model-writtenrun 9546 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund reported that GiST microvacuum could remove index entries still visible to standby snapshots because replay lacked recovery-conflict handling, and Alexander Korotkov undertook to fix it #1 #2.
- Alexander Korotkov supplied a patch modeled on B-tree and hash conflict handling, introducing a new WAL record and calculating the conflict horizon on the standby #3.
- Andres Freund objected to introducing a WAL record in minor releases and proposed calculating the horizon on the primary with an existing conflict record; Alexander Korotkov acknowledged that option but sought a less invasive backpatch #4 #5.
- Alexander Korotkov proposed appending conflict information to existing WAL records so older versions could ignore it, while Peter Geoghegan favored retaining the B-tree/hash approach on master #6 #7.
- Alexander Korotkov posted the compatible backpatch variant, keeping horizon calculation on the standby because moving it to the primary would introduce substantial new behavior into stable releases #8.
- Testing demonstrated that a patched standby canceled the conflicting transaction while an unpatched cascading standby replayed the same WAL but returned an incorrect count; Alexander Korotkov then submitted polished master and backpatch variants #9.
- The master variant landed unchanged as c952eae52a, with compatible backports through PostgreSQL 9.6, and Alexander Korotkov confirmed the push #10.
- Months later, Andres Freund supplied a separate conversion after noticing that GiST had missed the shared primary-side horizon work; the thread ends with his request for comments, and that attachment subsequently landed unchanged as b5f58cf213, including WAL description support and a WAL-version bump #11.