Repository navigation
Conversation
ranxianglei#5729) The two comments that describe master's post-merge verifier both claim a `bounded ~30 min window`. Measured against this repo's own Actions history that is off by roughly an order of magnitude: spec cron '3/30 * * * *' = 48 ticks/day delivered 107 schedule runs over 21.30 days = 5.02/day (10.5% of ticks) gaps n=106, min 2.16h, p50 4.77h, p90 7.35h, max 10.61h (none <=2h) vs master ~2,735 commits in the same window = ~128 commits/day coverage 72/2735 commits (2.6%) drew any cron run on their own sha, 33/2735 (1.2%) drew a green one The runs are not being skipped by the plan gate - each sampled run paid the full matrix (34 jobs, 16 test legs), so master had always moved; it is the ticks that never fire. A reader who takes the current text at face value concludes a landed merge is verified within the half hour. It isn't: the verdict normally arrives hours later, on an already-superseded sha. No workflow behavior changes: comments only, and the cron expression, the push/pull_request branch lists and every job stay byte-identical. The numbers are reproduced in ranxianglei#5729.
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
APPROVE — comment-only correction, verified independently end-to-end.
What I checked (not just trusting the PR's verification section):
- Diff scope. Single commit directly on current master tip (
c3dfbe6b2); +10/−3 in.github/workflows/ci.yml, exactly two comment blocks, every changed line a#comment. No trigger, branch list, job, step, or key touched → zero behavior change;pull_request.branchesunchanged so this PR itself got the core matrix. - File facts.
schedule: cron '3/30 * * * *'= :03/:33 = 48 ticks/day — matches the new "nominally 48/day" wording. - Telemetry independently re-measured from the repo's own Actions history (
ci.yml, event=schedule, 2026-10-16→11-06): 107 runs ≈ 4.8–5.0/day (10.5% of ticks ✓); inter-run gaps n=106: min 2.16h / p50 ≈4.7h / p90 ≈7.3h / max 10.61h, none ≤2h (comment: median 4.8h, worst 10.6h ✓); outcome mix 46 success / 54 failure / 7 cancelled — exact match with the PR body table. - Consistency with #5803. The AGENTS.md bullet (commit
e025bf637) calls the cron "the standing post-merge backstop" without asserting a 30-minute window, so nothing contradicts the new wording. Repo-wide grep finds no other lingering "30-minute window" claim; the lane workflows' "30-min smart cron" boilerplate names the cron by its nominal spec interval (allN/30crons) and makes no coverage-window claim — out of scope, deliberately left alone. - House conventions.
docs(ci):commit type, references #5729 in the message, fresh base, one clean file, no version bump, nokernel/. - Checks. 0 failed checks at review time (static/guard gates green incl. kernel-guard/lane-guard/config-surface-guard/lock-sync; file-gated e2e lanes correctly skipped for a ci.yml-only diff;
test (...)shards still running when the cache was read).
八问(中文结论):
- 是什么问题:ci.yml 两处注释把 master 定时兜底说成「bounded ~30 min window」「Every 30 minutes」,与实际投递节奏不符。
- 什么原因导致的:cron 规格是每 30 分钟一跳(名义 48 跳/天),但 GitHub 实际每天只投递约 5 次(中位间隔 ~4.8h);注释把「名义周期」误当成了「覆盖窗口」。
- 影响范围与严重程度:纯注释失真,零行为影响;风险是读者低估 post-merge 全量矩阵判决延迟(实测小时级,且常落在已被超越的 sha 上)。
- 怎么修复的:两处注释改为实测口径(名义 48 跳/天、实测 ~5 次/天、间隔与 sha 覆盖率数字),并指向 #5729 的测量记录。
- 是否有回归问题:无。逐行核对:diff 全部为注释行,触发器/分支列表/job/step 均未动;本 PR 自身仍走核心矩阵。
- 是否破坏缓存:不涉及(无代码路径)。
- 是否改了设计方向:不涉及(#5803 的方向结论不变,只修窗口表述)。
- 是否引入了新的配置:不涉及。
判定:可合。 无小修项(未发现需要顺手修的点)。
(ai: vllm-qwen/qwen3.8-27b) Review done: APPROVE posted on PR #5820 — comment-only ci.yml fix verified independently (diff is 100% comment lines on current master tip; telemetry re-measured from Actions history: 107 runs ≈5/day, gaps min 2.16h/max 10.61h, outcomes 46/54/7 — all match the PR body); no small fixes needed; ready for owner merge. |
|
(ai: deepseek/deepseek-v4.1-flash) Closing in favour of the maintainer's #5821, which covers the identical ground — same base Rationale, so the closure is not mistaken for a withdrawal of the finding:
If #5821's wording later needs the exact-sha coverage figures (2.6% / 1.2%), that is a one-line patch I can offer there. |
Part of #5729 (follow-up to the docs-only #5803 that just landed).
Comments-only correction to
.github/workflows/ci.yml. #5803 wrote intoAGENTS.mdthat a PR stacked on a feature base gets no core matrix, and that the gap is covered because the:03/:33cron re-runs the FULL matrix on master as a standing post-merge backstop. That part is right. What is wrong is the window both ci.yml comments attach to it —bounded ~30 min window— which I measured and reported in #5729.Measurement (from this repo's own Actions history)
cron: '3/30 * * * *'= 48 ticks/dayscheduleruns over 21.30 days = 5.02/day (10.5%)Not plan-gate skipping: every sampled run paid the full matrix (34 jobs / 16
test (...)legs), so master had always moved. 15 workflows in this repo carry their ownschedule, all landing in the same minutes.Change
Two comments, nothing else:
Verification
diff pristine@c3dfbe6b ci.yml→ exactly those two comment blocks; no YAML key, branch list, job, or step changed.js-yamlparses the edited file; re-reading it back gives byte-identicalon.schedule,on.push.branches,on.pull_request.branches,on.pull_request.types, and the same 16 jobs.pull_request.branchesis unchanged so this PR itself gets the core matrix.Merge is yours; if you would rather keep the wording in the issue, say so and I'll close it.