ci(ci): drop the classifier job whose only remaining reader was the arm gate #2408

Merged
jaylouisw merged 1 commit from ci/1121-1027 into main 2026-11-21 10:44:19 +00:00
Owner

The Classify the change job is gone. Its three outputs are answered where they are spent, and arm - the last reader - asks the liveness question itself.

Why it went

docker already classified its own diff (same scripts/classify-change.js, same arguments, one shallow fetch of the base commit, data_only=true stands its own steps down), and the three test legs read neither data_only nor code ("three versions, always" is on the matrix). By 2026-11-21 the only consumer left of the whole job was arm's live condition. So it bought one API call per run and cost a runner slot on every run: this runner takes four jobs at once, and ci.yml asked for five, so the fifth - always a test leg, i.e. the wall clock - waited for the classifier.

Baseline (before) - leg logs, Forgejo run indices

run event wall 3rd leg starts last leg ends suite line
9258 PR #2407 70.6 s +9.5 s +65.0 s 445 / 15024
9254 PR #2405 68.1 s +11.4 s +62.7 s 444 / 14983
9238 PR #2404 70.9 s +13.6 s +66.1 s 444 / 14983
9235 PR #2403 74.5 s +8.6 s +69.5 s 444 / 14983
9214 PR 217.1 s +14.0 s +67.9 s 444 / 14983
9250 push 75.9 s +22.1 s +72.2 s 444 / 14983
9246 push 107.9 s +15.7 s +69.6 s 444 / 14983
9230 push 66.2 s +4.1 s +62.0 s 444 / 14983
9225 push 89.9 s +27.9 s +84.0 s 444 / 14983

Pull-request runs: median wall 70.9 s, third-leg start +9.5 .. +14.0 s, last leg ends ~+66 s, against a ~45 s suite.

After

Run 9262 (this PR, ci.yml for #2408), offsets from the first job's received task line:

job start end duration
Server smoke and Docker build +0.0 s +25.5 s 25.5 s
Test (Node 18) +0.0 s +58.4 s 58.3 s
Test (Node 20) +0.0 s +57.3 s 57.3 s
Test (Node 22) +0.0 s +58.2 s 58.2 s
Arm auto-merge +60.3 s +62.2 s 1.9 s

wall 70.9 s (PR median before) -> 62.2 s, -8.7 s / -12%. Third-leg start +9.5 s -> +0.0 s. All four jobs start within 0.1 s; the legs themselves are unchanged (57-58 s here, on a host that was ~10% slower than the baseline runs: 50-52 s suites against 44-46 s).

Post-merge push run 9277 (main b0642b27, 446 suites / 15125 assertions): wall 74.3 s, docker +0.0 -> +18.3 s, legs +1.9 / +12.6 / +13.9 -> +58.6 / +68.7 / +69.0 s. The legs there queued behind the merge burst (close, audit, rebase, main-health all dispatch on the same push), so that run is not a clean before/after - the classifier is gone from it, but it was never what the legs were waiting for in that burst.

Coverage is unchanged

  • 444 suite(s) ... 14983 assertions passed, 0 failed in every leg of 9262 - the same counts main's runs 9254/9250 report (444 / 14983); 446 / 15125 after the two device/test merges in 9277.
  • Three Node versions, the HTTP smoke test, the image build, the Dockerfile checks and the prune all still run.
  • The classification rule, the script, the arguments and the failure direction are untouched (anything unreadable is a code change, so the full matrix runs).
  • arm still requires all three CI / Test (Node *) statuses present and every status on the head green; the liveness question it now asks itself uses the legs' rule and direction (only a positive merged:true / state:closed stands it down). 9262 armed this PR: #2408: armed for a squash merge of 333e7c7d ....
  • The only status that disappears is Classify the change, which gated nothing: branch protection on main has enable_status_check: false.

Workflow-only change: .forgejo/workflows/ci.yml (+52 / -250).

