model-writtenrun 3656 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2023 - Jan 2024 ยท 19 entries
- Melanie Plageman proposed marking removable line pointers LP_UNUSED during pruning of tables without indexes, avoiding a separate heap-vacuum WAL record; her example reported a 10% vacuum speedup and 40% fewer WAL bytes, alongside locking and truncation-bookkeeping prerequisites #1.
- Peter Geoghegan questioned reliance on cached visibility-map state, while Melanie explained the tradeoffs involving redundant VM updates and page-skipping logic; removing that state remained future work #2 #3.
- The buffer-lock prerequisite disappeared after an upstream fix, leaving two patches, and Melanie later confirmed that existing vacuum regression tests exercised the no-index deletion path #4 #5.
- January review focused on whether extra pruning branches justified optimizing an uncommon case: Robert Haas questioned performance and naming, while Andres Freund identified plausible bulk-loading workloads and suggested predictable branches with unlikely() hints #8 #9.
- Andres challenged the misleading replay parameter and scattered per-page processing, and Robert requested the broader refactoring that the optimization enabled; Melanie responded with VM relocation, pruning-state removal, and consolidated FSM updates #10 #15 #19.
- Melanie revised dead-item handling and temporarily dropped hastup removal after objections to moving truncation bookkeeping, while retaining the broader refactoring patches #22.
- Robert proposed deriving assertion behavior from the recorded dead-item count instead of enlarging WAL records or passing misleading index information during replay; Melanie implemented that approach and then added the requested assertion that an item was not already unused #32 #36 #39.
- After Michael Paquier checked the truncation paths and Andres relaxed his objection, Robert committed hastup removal from the restored series as e2d5b3b9b6, with a small adjustment for landing it independently #28 #35 #39 #43.
- Robert demonstrated that ordinary table-access locks cannot prevent concurrent index creation, so immediate removal remained disabled for on-access pruning; Melanie renamed the option mark_unused_now and added an explicit dead-or-unused wrapper in the remaining series #42 #44.
- Review of recordfreespace exposed confusion about existing FSM behavior, which Peter clarified as an attempt to record free space once per scanned page, normally on its last visit; a real inconsistency remained when index vacuuming was disabled in the no-prune path #45 #52 #54.
- Robert posted an independent FSM consistency patch, Peter caught its invalid use of pruning state, and Melanie reviewed the replacement output parameter and comments; the final revision landed unchanged as 45d395cd75 #56 #57 #64 #65 #66 #67.
- Melanie rebased the main series over that fix and returned to consolidating FSM updates in lazy_scan_heap(), abandoning the alternative of duplicating them inside both per-page helpers #63 #70.
- Robert suggested organizing processing around got_cleanup_lock, and Melanie adopted that structure so new or empty pages were handled before either pruning path; she also explained that PageRepairFragmentation already provided the line-pointer truncation behavior Robert feared losing #71 #72 #73.
- Discussion distinguished updating individual FSM entries from vacuuming the FSM itself, and Melanie added guards intended to preserve the existing cadence before Robert committed the unchanged pruning component as c120550edb #76 #81 #83.
- Robert and Peter explored advertising free space during the first heap pass, but Peter highlighted pages whose LP_DEAD stubs make their otherwise free space unusable until the second pass; broader changes to FSM timing and allocation policy remained outside this series #80 #86 #88 #89.
- Robert committed the VM-relocation component of #81 unchanged as cb970240f1 and its pruning-state removal component as e313a61137, with small variable-scope, initialization, comment, and formatting edits to the latter #90.
- Only FSM consolidation remained, and Melanie reposted it before revising its comments to explain the new conditional processing structure rather than merely combining old text #91 #92 #94.
- Final review clarified that failsafe activation or bypassing index vacuuming can leave some previously scanned pages without an FSM update; Peter explained the failsafe behavior, while Melanie and Robert distinguished the bypass case and left changing that policy for separate work #95 #96 #97 #98 #99.
- Robert incorporated the agreed comment changes into #94 and committed FSM consolidation as 5eafacd279; Melanie marked the entry committed, and the thread ended with the submitted work landed and broader pruning and FSM redesign ideas deferred #101 #102 #103.