model-writtenrun 8387 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Fujii Masao proposed removing non-fast promotion because the emergency fallback retained when fast promotion was introduced had survived for more than six years without an apparent continuing need #1.
- Robert Haas and Michael Paquier favored removal earlier in a release cycle, and Fujii accepted targeting PostgreSQL 14, although Alvaro Herrera and Andres Freund questioned the need to delay such a small change #2 #3 #4 #5 #6.
- Fujii submitted the initial removal patch for the first PostgreSQL 14 commitfest, adopting Andres's clearer term non-fast promotion and removing the fallback signal file and separate fast-promotion flag #7.
- Michael identified that sending SIGUSR2 alone would cease to promote a standby and suggested retaining that behavior through fast promotion; Fujii maintained that this undocumented interface should be replaced by documented promotion methods #8 #12 #13.
- Compatibility checks found that PAF, repmgr and the pgsql resource agent already used supported promotion interfaces, while a source inspection suggested Patroni did too, reducing concern about removing signal-only promotion #17 #21 #22 #27.
- Kyotaro Horiguchi questioned retaining an immediate checkpoint when the previous checkpoint record was unavailable; Fujii explained its crash-safety purpose, and Kyotaro clarified that always emitting an end-of-recovery record would simplify signaling without improving performance, a broader change that was left out #15 #16 #20 #25 #26.
- Alvaro's objection to an obsolete historical comment led Fujii to rewrite it and move it into CheckPromoteSignal() in the second patch #19 #23.
- Hamid Akhtar tested the second patch successfully with installcheck-world and recommended reusing CheckPromoteSignal(); Fujii incorporated that suggestion and reused RemovePromoteSignalFiles() in the third patch, while rejecting a merger of signal-checking functions because they serve different processes #29 #30.
- Kyotaro noted that the promoted flag could remain false even when promotion succeeded through the immediate-checkpoint path; Fujii discussed alternative names, but the submitted and committed code retained promoted #31 #32 #30.
- After Hamid's final checks found only patch-application offsets, Fujii committed the third patch unchanged apart from rebasing as b5310e4ff6 and announced completion, leaving no attached patch awaiting commitment #33 #35 #36.