model-writtenrun 19 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Greg Burd revived the CRC failure in #1 after reproducing standby failures during concurrent hash inserts and vacuum: consistency checking read a page held only by a pin, allowing its contents to change between CRC calculation and WAL copying, so he proposed suppressing consistency-check images for REGBUF_NO_CHANGE buffers.
- Michael Paquier reproduced the failure and supplied simpler comments in #2, accepting the replay check for missing images and recommending backpatching only through v17 because older branches lack REGBUF_NO_CHANGE; Greg accepted that scope in #4.
- Andres Freund objected to retaining hash's buffer-rule exception on master in #3, while Michael proposed fixing the immediate failure on all affected branches, then reworking hash and removing the exception in #8; Greg accepted responsibility for that rework in #10.
- Greg posted the independent hash rework in #13, moving primary-bucket block numbers into WAL record data so replay could acquire cleanup locks without registering unchanged pages, plus a removal preview explicitly marked not for merge; his tests found that simply locking the bucket at registration could deadlock, whereas the proposed design preserved standby scan serialization.
- Vignesh supplied an injection-point recovery test in #14, confirmed that the narrow workaround fixed the failure, and explained that its claims about never taking full-page images were too broad because ordinary backup images could still be generated.
- Michael committed his revision from #2 as 27caf4ca5a, with further comment and documentation edits but unchanged executable logic, and announced backpatches through v17 in #16; the hash rework and injection-point test were not included.
- Michael requested a separate thread with the original hash authors for the larger rework in #17, and Amit Kapila considered that approach reasonable while suggesting conditional logging of bucket block numbers to save space in #18.
- The broader debate about hash indexes also produced Jeff Davis's proposal to derive predicates usable by functional indexes in #20; shihao zhong identified equality-semantics complications in #21, and Jeff encouraged pursuing that work separately in #22.
- Vignesh submitted a documentation follow-up in #23 explicitly distinguishing consistency-check images from torn-page protection; the supplied thread shows no landing for that clarification or the separate hash rework.
- Vignesh reported in #24 that an updated primary using hash consistency checking could stop an older standby and asked for release-note guidance; Michael declined because the setting is a developer option in #25, and Vignesh accepted that conclusion in the final message #26, leaving the immediate fix landed and the separate follow-up patches uncommitted in this thread.