model-writtenrun 3143 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Apr 2021 · 5 entries
- Pavel Borisov proposed detecting visible duplicate heap entries behind unique B-tree indexes, including posting lists and page-boundary cases, with work credited to himself and Anastasia Lubennikova in #1.
- Mark Dilger proposed constructing duplicate data through catalog changes instead of recompiling the server in #3; Pavel adopted that regression-test approach in #4 and removed unstable LSN output in #5.
- Mark argued that callers should be able to avoid the additional heap-checking cost in #7 and #12, while Tom Lane questioned 32-bit portability in #11; Pavel added an optional argument and alternate expected output in #13.
- Peter Geoghegan recommended sharing the heapallindexed snapshot and testing with a deliberately faulty operator class in #14, then argued for immediate errors and against speculative corruption warnings in #15.
- Pavel defended retaining cross-page checks because a customer's corrupted index had required them, while offering to suppress uncertain warnings and follow existing error conventions in #16.
Dec 2021 - May 2022 · 3 entries
- After several months away, Pavel returned with #19, adding pg_amcheck support and documentation, changing tests to operator-class fault injection, sharing snapshots, following sibling links correctly, and replacing warning accumulation with immediate errors.
- Mark's review in #20 exposed missing attachments and compatibility problems with older remote amcheck versions; #21 restored the files and added version-aware client behavior, and Maxim Orlov supplied the remaining test cleanups in #23.
- Pavel fixed a declaration-order build problem in #25 and added previously omitted NULL-key handling and tests in #26.
Jul 2022 · 1 entry
- Following several upstream-driven rebases, Aleksander Alekseev reviewed the feature and supplied named SQL arguments, corrected documentation and comment cleanup in #32.
Sep 2022 · 1 entry
- Dmitry Koval corrected stale error-reporting comments and TAP conventions in #33, Karina Litskevich aligned initializer order in #34, and Maxim added Meson integration in #36.
Oct 2023 · 1 entry
- After another long gap, Alexander Korotkov rebased and polished diagnostics and tests in #39; that complete patch landed unchanged as 5ae2087202.
Mar - May 2024 · 14 entries
- Noah Misch's post-commit review found that uniqueness checking increased a large-table test from 58 seconds to 276 seconds and proposed lazy heap visibility checks in #41, later correcting his percentage calculation in #43.
- Peter accepted that the shared MVCC snapshot made lock coupling unnecessary and supported lazy visibility checks in #42, then identified problematic target-state replacement and silent handling of non-leaf siblings in #44.
- Peter Eisentraut requested friendlier option names in #46, and Alexander proposed grouping the last-visible-entry fields into a struct alongside the other fixes in #48.
- Pavel submitted five follow-up patches in #50, reporting that the optimization reduced his uniqueness-check run from 419 seconds to 171 seconds and also implementing the struct, sibling-page, error-reporting and naming proposals.
- Karina objected that the proposed option rename conflicted with existing naming conventions in #51; Alexander omitted that patch for lack of consensus and polished the remaining four in #55.
- Tom Lane challenged landing these changes after feature freeze in #56, and Mark distinguished the optimization from the refactorings and error-reporting improvement while raising specific code concerns in #60.
- Mark demonstrated that existing tests did not exercise the sibling-page path and traced subsequent accesses to state->target in #64 and #65; Tom called for tests addressing those concerns in #66.
- Alexander concluded that replacing state->target was wrong but had no visible consequence because the resulting low key was unused on leaf pages, and documented that reasoning without changing the code in #68.
- Pavel supplied an independent proposal to stop collecting unused leaf-page low keys in #69, explicitly keeping it outside the uniqueness follow-up series.
- Mark supplied a developer-only corruption generator in #74 and initially reported missed corruption, but traced that result to an unrecognized high-key error message in #77 and reported successful detection after correcting expectations in #80.
- The last-visible-entry refactoring from #68 landed as 532d94fec3, adjusted to apply without the deferred optimization.
- The sibling-page patch from #68 landed as 0b5c161248, preserving the target page while holding the sibling in a local variable.
- The non-leaf sibling error-reporting patch from #68 landed unchanged as 97e5b0026f, completing the three May follow-ups selected for PostgreSQL 17.
- The supplied record therefore accounts for the original feature and four follow-up commits, while the rename was dropped in #55, the separate low-key proposal in #69 has no demonstrated landing here, and the corruption generator in #74 remained explicitly outside the intended production changes.
Jul 2024 · 2 entries
- Alexander returned with only the remaining optimization, rebased over those changes, in #84 and proposed landing it for PostgreSQL 18.
- The thread ended with administrative discussion in #85 and #86, while the optimization was still proposed; the supplied commit record subsequently shows it landing unchanged as cdd6ab9d1f.