model-writtenrun 8846 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thunder reported that canceling vacuum after logging truncation but before physically truncating could leave the standby missing pages subsequently referenced by WAL, and proposed an inline patch holding cancellation interrupts around truncation #1.
- Michael Paquier supplied a failure injector demonstrating that errors beyond cancellation could trigger the same problem #4, while Kyotaro Horiguchi identified an indirect interrupt check supporting the original cancellation scenario #6.
- When Fujii Masao asked about a critical section #9, Michael argued that failed physical truncation would then crash the primary without necessarily preventing standby replay failure #10.
- Fujii proposed a developer setting to let recovery proceed past otherwise persistent invalid-page failures #11 and attached ignore_invalid_pages, defaulting off and downgrading the relevant PANIC reports to warnings when enabled #13.
- Michael's reviews prompted documentation formatting fixes and explicit PANIC terminology #14 #16, incorporated into Fujii's revisions without changing recovery logic #15 #17.
- Michael clarified that the setting ignores invalid page references rather than entire WAL records #18; Fujii proposed replacement wording #19, which Michael accepted before marking the patch ready for committer #20.
- Amit Kapila reopened the unresolved original truncation problem #21; after correcting a mistaken association with another issue, Michael remained uncertain about the solution #24, and Fujii suggested reconsidering interrupt suppression as a partial improvement #25.
- Fujii committed 41c184bc64 from #17 with light documentation and help-text refinements, then announced the landing #26.
- The thread ended with the recovery escape hatch committed but the underlying truncation bug unresolved: Andres Freund argued that discarded dirty buffers and WAL logging an unperformed truncation cause corruption worse than a PANIC restart, pointed to a broader proposal, and rejected interrupt suppression as useful in isolation #27.