model-writtenrun 189 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Radim Marek reported reproducible memory exhaustion and a hard allocation failure near 105 million concurrent updates or deletes, attributing the growth to retained combo CIDs during REPACK catch-up #1.
- Shihao Zhong independently reproduced the memory growth and reported that an external REPACK enhancement series targeting PostgreSQL 20 avoided combo CIDs by replaying original transaction IDs without per-change command-counter increments #2.
- Radim submitted a documentation warning for PostgreSQL 19 #3; Antonin Houska recommended distinguishing REPACK's performance purpose from VACUUM's wraparound protection #4, and Shihao incorporated that guidance into the next patch #5.
- Alvaro Herrera revised the locking, memory, and wraparound explanations #6, after which Antonin advised against an exact per-tuple memory figure and emphasized disk usage, long transactions, and catch-up locking while approving the VACUUM guidance #7.
- Alvaro reorganized the reference page into dedicated Notes subsections and added monitoring guidance #9, then moved the concurrent-operation restrictions and expanded MVCC warning into the same section after considering Thom Brown's separate report #10.
- Antonin identified confusing resource wording #11 #14, Radim proposed clearer descriptions of table copies and sort space #13, and Alvaro finalized that wording and moved the long-transaction warning into the general resource section #16, with Antonin endorsing the revised structure before commit #17.
- Alvaro committed the documentation reorganization as 53c8cfc253 from #10, with light wording, placement, formatting, and disk-storage clarifications, and announced the push in #18.
- After the documentation landed, Nathan Bossart relayed release-management concerns about the practical limit and asked whether an incremental fix could help #19; Alvaro suggested trying a huge allocation while noting uncertainty about risks and subsequent limits #20.
- Shihao corrected the failing allocation's location to the comboCids array and supplied an experimental huge-allocation patch that completed 110 million replayed updates correctly, while warning that it preserved linear memory growth and could expose systems to OOM kills #21.
- Radim considered the documented limit acceptable with operational planning #22, Alvaro opposed a localized workaround given the queued PostgreSQL 20 redesign #23, and Nathan accepted that conclusion #24, leaving the documentation committed and the experimental allocation change unadopted.