Repository navigation
ci: release Atlas consumers from unrelated shard waits - #444
Conversation
afflom
left a comment
There was a problem hiding this comment.
Source review of 63b9093 found no blocking issue. The 14 expanded shard executions preserve their names, environment, steps and artifacts. The committed plan's prerequisites are all reachable, the join still requires every shard including census-c/d, and failure propagation, honesty, same-run artifact custody and release-evidence behavior remain intact. RP-14's dynamic source assertions and RP-18's release gate do not depend on the old matrix grouping. Local syntax/equivalence checks and missing-prerequisite/incomplete-join negative probes pass.
The initial draft's honesty and exact-head reproducibility passed. After source review, the PR was marked ready to start the full acceptance run: https://github.com/UOR-Foundation/LexLean/actions/runs/42647185072 . Concurrency cancelled the superseded reduced draft run. This review does not substitute for full hosted acceptance, which remains pending; no merge or measured overall speedup is claimed.
|
The complete ready-event run 42647185072 is terminal: 44 jobs passed, and Intel macOS job129265555619 was cancelled. All Atlas shards and the final join, package acceptance, same-run honesty and exact-head reproducibility passed, but the cancelled host job means this PR has not met full acceptance and cannot merge. Fresh head63b9093/base4dc and the matching head/tested merge tree07a566 remain unchanged. The Intel job ran15:45:00–19:45:27 UTC against a240-minute workflow timeout; its terminal log says only “The operation was canceled.” That timing is consistent with the limit but does not establish the cause. The conformance run was incomplete; no failing-test result was observed. One focused terminal log and the full run/job receipt are preserved. No rerun, cancellation, source change or merge has been initiated; the source review remains scoped to the scheduling change. |
|
Release qualification is now consolidated in draft #450 at 25be1d8. It retains this PR's reviewed commit as a parent, carries the Atlas and BS-07 changes together with the Intel construction/partition fix, and targets release/v0.4. Keep this PR open and its history intact until the exact combined candidate passes full acceptance and is merged; do not launch another independent full gate here. #450 currently has source review and focused local tests only, with real Rust/Intel qualification still pending. |
|
The combined candidate #450 has now landed in release/v0.4 as cf76090 after all 50 exact full-gate jobs and exact-head reproducibility passed. Fresh readback confirms actual tree 0c24b969dcc5c02d9afb7616f97736827b7b09e6 equals the reviewed and fully tested tree. This preserves both source histories; GitHub now reports this original carrier merged. No separate rerun is needed. Acceptance: #450 (review) |
Change
Split the first Atlas matrix by the prerequisites already declared in
.github/verify-shards.json:cert4runs in its own job, socert4-replaystarts as soon as its input is ready.census-aandcensus-bremain one matrix.closurewaits for them andcert4.census-candcensus-drun in a separate matrix and still block the final join.There are still 14 expanded Atlas jobs, with identical check names, commands, environments, artifact names and retention settings. Every shard still runs once. The complete join, honesty dependency, clean-checkout policy, all acceptance tests and release-evidence checks stay intact. No semantic identity or generated record changes.
Reason
In successful run 42596285175,
cert4finished at 10:24:53 UTC andcensus-a/bby 10:31:48. Both consumers waited until 10:41:54 because the shared matrix also containedcensus-c. That is approximately 17 minutes of avoidable replay waiting and 10 minutes of closure waiting. These are observed dependency delays, not measured savings from this patch. Intel macOS dominated that run's total time; no overall CI speedup is claimed.Verification
Passed locally against base
4dc5b803c5b0771f9f19563373515d0a2c2f313a:The existing RP-14 and RP-18 implementation was reviewed; neither needs a source change. Cargo, Lean and
just vvwere not run in this workspace. Hosted CI remains required. This starts as a draft, whose reduced CI is a development signal only; full acceptance is still required before merge.Ownership
This branch owns only the Atlas scheduling change. The current open PR inventory has no overlapping scheduling carrier. The v0.5 integration and regeneration lane remains #405's.