model-writtenrun 6866 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Kyle Kingsbury reported apparent SERIALIZABLE write skew in a Jepsen workload using INSERT ... ON CONFLICT #1, and Andrey Borodin subsequently supplied a separate reproducer and tentative fix for serialization errors swallowed by ROLLBACK TO SAVEPOINT #2.
- Andrey identified that the workload's single-operation transactions were not explicitly using SERIALIZABLE #3, and Kyle confirmed that setting session-wide isolation eliminated his original anomaly #4; Andrey kept the independent savepoint issue open because rolled-back reads can still affect the enclosing transaction #5.
- Andrey then submitted a separate ON CONFLICT series that added predicate locking for conflicting rows when DO NOTHING or a rejected DO UPDATE writes nothing #6; he deferred another false negative confined to TEST_SUMMARIZE_SERIAL until production cases were addressed.
- Zsolt Parragi identified additional serialization-error paths and proposed a shared dooming helper #7; after Andrey questioned whether savepoint recovery was necessarily incorrect #8, Zsolt explained that rollback cannot undo observations #9, and Andrey accepted that reasoning and invited an expanded fix #10.
- After a gap, Zsolt supplied the expanded savepoint patch with separate macros for newly detected failures and already-doomed transactions, broader isolation tests, and documentation explaining why the top-level transaction must remain doomed #11.
- Andrey reviewed the expanded error paths, accepted conservative dooming after write failures, and reported successful isolation-suite tests on every supported branch, but requested an unrelated-row read to distinguish dooming from rediscovering the original conflict #12; Zsolt made that test change #13, and Andrey declared it ready for a committer #14.
- Dean Rasheed committed the ON CONFLICT fix from the later matching patch #1153687 in thread #253491, crediting Zsolt and Andrey: the implementation fetches the conflicting tuple with the query snapshot to close the concurrent-writer race and adds DO SELECT and injection-point coverage, with backpatching through PostgreSQL 14 fb60892f40; this corrects the previous attribution to the preliminary series #6.
- At the thread's end, Aleksander Alekseev rebased and approved the savepoint patch, suggesting a shorter helper name but leaving that choice to the committer #15; the ON CONFLICT fix had landed through thread #253491, while the savepoint fix still awaited commitment.