ExportXMLWordPrintableJSON

    • Type: Task
    • Resolution: Unresolved
    • Priority: Major - P3
    • None
    • Affects Version/s: None
    • Component/s: None
    • None
    • Storage Execution
    • Storage Execution 2026-11-23
    • None
    • None
    • None
    • None
    • None
    • None
    • None

      Make changes to replWriterThreadCount safe while a split transaction remains prepared. This expands the ticket from test coverage alone to the ordering coordination needed for SERVER-138432 alongside live worker resizing from SERVER-134783.

      Split commit/abort dispatch uses the original split's requester ID modulo the current worker count. Subsequent CRUD uses the document hash modulo the current count. After a resize, those routes can differ: with a four-worker pool, a document with hash 7 prepares on worker 3; after growing to eight workers, its split resolves on worker 3 while subsequent CRUD routes to worker 7.

      Draining the old worker queues ensures prepare has finished before replacing workers, but does not order the resolving batch against subsequent CRUD in the new pool. Classic's batch barriers provide that ordering automatically; POA needs explicit coordination.

      Implement that coordination before deciding whether a resolving batch can pipeline. Preserve the original split sessions rather than re-splitting prepared work, and retain normal fixed-worker-count pipelining.

      Acceptance criteria

      • Cover growth and shrinkage between prepare and resolution, including a non-divisor shrink such as 4 → 3.
      • Include a split assigned to a removed worker and multiple splits resolving on one replacement worker.
      • Test commit and abort separately, followed by writes to the same documents, verifying correct final contents and continued replication progress.
      • Preserve split-session lifetime and exactly-once release; verify subsequent reuse of the logical session.
      • Use deterministic synchronization to establish prepare completion, actual resize, resolution, and ordering of subsequent work.
      • Ensure concurrent worker-count parameter changes cannot invalidate the dispatch/serialization decision.
      • Add focused unit and integration coverage for the chosen coordination policy.

            Assignee:
            Ali Mir
            Reporter:
            Ali Mir
            Votes:
            0 Vote for this issue
            Watchers:
            1 Start watching this issue

              Created:
              Updated: