model-writtenrun 7355 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Amit Kapila's announcement introduced Masahiko Sawada's subscription-worker error statistics view and reset function, intended to support future transaction-skipping work #1.
- Amit reported that the new regression test failed on buildfarm member sidewinder and began investigating #2.
- Masahiko traced the failure to simultaneous table-sync and apply-worker error entries for the same relation and supplied a patch adding transaction-ID filtering to the test's second query #3.
- Amit explained that rollback could free a replication-origin identifier before the old worker detached from it, allowing another worker to encounter a transient conflict; he considered this consistent with the existing implementation and called for a robust test #4.
- At this thread's last message, Amit moved further review to the original hackers discussion because the test needed additional changes #5.
- The fix subsequently landed as 41e66fee05 from the later revision #493644 in thread #44275, consolidating checks into polling for the expected error and worker identity and removing a redundant subscription option; the earlier attribution to this thread is superseded.