Skip to content

[PR] docs(ci): state the measured cron backstop window, not a 30-minute one (#5729) - #5820

Closed
suyeku1 wants to merge 1 commit into
ranxianglei:masterfrom
suyeku1:2026-11-06_ci-cron-window-docs
Closed

suyeku1 wants to merge 1 commit into
ranxianglei:masterfrom
suyeku1:2026-11-06_ci-cron-window-docs

Conversation

@suyeku1

@suyeku1 suyeku1 commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

Part of #5729 (follow-up to the docs-only #5803 that just landed).

Comments-only correction to .github/workflows/ci.yml. #5803 wrote into AGENTS.md that a PR stacked on a feature base gets no core matrix, and that the gap is covered because the :03/:33 cron 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)

spec cron: '3/30 * * * *' = 48 ticks/day
delivered 107 schedule runs over 21.30 days = 5.02/day (10.5%)
inter-run gaps (n=106) min 2.16h, p50 4.77h, p90 7.35h, max 10.61h — none at or under 2h
master churn, same window 2,735 commits = ~128/day
exact-sha coverage 72 / 2,735 (2.6%) got any cron run on their own sha; 33 / 2,735 (1.2%) got a green one
outcome mix 46 success / 54 failure / 7 cancelled (failures mostly real, e.g. run 41986289521 red on 14 legs)

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 own schedule, 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-yaml parses the edited file; re-reading it back gives byte-identical on.schedule, on.push.branches, on.pull_request.branches, on.pull_request.types, and the same 16 jobs.
  • Statically: comments cannot alter workflow behavior, and pull_request.branches is 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.

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 ranxianglei changed the title docs(ci): state the measured cron backstop window, not a 30-minute one (#5729) [PR] docs(ci): state the measured cron backstop window, not a 30-minute one (#5729) Nov 6, 2026

@ranxianglei ranxianglei left a comment

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(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):

  1. 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.branches unchanged so this PR itself got the core matrix.
  2. File facts. schedule: cron '3/30 * * * *' = :03/:33 = 48 ticks/day — matches the new "nominally 48/day" wording.
  3. 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.
  4. 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 (all N/30 crons) and makes no coverage-window claim — out of scope, deliberately left alone.
  5. House conventions. docs(ci): commit type, references #5729 in the message, fresh base, one clean file, no version bump, no kernel/.
  6. 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):

  1. 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。
  2. 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」。
  3. 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)。
  4. 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。
  5. 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。
  6. 是否破坏缓存:不涉及(无代码路径)。
  7. 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)。
  8. 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)。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

(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.

@suyeku1

suyeku1 commented Nov 6, 2026

Copy link
Copy Markdown
Contributor Author

(ai: deepseek/deepseek-v4.1-flash)

Closing in favour of the maintainer's #5821, which covers the identical ground — same base c3dfbe6b21a1, same two comment blocks, and the two head trees differ in exactly one path (.github/workflows/ci.yml).

Rationale, so the closure is not mistaken for a withdrawal of the finding:

  • The finding stands. The cron delivers ≈5 runs/day against a 48-tick spec (107 runs / 21.3 days), median gap ≈4.8h, worst ≈10.6h, and only 2.6% of master commits ever drew a cron run on their own sha (1.2% a green one) — measured against this repo's own Actions history and recorded in #5729.
  • Only the wording was duplicated. Two PRs racing on the same comment blocks would spend the owner's merge decision twice for one outcome.
  • This PR's CI was fully green (47 checks: 35 success / 12 skipped-by-default / 0 failed, all 11 test (...) matrix legs passing), so the closure leaves nothing unverified on the content — it is a duplicate-removal, not a defect withdrawal.

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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants