model-writtenrun 231 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thom Brown reported a logical-decoding crash during concurrent repacking in #1, identifying unsuppressed TOAST changes and tuple-less UPDATE records as interacting problems when an output plugin requests heap rewrites.
- Antonin Houska supplied a reliable isolation-test reproducer and a heap_update() fix in #2, selectively propagating logical-decoding suppression instead of blindly forwarding update options; Thom confirmed that the fix stopped the crash in #3.
- Masahiko Sawada retained the reproducer as a registered regression test, merged it with Antonin's fix, and switched to HEAP_INSERT_NO_LOGICAL in #4; he also added a separate DecodeUpdate() fix that checks for a missing new tuple instead of introducing another WAL flag.
- Zhijie Hou approved both fixes and requested clarification about catalog updates in #5; Masahiko incorporated that comment and whitespace cleanup in #6, then endorsed the patches in #7, followed by Álvaro Herrera's approval in #8.
- Masahiko landed the two members of #6 separately as 7164dcb665 for TOAST suppression and 3a3524e9ac for tuple-less UPDATE filtering, preserving their code and tests and confirming the push in #9.
- Alexander Lakhin reopened the test issue in #10, reporting a buildfarm failure and reproducing inflated decoded-change counts locally under frequent autovacuum.
- Masahiko supplied a follow-up in #11 that enables skip-empty-xacts and disables transaction IDs in both decoding steps, preventing empty autovacuum transactions from affecting the count.
- The follow-up landed as adffcab923 with the attachment's test changes plus corrected expected output; Masahiko confirmed the push in the final message #12, leaving all submitted fixes landed.