model-writtenrun 9870 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tomas Vondra reported that VACUUM FULL and CLUSTER could exhaust logical-decoding memory because rewritten TOAST inserts were decoded while main-heap full-page records were ignored, and proposed suppressing those inserts and filtering tuple-less records before buffering them #1.
- Masahiko Sawada observed that the decoder change made the reorder buffer's null-tuple guard redundant #2; Tomas agreed #3 and posted a revision replacing that guard with an assertion and improving comments #4.
- Tomas committed #4 unchanged as f69c959df0 and announced backpatches through PostgreSQL 9.4 #5.
- Andres Freund questioned backpatching what he characterized as an optimization without advance notice #6; Petr Jelinek argued that out-of-memory failures could stop logical replication entirely and justified backpatching #7, which Andres accepted while requesting that rationale before committing and in the commit message #8.
- The thread ended with the fix committed and backpatched: Tomas acknowledged omitting advance notice and reiterated that replication failure made this a bug fix, with no further patch proposed #9.