model-writtenrun 7381 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sergey Bernikov reported that updating a published row after deleting its subscriber-side counterpart crashed logical replication into a partitioned table; Tom Lane reproduced the repeated crashes on both development HEAD and PostgreSQL 13 #1 #2.
- Amit Langote diagnosed apply_handle_tuple_routing() continuing after failing to find the target tuple and supplied a patch that returns immediately in that case #3.
- Michael Paquier requested regression coverage, Amit questioned how to test a DEBUG1-only outcome, and Tom chose to raise log verbosity and inspect the logs because none of the missing-tuple paths had coverage #4 #5 #6.
- While adding tests, Tom discovered two more crash bugs involving partition attribute-map ownership and mapping a NULL old tuple; he committed those fixes together with Amit's patch and new tests as ab55d742eb, also distinguishing partitioned cases in the debug messages #7.
- Michael subsequently questioned reload synchronization and whole-file log matching, explaining that future tests could produce overlapping messages and false matches #8 #10.
- Tom reasoned that subsequent connections and replication activity provided sufficient reload ordering, so Michael agreed to leave reload handling unchanged unless buildfarm failures appeared #11 #12.
- Michael nevertheless posted a separate patch to scan only newly generated log content, and Tom identified that its log-position capture needed to move after a subscriber SQL command to preserve the reload-ordering argument #12 #14.
- Michael accepted that correction and committed the improvement across all three test groups, with explanatory comments, as bfd96b7a3d; his final message confirmed completion, leaving both fixes landed and no outstanding patch in the thread #15 #17.