model-writtenrun 7923 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut introduced a patch developed with Simon Riggs to keep standbys available after primary parameter increases: warn about mismatches, continue replay, and pause only if shared-memory resources actually run out, particularly to avoid cascading standby shutdowns after failover #1.
- A related sizing question remained separate: Peter observed that autovacuum_max_workers contributes to shared-memory sizing without being tracked in WAL, Michael Paquier favored tracking individual settings for diagnostics, and Peter suggested tracking combined MaxBackends instead; the supplied thread records no implementation of either option #1 #5 #6 #7.
- Alvaro Herrera questioned what would happen after users resumed recovery, and Peter clarified that resuming would retain the existing fatal shutdown behavior rather than allow replay to continue #8 #10.
- Masahiko Sawada argued for repeated warnings once recovery was paused, while Kyotaro Horiguchi's testing showed that lock-exhaustion advice could name the wrong parameter; Peter incorporated minute-by-minute warnings, less specific advice pointing to earlier diagnostics, and expanded documentation #9 #12 #13 #14 #16.
- Masahiko reviewed the revised loop and requested recovery-pause wait-event reporting, Sergei Kornilov supported that addition and asked about a function prototype, and Peter committed the patch with wait-event instrumentation and small comment corrections as 246f136e76 #17 #18 #19.
- Peter returned in June with a substantially simpler replacement that pauses immediately on an incompatible parameter-change record, splitting out the macro-to-function conversion and abandoning the wait-until-resource-exhaustion approach; he omitted configurability because it had attracted little interest #20.
- The June replacement starts from the original error-producing parameter check, showing that the March implementation was no longer in its base, but the supplied thread does not explain the intervening removal or its rationale #20.
- After a minimal July rebase, Sergei reviewed the simplified patch in November and requested restored wait-event reporting, repeated parameter details in the eventual fatal error, and consistent pause wording #21 #22.
- Peter landed the unchanged preparatory conversion from the July series as b5acf10cfc, then submitted the remaining recovery-pausing patch with Sergei's diagnostic improvements, retaining the word 'unpaused' because recovery would not actually continue #21 #23.
- Sergei's testing exposed a pause before the standby could accept connections, and he proposed handling promotion requests explicitly; Peter restricted pausing to active hot standby and added a one-time warning that promotion requires corrected settings and a restart #24 #25.
- Sergei reported successful installcheck-world, feature and documentation testing of the final version, and Peter committed it unchanged apart from rebasing as 15251c0a60; the thread ends with the revised series fully landed, preserving read access during the pause while requiring configuration changes and restart before recovery can continue #26 #27.