The `Classify the change` job is gone. Its three outputs are answered where they are spent, and `arm` - the last reader - asks the liveness question itself. ## Why it went `docker` already classified its own diff (same `scripts/classify-change.js`, same arguments, one shallow fetch of the base commit, `data_only=true` stands its own steps down), and the three test legs read neither `data_only` nor `code` ("three versions, always" is on the matrix). By 2026-11-21 the only consumer left of the whole job was `arm`'s `live` condition. So it bought one API call per run and cost a runner slot on every run: this runner takes four jobs at once, and ci.yml asked for five, so the fifth - always a test leg, i.e. the wall clock - waited for the classifier. ## Baseline (before) - leg logs, Forgejo run indices | run | event | wall | 3rd leg starts | last leg ends | suite line | |---|---|---|---|---|---| | 9258 | PR #2407 | 70.6 s | +9.5 s | +65.0 s | 445 / 15024 | | 9254 | PR #2405 | 68.1 s | +11.4 s | +62.7 s | 444 / 14983 | | 9238 | PR #2404 | 70.9 s | +13.6 s | +66.1 s | 444 / 14983 | | 9235 | PR #2403 | 74.5 s | +8.6 s | +69.5 s | 444 / 14983 | | 9214 | PR | 217.1 s | +14.0 s | +67.9 s | 444 / 14983 | | 9250 | push | 75.9 s | +22.1 s | +72.2 s | 444 / 14983 | | 9246 | push | 107.9 s | +15.7 s | +69.6 s | 444 / 14983 | | 9230 | push | 66.2 s | +4.1 s | +62.0 s | 444 / 14983 | | 9225 | push | 89.9 s | +27.9 s | +84.0 s | 444 / 14983 | Pull-request runs: median wall **70.9 s**, third-leg start **+9.5 .. +14.0 s**, last leg ends ~+66 s, against a ~45 s suite. ## After Run **9262** (this PR, ci.yml for #2408), offsets from the first job's `received task` line: | job | start | end | duration | |---|---|---|---| | Server smoke and Docker build | +0.0 s | +25.5 s | 25.5 s | | Test (Node 18) | +0.0 s | +58.4 s | 58.3 s | | Test (Node 20) | +0.0 s | +57.3 s | 57.3 s | | Test (Node 22) | +0.0 s | +58.2 s | 58.2 s | | Arm auto-merge | +60.3 s | +62.2 s | 1.9 s | **wall 70.9 s (PR median before) -> 62.2 s, -8.7 s / -12%. Third-leg start +9.5 s -> +0.0 s.** All four jobs start within 0.1 s; the legs themselves are unchanged (57-58 s here, on a host that was ~10% slower than the baseline runs: 50-52 s suites against 44-46 s). Post-merge push run **9277** (main `b0642b27`, 446 suites / 15125 assertions): wall 74.3 s, docker +0.0 -> +18.3 s, legs +1.9 / +12.6 / +13.9 -> +58.6 / +68.7 / +69.0 s. The legs there queued behind the merge burst (`close`, `audit`, `rebase`, `main-health` all dispatch on the same push), so that run is not a clean before/after - the classifier is gone from it, but it was never what the legs were waiting for in that burst. ## Coverage is unchanged - `444 suite(s) ... 14983 assertions passed, 0 failed` in every leg of 9262 - the same counts main's runs 9254/9250 report (444 / 14983); 446 / 15125 after the two device/test merges in 9277. - Three Node versions, the HTTP smoke test, the image build, the Dockerfile checks and the prune all still run. - The classification rule, the script, the arguments and the failure direction are untouched (anything unreadable is a code change, so the full matrix runs). - `arm` still requires all three `CI / Test (Node *)` statuses present and every status on the head green; the liveness question it now asks itself uses the legs' rule and direction (only a positive `merged:true` / `state:closed` stands it down). 9262 armed this PR: `#2408: armed for a squash merge of 333e7c7d ...`. - The only status that disappears is `Classify the change`, which gated nothing: branch protection on `main` has `enable_status_check: false`. Workflow-only change: `.forgejo/workflows/ci.yml` (+52 / -250).
ci(ci): drop the classifier job whose only remaining reader was the arm gate
All checks were successful
CI / Server smoke and Docker build (pull_request) Successful in 29s
PR rebase / Rebase onto main and arm auto-merge (pull_request) Successful in 6s
CI / Test (Node 20) (pull_request) Successful in 59s
CI / Test (Node 22) (pull_request) Successful in 59s
CI / Test (Node 18) (pull_request) Successful in 1m0s
CI / Arm auto-merge (pull_request) Successful in 3s
333e7c7d76
`Classify the change` published `data_only`, `code` and `live`.  Every one of
those answers is now asked by the job that spends it: `docker` classifies its
own diff with the same script and the same arguments and stands its own steps
down on `data_only=true`, the three test legs never read either output, and
`live` is asked by the legs and by `docker` in their first step.  That left
`arm` as the only reader, and it now asks the same question itself before it
reads the statuses - same rule, same direction (only a positive `merged:true`
or `state:closed` stands the job down).

So the job bought one API call per run and cost a runner slot on every run.
This runner takes four jobs at once and ci.yml asked for five, so the fifth -
always a test leg, which is the wall clock of the run - waited for the
classifier to finish.

Measured through the leg logs of nine consecutive runs (PRs 9258, 9254, 9238,
9235, 9214; pushes 9250, 9246, 9230, 9225): the third leg started a median
9.9 s after the first job and the last leg ended +67.9 s in, against a ~52 s
suite, for a median 74.5 s run.  No suite, assertion, runtime, smoke test,
image build or status in the arm gate changes.
forgejo-actions scheduled this pull request to auto merge when all checks succeed 2026-11-21 10:39:17 +00:00
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
jaylouisw/packetforge!2408
No description provided.