model-writtenrun 6405 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Geoghegan opened the thread by proposing documentation explaining how autovacuum logging and heap fillfactor could help users understand and improve HOT effectiveness in PostgreSQL 14 #1.
- Andres Freund warned that bypassing index vacuuming could hurt bitmap-scan workloads and lacked an escape switch, while Michael Paquier advocated configurable thresholds; Peter accepted the need for a switch but preferred generalizing INDEX_CLEANUP to preserve flexibility for future optimizations #2 #4 #7.
- Michael submitted separate scale-factor GUCs for manual VACUUM and autovacuum, and Justin Pryzby identified an incorrect fallback variable, configuration-context concerns, and documentation errors #9 #10.
- Peter argued that exposing the threshold offered uncertain benefits and additional risk, while Masahiko Sawada supported an AUTO/ON/OFF-style switch because users would struggle to tune the proposed scale factor safely #11 #14.
- Peter replaced the GUC approach with a proof of concept making vacuum_index_cleanup accept AUTO, ON, and OFF; Masahiko requested preservation of yes/no spellings and explicit AUTO support in VACUUM commands #17 #18.
- Michael continued to request a system-wide disable switch, but Masahiko and Peter favored relation and command controls consistent with other VACUUM options, so the subsequent patches omitted a GUC #19 #20 #21 #23 #24.
- Peter's next revision distinguished explicit AUTO from an unspecified option, added vacuumdb --force-index-cleanup, documentation, and tests, and left forced heap truncation out because waiting indefinitely for an exclusive lock would be dangerous #25.
- Masahiko caught the new vacuumdb option emitting FALSE instead of TRUE and requested enum naming, help, completion, and documentation fixes; Justin supplied wording corrections, which Peter incorporated into the final posted revision #26 #27 #28.
- Peter committed 3499df0dee from #28, expanding the documentation and additionally refactoring internal truncation and failsafe state, then announced the push #30.
- After the commit, Mark Dilger requested a reproducible pathological workload to justify broader controls, and Peter emphasized the command option's usefulness for testing; the thread ended with the switch landed, no system-wide GUC, and no submitted HOT tutorial addressing the original documentation proposal #31 #32 #1.