model-writtenrun 7374 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro reported that the index-only-scan comment incorrectly claimed tuple predicate locks required xmin, and Heikki Linnakangas explained that updated tuples would prevent the visibility-map fast path, making tuple locking plausible #1 #2.
- Thomas separated the immediate comment correction from the behavioral change, which needed a TID-based locking interface, and posted the comment-only patch #3 #4.
- Ashwin Agrawal reviewed the comment correction and discussed unnecessary conflicts from index-page locking; Thomas committed the patch unchanged as 74b7cc8c02 and confirmed the push #5 #6.
- Thomas identified an existing README-SSI proposal to avoid index predicate locks for successful equality searches covering an entire unique key #7.
- After nearly two years, Thomas posted a separate behavioral patch using tuple-level SIREAD locks for index-only scans, with isolation tests demonstrating both avoided false conflicts and preserved true-conflict detection #8.
- Thomas then retained that patch and added an experimental B-tree optimization that deferred index-page locks until the executor knew whether a unique-key search found a visible tuple, while leaving planner integration and split handling open for further work #9.
- David Rowley's reviews led Thomas to replace attribute sorting with a Bitmapset and exclude deferred unique constraints; Thomas demonstrated a phantom-search hazard addressed by the latter check, while remaining uncertain whether further restrictions on self-written tuples were needed #10 #11 #12 #13.
- Daniel Gustafsson later reported that the series no longer applied and requested a rebase, then marked it Returned with Feedback; the comment fix had landed, while both behavioral patches remained uncommitted in the supplied thread #14 #15.