model-writtenrun 7802 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Apr 2019 · 6 entries
- Pavan Deolasee revived Jeff Janes's earlier work from thread #33624 with #1, proposing to scan completed COPY FREEZE pages and set both page and visibility-map flags, including special handling for the final page.
- Kuntal Ghosh proposed marking new pages during insertion in #2 and #4, while Jeff Janes's TOAST test in #5 exposed skipped blocks that Pavan traced to relcache invalidation and loss of the saved insertion target in #6.
- Masahiko Sawada questioned the rescan and its cost; Pavan demonstrated that ordinary and frozen inserts can share a page in #10 and measured a large reduction in subsequent VACUUM time in #13.
- Masahiko's detailed review #15 led Pavan to add pg_visibility tests and avoid unnecessary buffer dirtying in #16.
- Andres Freund objected to heap-specific hooks in COPY and the extra scan, locking, and logging work in #20 and #22, then proposed moving visibility handling into heap insertion in #26.
- After Tom Lane questioned overhead in frequently used insertion routines in #29, Andres supplied the heap_multi_insert-only prototype #34, using insertion WAL flags and retaining the visibility-map pin across pages while leaving single-row insertion and WAL-interface cleanup unfinished.
May - Jun 2019 · 1 entry
- Pavan stepped back because of other priorities in #37, and after a long gap Ibrar Ahmed resumed the work in #39, rebasing Andres's prototype, restoring Pavan's tests, and measuring its performance.
Jul - Aug 2020 · 3 entries
- Daniel Gustafsson reported an uninitialized recptr warning in #42; Anastasia Lubennikova took over and fixed that case in #43.
- Robert Haas's concerns about duplicate flags and the WAL workaround in #44 led Anastasia to adopt advance visibility-map pinning and ordinary visibilitymap_set logging in #45.
- Ibrar initially measured a COPY slowdown in #46, but repeated tests by Anastasia in #52 and Ibrar in #53 found little difference in COPY time and substantially faster VACUUM.
Oct - Nov 2020 · 1 entry
- Tatsuo Ishii independently tested the patch in #55, and Tomas Vondra combined larger benchmarks with detailed code review in #56, leading to stronger assertions and comment cleanup in #57.
Jan 2021 · 6 entries
- The single-row path remained unresolved until Tomas demonstrated in #59 that the newer implementation left all TOAST pages without visibility flags; Anastasia added heap_insert handling in #60 and tested ordinary, TOAST, and trigger workloads in #62.
- Tomas confirmed the TOAST improvement but found extra allocated pages in #63; Anastasia fixed last-page reuse when the free-space map is bypassed and consolidated the implementation in #64.
- Tomas announced the landing in #66, but 7db0cd2145 actually used the older submission #59, with a comment addition and a replay hunk moved to the single-insert routine, leaving the later TOAST and page-reuse fixes out.
- Tom Lane reported CLOBBER_CACHE_ALWAYS buildfarm failures in #69, and Tomas acknowledged in #70 that he had accidentally committed the older version despite reviewing the newer one.
- Tomas restored the missing portions of #64 in 39b66a91bd, covering single-row visibility handling, multi-insert replay, and last-page reuse, then confirmed additional testing in #71.
- At the thread's end, the implementation and corrective commit had landed, while Tomas's additional TOAST regression test in #71 remained a proposal with no landing shown.