model-writtenrun 5626 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2021 - Apr 2022 ยท 25 entries
- Peter Geoghegan opened the thread with #1, proposing a common definition of scanned heap pages so VACUUM could collect existing LP_DEAD items and preserve opportunities to advance relfrozenxid even when a cleanup lock was unavailable.
- Andres Freund's initial review in #2 supported the direction but challenged the patch's breadth, requested stronger testing, and suggested tracking the oldest actually observed transaction ID instead of relying entirely on fixed freezing cutoffs.
- Andres also argued that cleanup-lock failures still needed useful diagnostics and warned about the WAL cost of eager freezing in #4 and #6; Peter responded with separate missed-dead-tuple instrumentation and no-pruning FSM processing in #8.
- The series expanded in #9 to track the oldest remaining XID and MultiXact independently of tuple-freezing cutoffs, alongside a prototype that froze early after pruning or when a page could become all-visible.
- Masahiko Sawada reviewed scan-state organization and last-page handling in #10; Peter accepted moving skipwithvm into LVRelState but defended always scanning the last page as a worthwhile simplification in #11.
- Robert Haas challenged proposals to remove freezing settings or delay cleanup-lock waits in #15, #18, and #20, emphasizing irregular vacuum scheduling and workloads where waiting is necessary; Peter subsequently proposed retaining a configurable escape hatch in #34, which Robert welcomed in #35.
- Peter changed the freezing policy in #39 to operate on whole pages instead of freezing merely because pruning dirtied a page, and added an experimental FSM heuristic intended to preserve settled pages while reporting long-running TPC-C measurements.
- John Naylor asked about space utilization in #40, while Robert questioned the FSM heuristic's workload assumptions and requested broader testing and configurability in #42 and #44.
- The first three components of #39 landed as scan simplification 44fa84881f, instrumentation 872770fd6c, and cutoff consolidation efa4a9462a, with unchanged instrumentation and light cleanup elsewhere; Peter resent the remaining three patches in #54.
- After concerns about the project's scope, Peter proposed targeting only observed-horizon advancement for PostgreSQL 15 in #55; Robert and Andres supported reducing scope in #56 and #60.
- Robert requested clearer input/output contracts for horizon tracking in #58, and Andres requested isolation tests for unavailable cleanup locks in #59, leading Peter to investigate adversarial lock-failure testing.
- That testing exposed an infinite loop involving pg_surgery-created orphaned HOT tuples in #63; Peter's initial repair idea proved insufficient, and Andres reproduced and explained the triggering sequence in #69 and #72, leaving pruning robustness as separate work.
- The revision in #82 dropped the FSM experiment, reimplemented page freezing with separate freeze/no-freeze tracking states, and introduced a fix for non-aggressive VACUUM losing horizon advancement when visibility-map bits changed concurrently.
- Andres's detailed review in #83 identified MultiXact handling concerns and reiterated the risk of excessive WAL and MultiXact creation; Peter agreed in #89 that observed-horizon advancement could ship independently of page-level freezing.
- Peter developed the skipping fix into a range-based helper in #92, then added extensive no-cleanup-lock isolation tests and removed page-level freezing from the submitted series in #93, deferring that work to PostgreSQL 16.
- A proposed change making DISABLE_PAGE_SKIPPING stop forcing aggressive VACUUM met objections from Robert in #94 and #96, with Thomas Munro also describing confusing existing behavior in #98; Peter agreed to drop the proposal in #97.
- Robert's review in #100 and #102 pushed for a clearer explanation of the skipping race and an understandable separation of behavior changes from refactoring; Peter retained a single skipping patch but revised its explanation and added a separate resource-allocation cleanup in #104.
- Further review of function contracts and documentation in #105 led to the clearer heap_tuple_would_freeze() name and other editorial improvements in #107.
- Justin Pryzby reported a concrete assertion failure in #108; Andres reproduced it in #114 and traced it in #119 to shared-relcache horizons differing from the database-local catalog values.
- Peter added a temporary PANIC diagnostic in #113, but Andres established in #129 that the reproduced assertion failure was not an actual future-valued catalog horizon; the final series in #131 removed the unreliable assertions, retained separate warnings, and left the underlying shared-relcache issue for separate work.
- The logging documentation tip from #131 was extracted and reworded into 05023a237c before the main horizon-tracking change landed.
- The final horizon-tracking implementation and tests from #131 landed with light cleanup as 0b018fabaa, followed by the range-based visibility-map skipping component as f3c15cbe50.
- The resource-allocation component from #131 landed as c42a6fc41d, retaining the final progress reports that the submitted patch had proposed removing.
- Andres reviewed and accepted the final warnings in #135, and Peter confirmed their landing in #137 as e83ebfe6d7.
- The reduced submitted series was fully landed, while page-level freezing and broader policy changes remained future work; the closing discussion in #132, #133, #138, and #139 explored accumulated VACUUM statistics, worker utilization, and taking an upfront visibility-map snapshot to bound scanning work.