complete and holding (nothing to re-do).
The conflict was already resolved and the PR is in sync with master (full 八问 in my previous comment). I did not gh-poll per policy — the mirror cache will reflect the final CI result as the run completes. |
[ework] CI fix verified green ✓ |
(ai: vllm-qwen/qwen3.8-27b) Review verdict: do NOT merge — this PR is a no-op. Its fix is already in master. Please close it as superseded by Critical finding (verified against the pushed head The PR's entire contribution — the #3923 same-PID container-restart guard — is already merged into master via commit After that, the merge-guard / ci-engineer syncs repeatedly merged master back into this branch, so the two trees converged. The PR now has zero net changes:
Verdict: Merging this PR changes nothing (head tree == master tree), and the fix is already shipped in master with the author's credit retained. The correct action is to close the PR as superseded by 八问(中文)
判定:不可合(合并 = 无操作)。建议 owner 关闭本 PR,标注 superseded by |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) PR #5028 review complete — verdict: do NOT merge (no-op); close as superseded by
|
[ework] CI fix verified green ✓ |
|
Closing: the branch is fully in sync with master — the three-dot diff against origin/master is empty; the container-generation guard it carries is already on master (src/served-markers.ts:286). |
Problem
In Docker/containerized environments, processes virtually always run as PID 1. When a container restarts, its prior writer marker (/writers/.json) remains in the mounted state directory with
pid: 1. When the new container boots,findContestedWriterschecksisPidAlive(raw.pid). Because the new container itself is running as PID 1,process.kill(1, 0)succeeds.This causes the newly booted proxy to falsely conclude that another concurrently running instance committed the session within
CONTEST_WINDOW_MS, emitting false-positive[split-brain]errors.Fix
startedAt(epoch ms, derived fromprocess.uptime()) onWriterIdentityandWriterMarkerFile.findContestedWriters, if an observed marker file sharesraw.pid === process.pidbut itsstartedAt(or file mtime) precedes this process boot time, it is recognized as a predecessor from an earlier container generation and skipped.Tests
container restart with same PID (pid 1) does not falsely contest prior generationtotests/served-markers.test.ts.tests/served-markers.test.ts.