model-writtenrun 7057 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Aug 2024 · 14 entries
- Melanie Plageman opened the thread with a reproducer and fix for vacuum errors caused when a reconnecting standby moved GlobalVisState behind the fixed OldestXmin cutoff, leaving dead tuples for freezing; older branches could instead loop indefinitely #1.
- Peter Geoghegan supported enforcing the OldestXmin pruning floor, while Heikki Linnakangas asked that the initially experimental reproducer become a permanent regression test #2 #4.
- Robert Haas and Noah Misch questioned using separate pruning and freezing horizons, but Andres Freund explained that preventing backward movement could cause persistent standby conflicts and that global visibility objects could not cleanly hold a vacuum-specific cutoff; Robert consequently accepted Melanie's local safeguard #8 #20 #30 #32 #33.
- Alena Rybakina's proposed additional multixact condition was dropped after she traced the apparent performance problem to her laptop #3 #5 #27.
- Melanie made the test smaller and more deterministic by removing the standby cursor and waiting for a buffer-pin conflict, then added an explicit intermediate index-vacuum check to prevent false successes #34 #35.
- Peter verified that the test failed without the fix and passed with it, requested clearer comments, and recommended committing the test separately so buildfarm instability could be handled independently; Melanie incorporated that feedback #36 #37.
- A subsequent CI failure showed that the dataset did not always force two index-vacuum passes, prompting discussion of additional rows and VERBOSE diagnostics #38 #39 #40.
- The production fix landed unchanged from its member of #37 as 83c39a1f7f, while its test landed separately as aa607980ae with platform-dependent row counts and additional logging.
- Tom Lane reported buildfarm timeouts and a 908-second successful run with an enlarged timeout, starting debate over opt-in testing, slow-machine exclusions, injection points, and reducing the test's memory requirement #41 #42 #48 #52 #54.
- Melanie found both slow standby catch-up on 32-bit machines and failure to reach an intermediate index vacuum on a 64-bit machine, so she reverted the test on master and PostgreSQL 17 as 1a3e90948b, retaining the production fix and older-branch tests #58 #60.
- Masahiko Sawada proposed generating dead tuples with DELETE to reduce WAL, then diagnosed the unexplained 64-bit failure as dependence on assertion-build memory overhead rather than machine speed #64 #68.
- John Naylor supplied a separate experimental patch lowering maintenance_work_mem to 64kB and creating radix-tree slab contexts lazily; Melanie tested it successfully with 9000 rows and reported a substantial runtime reduction #70 #72.
- Melanie adopted Masahiko's DELETE suggestion for the surviving older-branch test, producing 571e0ee40e, and corrected an independently discovered PostgreSQL 16 Meson registration omission in 9d198f4d3e #71.
- John discovered that slab-context creation already avoided allocating backing blocks, dropped the unnecessary radix-tree changes from #70, and landed the 64kB minimum with matching autovacuum handling and a shared-memory caution as bbf668d66f #78 #79.
Jun 2025 · 4 entries
- After approximately ten months without restored coverage on newer branches, John reopened the issue and Melanie revived the test using DELETE, the lower memory minimum, and 9000 rows #81 #82 #84.
- John's cross-platform testing helped expose a new stability problem: at smaller row counts, the cursor could choose an index-only scan and pin an index page, preventing index vacuum from finishing; Melanie forced sequential scans and disabled concurrent I/O, read combining, and parallel vacuum #85 #87 #88.
- John recommended using 2000 rows everywhere and simplifying the comments, Melanie incorporated those changes, and John approved the final patch #89 #90 #91.
- Melanie reported the restored test pushed to master and PostgreSQL 17, with the supplied restoration commit matching #90 exactly; the thread ends awaiting buildfarm results rather than further patch review 303ba0573c #92.