Skip to content

FLOW-003 QM-09: compile-once — one --workspace test build, nextest archive, K shards fanned across all runners (FLOW-003 v2.1 §row QM-09) #4527

Closed as not planned

Description

@noahgift

Operator 2026-09-27, trunk-based target design. Order of work: measure (#4513), then build once (QM-09), then generated shared files (#4526), then tiers (QM-10/11), then retire folds once main has been green 7 days in a row. Spec: docs/specifications/FLOW-003-queue-and-release-cycle-model.md (branch flow-003-v2.1). Targets: CI p50 ≤10 min for code and ≤2 min for docs; lead time under 1 h; one build per PR. This row changes .github/workflows. The operator's order of work authorizes the change; its PR still carries a quorum and a before/after timing receipt.

Activity

  1. added this to the 0.70.1 milestone on Sep 27, 2026
  2. 22 remaining items

  3. noahgift commented on Oct 9, 2026

    @noahgift
    ContributorAuthor

    Reopened. This issue carries no agent marker, so it may have been filed by a human, and agents never close a human-filed issue. A 3-lane review quorum (3/3, planted control passed) ruled to reopen it and to move or re-home it within the epic budgets instead of parking it.

    Agent: aprender-a3

  4. removed
    parkedClosed over epic budget (APR-EPIC-001 rule 5). Reopen = pull.
    on Oct 9, 2026
  5. noahgift commented on Oct 9, 2026

    @noahgift
    ContributorAuthor

    Re-homed under ISSUE-TREE-001 (epic #4540 had no live owner). Owner: la-0.72, the live look-ahead session for 0.72.0, which matches this issue's milestone. Replaces owner:aprender-a7.

    Agent: aprender-df

  6. noahgift commented on Oct 9, 2026

    @noahgift
    ContributorAuthor

    QM-09 compile-once: paired full-tier proof and decision (Refs #4505)

    BEFORE run 37966946909 @af58dfccff (main) vs AFTER run 37966950775 @ddd8311bcc (branch la-72/4527-compile-once). Both were workflow_dispatch, tier full on all 3 shards.

    • Verdict set: identical.
      • Lib junit: 83994 / 83994, all pass, no duplicates.
      • Also identical: nextest partition summaries, the explicit list (102 ids, 664 run), cargo-test lines (6120) and x86-main guard rows (122).
      • 4 guard rows fail the same way in both dispatch runs: ratchet guards that cannot resolve a base on workflow_dispatch. Main's push run at af58dfc is green.
    • Wall time: slower.
      • Run wall: 38 → 72 min.
      • Shard + build runner-seconds: 5572 → 6994 (+25%).
      • Each shard spends 1021–1068 s fetching the 891 MiB nextest archive and saves about 620 s of compile.
      • The fetch is a single-stream GET of the artifact blob, measured at about 1.1 MiB/s per stream. 4 parallel ranges ran 3.85x faster.
    • Decision (quorum, 3/3): stop.
      • A parallel ranged fetch would only bring wall time to about even, because the build job runs ahead of the shards. It would cut runner-seconds by about 10%, which is not worth another sign-off round plus a paired run.
      • The branch stays at ddd8311 with no PR.
      • Revisiting this needs a faster artifact path, for example a runner-local cache.
  7. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    merged into #5318 (master ticket for epic #4540; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions