Repository navigation
ci: Atlas shards drop their own part from an earlier attempt - #560
Conversation
"Re-run all jobs" downloads atlas-* including the part the shard itself uploaded in an earlier attempt; the shard then refuses the non-empty output directory (LLI9001: census-d 133274946091, cert4-replay 133303022254 on run 43799050127). Remove that directory after placing the parts; RP-14 pins the step in every downloading job. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013Asa3QVAsEzzaQPMzJnNz1
afflom
left a comment
There was a problem hiding this comment.
Bounded source review at47f3b1e found no concrete introduced defect. The seven added cases remove only the current shard's output directory; the verifier's empty-output check remains. Fixed shard names and RP-14's dependency-order checks preserve required imports for the stated all-jobs rerun. Join is unaffected.
The new assertion pins source text; it does not demonstrate actual rerun artifact selection. Please retain complete single-attempt behavioral evidence before claiming recovered same-attempt execution. Failed-jobs-only reruns remain a separate scope. Current CI43870100877 and repro43870100817 are incomplete; CI includes three Intel skips. No full acceptance or retry authorization is implied. Receipt /workspace/work/pr560-47f3-source-review-20261116.json, SHA25632f20d198e81c7317534b65dd3b6f9d651a3ea426096002013d6d71c58915afd. Existing CI owner retained; no competing patch or repeated tests.
# Conflicts: # CHANGELOG.md
afflom
left a comment
There was a problem hiding this comment.
The merge with release 2e1ff2f preserves the reviewed fix: e9075c2 has ordered parents 47f3b1e and 2e1ff2f, tree 46604d6d. Against the current release, only the original seven own-shard removals, RP-14 assertion and four-line changelog note remain. The workflow blob is unchanged, the RP-14 additive patch is identical, and the arrays changelog entry is retained. No new source defect was found in this bounded merge review.
The behavioral limitation in review5920283939 remains: the string assertion and a fresh successful run do not establish recovery across attempts. Existing CI owner retains that evidence and final-head/actual-tree qualification. Exact repro43897548707 passed; CI43897548879 remains incomplete. No retry or repeated gate was started.
Receipt pr560-e907-delta-review-20261116.json; SHA256 6221ffedae023386bde1c95139517820940fa86873d2111bec2c53875e509539.
afflom
left a comment
There was a problem hiding this comment.
Actual landing verified: merged0c6eeb9a has ordered parents2e1ff2f5 and e9075c2. Its tree46604d6d equals the reviewed head and checked synthetic merge98b187d6. Final CI43897548879 attempt1 completed71 successful jobs/831 successful steps plus3 Intel job skips. Actual repro43920080311 passed; actualCI43920080624 remains pending.
This credits tree preservation and the completed fresh-run scope. The PR body's all-jobs recovery claim still needs evidence of actual rerun/artifact selection; fresh attempt1 and the RP-14 source assertion do not establish it. The Intel skips remain explicit.
Forward43920080440 failed again at17:38:47Z with index.lock present and exit128; regeneration was skipped. New log133484694086 is preserved once (SHA25661ae7276…). This is a recurrence of the existing forward-owner blocker, not a new proven cause or permission for lock deletion/retry. Prior source and originalR3/STOP obligations remain; no full acceptance or issue closure.
Actual review53c90f11ced991016149464dca21ca6d15b23a27411aff27a722399ccc6e6b62; forward receipt7647fedcd5c0ae48bc3bf1c237e090be11f7e8f9b43012cb0b260cb60656a7e9.
afflom
left a comment
There was a problem hiding this comment.
Terminal supplement to review5925116971: actual 0c6eeb9 push CI43920080624 attempt1 completed successfully at20:59:20Z: eight successful jobs/all63 steps, with15 skipped jobs. Intel construction, moves, print, lower, validate and x86_64-apple-darwin now complete, alongside plan/honesty. This adds actual-tree Intel evidence to the retained final-head71/831 and verified tree equality; it is not a full all-platform gate in this run. No canceled jobs or attempts were aggregated.
The all-jobs rerun recovery/artifact-selection claim remains unproved by these fresh attempt1 runs. The forward index.lock blocker and existing source/R3/STOP obligations remain. No rerun, cleanup, full acceptance or issue closure is implied. Receipt pr560-0c6-terminal-intel-20261116.json preserves the terminal inventory.
Problem. "Re-run all jobs" can't recover a gate. Each Atlas job downloads
pattern: atlas-*, which also pulls the part that job uploaded in an earlier attempt. The shard then refuses the non-empty output directory withLLI9001. On run 43799050127, attempt 3, this hit census-d (133274946091) and cert4-replay (133303022254). The census-d log shows it downloadingatlas-part-census-dfrom the previous attempt.Fix. After the parts are placed, each shard job removes
ci-parts/<own shard>. A job's own earlier output is never its input. The join produces no part and is untouched. RP-14 now pins the step in every downloading job; I checked that removing it fromatlas-census-tailfails RP-14.Not changed. Intel lanes still refuse a construction from an earlier attempt, as SPEC §8.3 requires ("no … artifact from another run or attempt is accepted"). With this fix, "Re-run all jobs" re-runs the construction as well, so it recovers both Atlas and Intel without a new dispatch.
Generated by Claude Code