Repository navigation
ci: perf-ab lock-up detector, run budget, and a 45% faster latency replay - #621
Merged
Merged
Conversation
…'s drain window
Run 39897074404 was cancelled by the 90-min job timeout one rep short of
max (complete=false: no verdict) though nothing hung: 858 children,
5.4-9.8 s each. Every row stayed inconclusive, so all 12 extended to 25
reps: 12 x 3 arms x 25 x ~6.3 s > 5400 s.
- Lock-up detector: each measured child gets PERF_AB_CHILD_TIMEOUT_S
(180 s, ~18x the slowest observed). On expiry gdb dumps every thread
(the child opts in with PR_SET_PTRACER; /proc fallback), the child is
killed, and the run fails with the stacks in the artifact and summary.
- --budget-minutes (workflow: 80): adaptive extension starts another step
only if its projected time fits, so a run ends complete and judgeable.
- The latency replay emits only p99, sampled while workers run, yet slept
through twice the release bound (2.6 s) per child. Skip it: 6.3 s ->
3.4 s per measurement on large-class/8. latency_protocol -> v2 so old and
new artifacts never pool.
Local-Gate: v1 tree=ca9f9717a102b8ed346d7b2718dee3bf929d1ddd secs=147 lanes=lint:run
Ci-Attestation: {"at":1792735169,"gate":"python/all/lint","host":"linux-x86_64","key":"c8ad26cfa5f66ca93c0eb7c02afd7c96b51050de8b08d3f089070facf4a24f14","lane":"lint","parents":["5a5e97e16cf872bb5d66220bbb8827db3e1822dc"],"secs":147,"stamp":"f40c2c83d3c29969d6c2e920a1188f70","tree":"ca9f9717a102b8ed346d7b2718dee3bf929d1ddd","v":1,"via":"run"}
This was referenced Oct 23, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why #616's V7 campaign could not reach a verdict, and the fix. Related: #614, #616, #620.
Diagnosis: not a lock-up, a time budget
Run 39897074404 was cancelled by the 90-min job timeout. Its checkpoint was
complete: false, so no verdict could be drawn. Nothing hung: its 858 measured children took 5.4–9.8 s each, with < 0.1 s spread within a row. Every row stayed INCONCLUSIVE, so the adaptive loop took all 12 rows to 25 reps: 12 × 3 arms × 25 × ~6.3 s ≈ 5,700 s > 5,400 s. The build takes < 1 min; the rest is measurement.About 5.2 s of each ~6.3 s measurement was sleep. Each child samples RSS for 2 × the 1,300 ms release bound, including the latency replay, which reports only p99, recorded while workers run.
Changes (ci/ only, no allocator code)
PERF_AB_CHILD_TIMEOUT_S(180 s, ~18× the slowest observed). On expiry gdb dumps every thread's backtrace; the child opts in withPR_SET_PTRACER, because Yama blocks a sibling from attaching, and there is a/procfallback. The child is then killed, and the run fails with the stacks in the artifactfailurefield and the step summary. Previously a hang silently used up the 90 minutes and left nothing.--budget-minutes(workflow passes 80). Adaptive extension adds a 5-rep step only if its projected time (5 × slowest observed rep per pending row) fits, so runs end complete and judgeable with honest INCONCLUSIVE labels. Rows the budget stops are recorded asstopped_by_budget.latency_protocolis bumped to v2, so v1 and v2 artifacts are never pooled. A full 12 × 25 campaign is now ~53 min.Tests
4 new tests (budget, child output/failure, lock-up → stack dump + kill), plus the existing perf-ab suites: 95 pass. Verified locally that the gdb dump attaches and prints
#0…frames.Not done: sharding rows across parallel runners. Each row is judged on its own, so it is possible, but the dispatcher's artifact pooling would need changes. With these fixes it isn't needed to fit the budget.
🤖 Generated with Claude Code