Skip to content

Record the parallel failure-order marker by default (phase 2 of #307) #351

Open

Description

@rmcdaniel

Follow-up to #307 / #348 (two-phase rollout).

#348 ships the parallel failure-order marker (durable_workflow.parallel_failure_order) as read always, write opt-in:

  • Worker(record_parallel_failure_order=True) or DURABLE_WORKFLOW_RECORD_PARALLEL_FAILURE_ORDER=true turns writing on.
  • The default is off, because published 2.20.0 and earlier cannot replay a marked run. The Server marks the workflow task replay-blocked and the run waits for repair.

Required outcome (next minor after the release that includes #348): writing defaults to on, so new runs raise an all() group's first recorded member failure by default.

Acceptance

  • The default flips to on. false still turns it off.
  • The CHANGELOG states:
  • The default-path fixtures (failure-order-default-first-task-sends-released-commands, parallel-group-metadata-emission and the completion-limit fixtures) are superseded properly at the next protocol version, or kept with recording explicitly set to off. They are not rewritten.
  • Installed acceptance: a marked run on a Server built from main with a worker of the new default, and a mixed fleet in which every worker runs the Opt-in first-recorded-failure ordering for all() with a per-run marker (#307) #348 release or later.
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

    enhancementNew feature or requestpriority:P2Normal-priority product work

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions