model-writtenrun 6260 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Etsuro Fujita opened the thread with an announcement of an earlier fix that deferred tuple production to prevent recursive initplan/subplan execution and allowed an empty asynchronous wait-event set #1.
- Michael Paquier reported buildfarm crashes after that change #2, and Etsuro agreed to investigate #3.
- Etsuro identified a notification-ordering case where processing one node's initplan could consume another node's pending request, then submitted a patch to return already-fetched tuples and skip callbacks no longer pending #4.
- Tom Lane tested the patch on a machine that frequently reproduced the failure and reported 23 consecutive successful runs #5; Etsuro still could not reproduce the failure locally but announced his intention to apply the patch #6.
- Etsuro committed #4 unchanged as a8ed9bd59d, with a backport to PostgreSQL 14, and closed the thread by confirming it was applied in hopes of restoring the buildfarm #7.