Skip to content

Windows claim stress test can read an incomplete claim record #943

Description

@janicduplessis

Problem

The blocking Windows unit lane can refuse a claim during the real-process stress test, even though the test expects all processes to cycle the same claim set without overlap or a refusal.

Evidence

  • PR #942 CI run 35552971373, attempt 1 failed packages/stim-cli/src/__tests__/ownership-claim.test.ts at line 569 in many processes cycling the same claim set never overlap and never surface a raw errno.
  • At iteration 15, one child returned STIM_CLAIM_REFUSED for build.lock/shared/<token>.claim because its record is missing, truncated or not valid JSON. The same test passed on this PR's Node 22 and Node 24 lanes, and the PR changes only website/docusaurus.config.ts.
  • A previous intermittent failure of this test was tracked in Windows claim stress test still refuses once in a while after #913 #925 and closed after a different mitigation; that issue did not identify this record-read reason.

Cause

Unknown. Determine whether a Windows reader can observe the claim before its record is fully published, or whether this is a transient read/delete state that should be retried while preserving the ownership safety rules.

Fix idea

Reproduce the exact refusal on windows-latest, trace the record publish/read path, and fix the race without allowing a second owner. Verify repeated real-process stress runs and the full blocking Windows unit lane.

Activity

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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions