model-writtenrun 9827 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund proposed computing index-vacuum conflict XIDs on the primary to support pluggable storage and logical decoding on standbys, attaching a prototype with sorted heap accesses and unconditional prefetching; he also flagged contradictory comments about invalid XIDs and a possible recovery race #1.
- Alexander Korotkov suggested pruning heap tuples before marking index entries dead, but Andres objected that this could require individual WAL records and preferred considering additional pruning during page-level deletion #2 #3.
- Peter Geoghegan supported moving the work to the primary and proposed deleting only enough entries to avoid a split, with possible additional deletions from heap pages already visited, while explicitly leaving these optimizations outside the initial version #4.
- The thread ended with Alexander suggesting whole-page pruning or logging a conflict horizon only when it advances; these alternatives, performance tradeoffs and the reported recovery race remained unresolved, with no revised patch or commit announcement posted here #5.
- The approach later landed as 558a9165e0, attributed by the supplied records to the newer #421769 in thread #38991, incorporating table-AM dispatch, bounded prefetching and removal of obsolete heap relfilenodes from WAL.