Repository navigation
Conversation
|
Could a maintainer rerun the failed Windows Node 22 shard 3 on 0d2ee1d? The failure is the duplicate-start log assertion in This PR only adds the post-body ownership recheck in |
(ai: vllm-qwen/qwen3.8-27b) Duplicate check: no duplicate — searched open+closed issues/PRs for these test files and the flake signature; nothing tracks them yet. Labels: I verified the claim independently before acting, rather than rerunning on request alone:
Action taken: reran the failed jobs on run 42749373752 (rerun accepted, run now queued) — my own judgment based on the evidence above that this is a single-shard timing flake unrelated to the diff, so no startup/WebSocket changes are warranted before seeing the rerun. If shard 3 reproduces the same failures on the rerun, that becomes a genuine signal and the two tests' timing robustness needs its own investigation (which would be filed separately, per the one-issue-per-mechanism rule). Merge remains human-only as always. 结构化八问(中文):
结论:重跑已触发,等结果即可;暂无需改动启动或 WebSocket 路径。 |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Outcome: rerun is green. This confirms the flake diagnosis from my analysis in the previous comment: the original failure was runner-load timing noise in Remaining: owner review and merge (human-only). If those two tests start failing repeatedly on unrelated PRs, their timing robustness deserves its own issue — nothing indicates that yet. |
|
Landed via batch PR #6439 (merged) — closing as superseded. |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Acknowledged — the keepalive fix landed via batch PR #6439, so this PR is superseded. Thread recap for the record: the only blocker was the Windows Node 22 shard-3 flake (twin-start timing assertion + three |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Status report after the session restart — nothing was lost; all prior work is intact in the thread. Live-verified current state:
Nothing outstanding on this thread from my side: no code changes were ever needed beyond the original one-line |
Problem
A real request can replace a session's keepalive snapshot while an older ping awaits its response body. The existing ownership check ran before response.text(), so a stale HTTP 429 or cache miss could subsequently disarm the new snapshot and cancel its timer.
Changes
Recheck that the registry still owns the same entry immediately after body consumption. Superseded replies cannot apply stop/success decisions to a newer snapshot. Current-entry hit, miss, error and scheduling behavior remains unchanged.
Regression evidence
At base 541cf26, deterministic deferred-body cases reproduce both failures: an old 429 and an old cache miss delete the replacement 1-hour snapshot. A stale cache-hit control passes.
The fork commit has the identical Git tree (8c085ce01475d6bf80eff1580a73889f3be95c21) as the locally verified commit d327946f61cd6f8323526f880d979316144895ef. The full remote CI verdict remains separate from these targeted local checks.
Prior work
Screened all 3,221 existing upstream PRs, including every Afloat16 PR (#6046–#6055). #3390 introduced keepalive; #5669/#5672 integrated through #5847 fix default TTL detection and timer liveness. This change addresses the separate body-completion ownership race.
Closes #6077