model-writtenrun 8082 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas opened #1 with three patches exposing oldSnapshotControl, adding SQL inspection of the time-to-XID map, and repairing its bucket-advancement arithmetic and head timestamp; he wanted more testing before committing.
- Andres Freund argued for finer time bins to make edge cases testable in #2 and rejected moving only the head timestamp backward in #4, while Robert preferred avoiding the code churn of changing binning in #3.
- Thomas Munro supplied a timestamp-control function in #5 that moved the effective clock into the future, then combined it with Robert's patches, CLOG-linked map truncation, and draft TAP tests in #6.
- Dilip Kumar challenged Thomas's diagnosis of a large-time-jump bug in #7, which Thomas accepted in #10; Andres meanwhile cautioned in #8 that controlled timestamps helped known-case testing but did not replace finer bins for discovering unknown problems.
- Thomas reworked the testing approach in #10, replacing custom interactive Perl machinery with existing isolation tests, removing the special zero-threshold test branches, and placing clock manipulation in a dedicated test module.
- Dilip proposed memcpy for copying the map under its lock in #9; after Robert reported a qualifier problem in #14, Dilip supplied a cast-based solution in #16, but the eventual extension retained the loop.
- Thomas and Dilip discussed proving actual tuple removal in #12 and #13, leading to #15, which used pg_visibility to verify early pruning and reused the introspection extension instead of a duplicate test helper.
- Thomas added map-maintenance tests validating Robert's fixes in #17; Dilip requested stronger checks of which slots received new XIDs in #18 and supplied a separate fast-path comparison change in #19.
- After nearly four months, Thomas rebased the series in #20 while asking Dilip to explain the comparison change, which remained outside the series; a further rebase in #21 and Makefile correction in #22 restored compatibility.
- Michael Paquier urged progress on the bug-fix subset in #23, and Robert decided in #24 to land the first three patches on master because backpatching the fix without observability would make failures difficult to investigate.
- Robert posted the landing subset with PGDLLIMPORT and documentation in #25, and Hamid Akhtar reported make installcheck-world passing in #27.
- Robert committed the three patches from #25 without substantive diff changes as f5ea92e8d6, aecf5ee2bb, and 55b7e2f4d7, crediting Thomas, Dilip, and Hamid as reviewers.
- In the commit announcement #28, Robert asked whether Thomas's testing functions could be made safe enough for contrib/old_snapshot, avoiding a second loadable module and potentially helping users.
- Thomas resent the three remaining patches in #30: the test rewrite, the XID-wraparound truncation fix, and a smoke test for the committed mapping repair; he was still considering safe public test functions and noted unresolved missing TestForOldSnapshot() calls.
- At the final message #31, Anastasia Lubennikova found the entry still waiting on the author and asked whether the remaining work needed review, leaving the test and wraparound portion uncommitted in this thread.
- The later feature-removal commit f691f5b80a cites this discussion as background, but its matching implementation came from #536477 in thread #47312, rather than these repair patches.