model-writtenrun 5696 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2021 - Mar 2022 · 14 entries
- John Naylor opened the discussion after a single-user vacuum took about three days, proposing to skip tables that did not need emergency freezing so the database could return online sooner #1.
- Andres Freund objected to silently changing all single-user vacuums because they also serve space reclamation and repair, and advocated explicit table selection and recovery without restarting the server #5 #12.
- Peter Geoghegan supplied a minimal patch forcing failsafe behavior in single-user mode, but Andres pointed out that it would also affect initdb and emergencies unrelated to wraparound, steering development toward an explicit command #15 #16.
- John's replacement prototype introduced VACUUM LIMIT to select old relations and skip nonessential maintenance; Peter proposed fixed one-billion-XID/MXID cutoffs to avoid depending on potentially unsuitable autovacuum settings, and John accepted that direction #27 #28 #31.
- Masahiko Sawada and Nathan Bossart argued for reusable age selectors and configurability, while John prioritized a simple emergency command; Michael Paquier's objection to new grammar led John to implement a parenthesized VACUUM option #30 #34 #39 #42 #44.
- Justin Pryzby corrected a bitmask bug and requested explicit handling of relation lists; John then named the option EMERGENCY, rejected relation lists, added documentation and error hints, and put age cutoffs in VacuumParams #46 #47.
- After John reconsidered processing the oldest tables first, Justin supplied a complete three-patch series containing John's implementation, a compile fix, and age ordering, together with TOAST handling and regression checks #50 #51.
- Robert Haas challenged the premise that wraparound exhaustion required single-user mode, explaining that ordinary reads and VACUUM could continue unless they allocated XIDs; John subsequently confirmed that the reported failing SELECT called an extension function that performed writes #52 #54 #66.
- Andres proposed better diagnostics identifying transactions, replication slots, or databases holding back the horizon, alongside restricted emergency XID allocation, but these remained discussion proposals #62.
- John's debugger tests confirmed that ordinary VACUUM allocated an XID for truncation and that PostgreSQL 14's failsafe already disabled that operation; he consequently shifted attention toward correcting recovery advice and potentially ordering ordinary database-wide VACUUM by age #67 #69.
- Peter committed the discussion-derived explanation of why failsafe vacuum must avoid truncation as 988ffc3063, crediting John's report and announcing the comment update in #72; none of the emergency-command implementation landed in that commit.
- Masahiko then tested recovery beyond xidStopLimit and found that autovacuum continued past failures on younger tables and eventually processed the failsafe-eligible tables, strengthening the case for recovery in normal operation and clearer documentation #75 #77.
- Robert emphasized that single-user mode bypasses wraparound protection, and Noah Misch identified an additional multixact complication: VACUUM may need new multixacts unless relevant running and prepared transactions have been resolved #85 #90.
- Peter later proposed an independent final failsafe check before truncation to cover unusually lengthy index cleanup; John considered it reasonable but reported that illness and the broader remaining work prevented completing the recovery changes for PostgreSQL 15 #94 #95.
Jun 2022 · 1 entry
- The thread ended with Justin recalling earlier efforts to remove the single-user recommendation and Peter discussing the lack of known production failsafe reports; the comment clarification was the only supplied landing, while the emergency command, revised recovery guidance, and final-check proposal had no demonstrated completion here #96 #97.