model-writtenrun 7364 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr 2020 · 9 entries
- Andres Freund opened the investigation after snapshot-scalability work exposed unprotected XID wraparound and inconsistent time-to-XID mapping, potentially allowing pruning to produce silently wrong query results #1.
- Robert Haas considered the mapping defects fixable and opposed removing substantial code from stable branches, while Andres clarified that disabling the setting was a possible back-branch response and identified further locking, WAL-position, and testing concerns #2 #4 #10 #14.
- Andres supplied and refined debugging instrumentation to expose map contents and pruning-horizon changes, without yet proposing a production fix #3 #12.
- Kevin Grittner explained that deployed applications depended on the feature to prevent bloat from long-lived ODBC cursors, emphasizing that delayed cleanup was acceptable but silently inconsistent results were not; Robert proposed an introspection module and simulated time advancement to make investigation easier #15 #17 #18.
- Peter Geoghegan independently reproduced changing query results inside a serializable transaction, Andres traced this to missing checks in heap_hot_search_buffer() and identified additional unchecked access paths, and Kevin confirmed the index-scan failure existed back to the feature's introduction #21 #23 #26 #29.
- Andres found that the existing tests never actually removed tuples through the feature and argued for removal or redesign, while Peter proposed a simpler timer that would release old snapshots' xmin protection and reject their subsequent data access #34 #35 #36.
- Amit Kapila argued that prospective maintainers should have time to attempt repairs, and Thomas Munro subsequently prototyped Peter's timer design, explicitly leaving unregistered snapshots, check placement, timer overhead, and usefulness unresolved #37 #42.
- Thomas pursued a separate targeted repair that discarded obsolete map entries before CLOG truncation, then combined it with Robert's mapping fixes and introspection patches, a clock-manipulation helper, and TAP coverage #43 #44.
- Robert explained why user-facing introspection belonged in contrib, Thomas distinguished that from his test-only clock manipulation, and Thomas planned to continue the repair and testing work in Robert's separate thread #45 #46.
Jun 2021 · 6 entries
- After more than a year, Tom Lane revived removal as the alternative to a maintained repair; Andres supported removal, Peter proposed a firm deadline and an accountable maintainer, and Tom targeted a fixed-or-removed feature for PostgreSQL 15 #48 #49 #50 #51 #56.
- Robert clarified that a time-map repair had already landed separately and offered bounded bug-fixing work, but Andres required meaningful tests, a thorough review, and a documented checking model, while Thomas said the wraparound and testing work had stalled because he could not establish confidence in the whole mechanism #53 #59 #60.
- Noah Misch emphasized compatibility and existing user needs, while Greg Stark, Tom Lane, Stephen Frost, Peter, and Andres explored alternatives involving declared table access or finer snapshot-aware cleanup; update-chain correctness and implementation cost remained unresolved #61 #65 #66 #67 #68 #69 #71 #76.
- Thomas returned to the timer experiment with a substantial rework that tracked timed snapshots separately and checked heap visibility and visibility-map access, but still described it as incomplete and left catalog snapshots, TOAST, usefulness, and overhead open #78.
- The supplied thread ends with Greg defending the original use case and explaining why improved vacuum tracking or declared table-access restrictions would not directly replace it; no completed replacement or removal was announced within these messages #79.
- The later outcome was Thomas's September 2023 removal commit f691f5b80a, citing unresolved correctness and performance problems and the absence of an active repair plan; the supplied verdict for thread #47312 identifies its source as the later removal patch #536477, so this thread contributed motivation rather than the committed implementation.