model-writtenrun 8651 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - Apr 2019 · 3 entries
- Darafei Praliaskouski opened the thread with #1, proposing that autovacuum count inserts as well as other changes so append-only tables receive visibility-map maintenance and avoid an enormous first anti-wraparound vacuum.
- Alvaro Herrera argued for separate visibility and freezing actions in #2, while Masahiko Sawada distinguished those operations from expensive index cleanup and truncation in #9 and #12.
- Chris Travers considered skipping index cleanup or introducing a freeze worker in #25, but Tomas Vondra warned against multiplying cleanup-worker types in #26; Darafei instead proposed preserving ordinary vacuum triggers and adding an insert-specific trigger in #28.
Jul 2019 · 1 entry
- Masahiko described a separate insert counter and ordinary-vacuum precedence in #31, but the architecture remained unsettled through #32 and #35, with partial scans and visibility-map updates during foreground operations remaining proposals.
Mar - Apr 2020 · 20 entries
- Laurenz Albe revived the work in March 2020 with #36 and #37, replacing the original shared counter with a separate insert threshold and initially skipping index cleanup; #38 added age-zero freezing to make these vacuums help future wraparound work.
- David Rowley's review #39 requested configurable scaling, public statistics, documentation, and protection against index bloat; Laurenz's revisions #41 and #44 added those facilities and restored normal index cleanup after further discussion in #42 and #43.
- David and Masahiko found documentation, logging, regression-output, and multixact-freezing omissions in #46 and #49, which Laurenz corrected in #50.
- The default-setting debate weighed frequent, manageable vacuums against index-only-scan benefits and total I/O, with David demonstrating scan benefits in #58 and Darafei favoring a fixed insert threshold to process recently written data in #66.
- Justin Pryzby's testing in #70 challenged forced freezing on insert-mostly tables, and Andres Freund explained the associated write and WAL costs in #71 and #81.
- Laurenz tried opportunistic per-page freezing in #86, but Justin found an incorrect condition and tuple-order dependence in #87; Laurenz removed special freezing entirely in #88, leaving ordinary vacuum triggered by inserts.
- Justin's analysis of stale reltuples estimates in #89 and #93 led to the larger scale-factor limit in #94, while David supplied a documentation and 0.01-default fixup in #106 that Laurenz folded into the rebased #109.
- Andres explained in #112 and #118 that frequent insert vacuums could repeatedly scan entire indexes, persuading Laurenz to adopt the more conservative 0.2 scale factor in #122 and leave freezing improvements for separate work.
- Justin questioned inconsistent limits in #124, Andres requested an explicit off switch in #127, and Alexander Korotkov identified the reloption-default conflict in #128; David resolved these in #131 with a -1 disable value, an internal -2 inheritance value, and a 1000-row default threshold.
- After explaining the internal sentinel in #133, David committed #131 as b07642dbcd, adding documentation, explanatory comments, and a catalog-version bump, and announced the landing in #135.
- David and Tom Lane reported intermittent partition_aggregate and stats_ext failures in #136 and #137; David proposed telemetry in #139 but worried that it would introduce failures without proving the cause.
- David instead committed explicit VACUUM ANALYZE operations for pagg_tab_ml and mcv_lists in 2dc16efedc, explaining the departure from the telemetry patch in #140.
- Amit Kapila reported another stats_ext failure in #142, prompting David's broader investigation in #145 and #446456 and additional explicit vacuums for other statistics-test tables in 24566b359d.
- While investigating continued failures, Dean Rasheed found that statext_store() looked up its target tuple before locking the catalog in #148; Tom implemented the proposed ordering fix in fe3036527a but explained in #150 that this race did not account for the buildfarm failures.
- Tom developed a concrete explanation in #152: concurrent non-aggressive autovacuum could skip buffer-locked pages and slightly change reltuples, disturbing exact estimates and the ordering of equally costly plan branches.
- After David identified another partition-test failure in #153, he committed cefb82d49e, disabling autovacuum on the test tables and reverting that test's earlier explicit-vacuum workaround.
- Tom objected in #155 that disabling autovacuum hid fragile test design, and posted #156 to give partitions different row counts so their plan ordering would remain stable and exercise cost ordering.
- David acknowledged looking over the replacement in #157, and Tom landed it unchanged as 7cb0a423f9, confirming the push in the thread's last message #158.
- The supplied later commit 18d85e9b8a corrected one remaining equal-size split at the upper level of pagg_tab_ml after another buildfarm failure, extending the partition-boundary approach beyond the thread's last message.
- By #158, the insert-trigger feature and the posted partition-test replacement had landed, the diagnostic attachment had been abandoned, and special freezing remained deferred; the later follow-up 18d85e9b8a shows that test stabilization continued afterward.