Repository navigation
ci: regenerate.yml measures stack bounds with no core handler - #516
Conversation
Each probe of the stack-bound bisection that runs out of stack aborts, and the runner pipes every core to apport or systemd-coredump, which copies the whole probe process; for BS-69 and BS-70 dozens of them exhausted the runner's memory (24 GB of swap) and killed the job. The measuring step now sends cores to a plain file under a zero limit, so no core is written and no handler runs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
Bounded source review finds no new implementation blocker in this six-line regenerate.yml change on actual b8b6. The stack-bound harness judges successful exit plus its work marker; it does not use core files or the core-dump status bit. Setting a file core pattern with a zero core limit changes crash handling without changing that inspected predicate, workload or bisection. This is source credit, not independent confirmation that every numeric bound is unchanged or that the older run's termination was caused by a particular memory consumer.
The separate a53a2ee finish run now provides useful scoped evidence: BS69 reports44813 native comparisons agreeing and BS70 reports4801, then both fail the normative-host pinned Lean4.32.1 prerequisite. Their stack-bound artifacts14026429747/14027955039 are retained; full-case/Lean acceptance did not complete. Existing owner should disposition those already-produced measurements against exact inputs and fix the missing prerequisite for remaining qualification, without blindly repeating native work or weakening the host assertion.
Correction to my earlier #504 comment6845958314: the failure-status concern was wrong. The logged shell is bash --noprofile --norc -e -o pipefail; set -uo pipefail leaves inherited errexit enabled. Both observed failures exit101 before trailing diagnostics. No status-propagation patch or rerun is owed; the earlier comment is corrected explicitly.
This PR's current full CI is still queued and exact repro passed. Final-head full acceptance and merged-tree equality remain required. No local sysctl, producer, test, cleanup, CI action or competing patch was run; the two previously unread finish-job logs were fetched once and retained.
afflom
left a comment
There was a problem hiding this comment.
Final-head acceptance is complete at4261c903f5a5f5c5b285b5342b4da7cd6be4ed80. Full PR run43401523632 passed all70 dynamically materialized jobs and772/772 steps, including same-run honesty, Atlas join, required conformance partitions and all hosts. Exact-head reproducibility43401523775 passed its job and7/7 steps. No canceled/manual/different-head results were combined.
The independent final audit573bdd43 reuses source review5889251388 and confirms current release baseb8b6e86a and synthetic55f0efaa ordered parents[b8b6,4261]. The reviewed head and tested synthetic merge share treed40fe2632a45f3241dfc3a86e28e213a59db8a67. No unresolved concrete source blocker was found for this six-line workflow repair. Acceptance is scoped to eliminating piped core handling while retaining the inspected stack-bound predicate; it does not independently establish the older failure cause or equality of every historical numeric measurement.
Proceeding with the authorized merge-commit delivery after final head/base readback, followed by actual merged-tree verification. No producer, local test or CI gate was repeated. No broader issue is closed by this workflow-only change; #405 retains required forward integration.
|
Delivered to release/v0.4 as15d66a764ea902686c6275543b21d28b3d03130f after final acceptance review5890861473. Actual ordered parents are b8b6e86 +4261c903f5a5f5c5b285b5342b4da7cd6be4ed80. Actual merged tree d40fe2632a45f3241dfc3a86e28e213a59db8a67 exactly equals the reviewed head and qualified synthetic merge. Full70/70 jobs772/772 steps, same-run honesty and exact reproducibility passed before merge. No checks or producers were repeated. This completes the workflow repair only; IR proof/product acceptance and broader issues remain open. Required actual-history forward integration stays with #405's existing owner. |
regenerate.yml's stack-bound step bisects each recursive package's harness. Every probe that runs out of stack aborts, and the runner pipes each core to a handler (|/usr/lib/systemd/systemd-coredump ...or apport). Each handler copies the whole probe process.For BS-69 and BS-70 (the IR track, #484 stack), dozens of these handlers piled up and killed the job. Run 0ac1b85's per-minute process log shows only apport processes at the top while swap climbed to 23.9 GB.
ulimit -c 0alone does not stop it: it was already 0, and piped handlers run regardless.Change: the measuring step first runs
sudo sysctl -w kernel.core_pattern=coreandulimit -c 0. With a file pattern and a zero limit, no core is written and no handler starts. Nothing else changes, and the recorded bounds are the same.The IR track's run a53a2ee applies the same two lines to measure BS-69 and BS-70.
Not in this PR:
ci.yml's PD-08 run also aborts probes on purpose. Changingci.ymlmeans updating RP-14's line-by-line checks, so it gets its own PR.🤖 Generated with Claude Code