model-writtenrun 9347 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2018 · 3 entries
- Masahiko Sawada proposed FREEZE_ONLY to accelerate emergency wraparound prevention by skipping index and second-pass heap vacuuming while retaining freezing and HOT pruning, and supplied initial benchmarks #1.
- Nathan Bossart questioned how retained dead tuples would allow the freeze horizon to advance and suggested separating the behavior from FREEZE, while Robert Haas argued for an independent option naming the index work being skipped #4 #8 #9.
- After an October rebase and a gap until January, Sawada replaced FREEZE_ONLY with DISABLE_INDEX_CLEANUP and reclaimed dead tuple storage while preserving dead line pointers, avoiding the proposed need to freeze xmax #3 #15.
Jan - Jun 2019 · 21 entries
- Bossart tested HOT and index cases and requested clearer behavior for tables without indexes and consolidation of duplicated page-vacuum code; Sawada incorporated those changes #16 #18.
- Haas then showed that ordinary pruning already performed the necessary tuple removal and that newly dead tuples could wait for the next vacuum, leading Sawada to remove the special page-vacuum path and simply discard collected dead-tuple references when skipping indexes #19 #23 #26.
- Bossart's reporting review and Alvaro Herrera's objection to a no-index NOTICE led to quieter behavior, revised statistics and an added vacuumdb companion patch #27 #29 #30.
- Further testing by Bossart and Kyotaro Horiguchi exposed misleading pruning counts and vacuumed-page accounting; Sawada revised the counters and added a vacuum_index_cleanup reloption so autovacuum could use the feature #32 #39 #40 #41.
- Haas's review simplified strategy selection through useindex and established positive INDEX_CLEANUP boolean syntax, which Sawada implemented on top of the separately developed boolean VACUUM-option infrastructure #42 #43 #46 #47.
- Sawada proposed automatically disabling index cleanup for anti-wraparound vacuums, but Haas argued that this could worsen bloat and deferred any adaptive default until a later release, leaving explicit configuration in control #48 #49 #50.
- Haas supplied replacement documentation and identified that an explicit command option must override the table default; the resulting implementation used an enabled/disabled/default enum resolved separately for each relation #51 #53 #54.
- Darafei Praliaskouski questioned the explanation of skipped heap work, while Horiguchi reviewed enum values and argued that INDEX_CLEANUP should be ignored with FULL; Sawada revised the patch and separately proposed relaxing DISABLE_PAGE_SKIPPING with FULL #55 #56 #58 #59.
- After another review of dead-tuple accounting and enum initialization, Sawada posted the main patch alone, and Haas committed it with documentation edits as a96c41feec #60 #62 #63.
- The main landing did not include the vacuumdb companion or the proposed DISABLE_PAGE_SKIPPING/FULL relaxation, and this thread provides no later landing confirmation for those additions #59 #62 a96c41feec.
- Bossart supplied an inline correction adding the missing word “option” to the new documentation, which Haas committed as 6665305e17 #64 #66.
- Sawada then reported Fujii's observation that the reloption lacked TOAST support and submitted a small follow-up adding toast.vacuum_index_cleanup; Michael Paquier recorded it as an open item #69 #70.
- Tom Lane reported a buildfarm assertion failure caused by counting retained dead tuples when cleanup was enabled, prompting Sawada's initial condition-flip fix #71 #73.
- Lane reproduced the race and argued that tuples retained when cleanup is disabled must follow the normal retained-tuple freezing path; discussion with Haas, Herrera and Andres Freund examined freeze-horizon safety and exposed misleading tuple-versus-line-pointer reporting #74 #76 #82 #85 #89.
- Sawada combined the correctness fix with byte-based reclamation reporting, but Haas questioned the usefulness and scope of that reporting, so Sawada split the work into separate correctness and reporting patches #95 #96 #97.
- Haas subsequently reported committing the retained-tuple logic correction and removing both extra dead-item counters; he later explained why retaining a tuple without sending it through freezing could violate the intended invariants #101 #105.
- The safety investigation also revealed confusing but non-buggy xmin_frozen handling; after Freund proposed a simplification, Herrera posted a separate heap_prepare_freeze_tuple cleanup and committed it with an explanatory comment as 2bf372a4ae #88 #98 #106 #107.
- For reporting, Haas proposed separate counters for removed tuples and recycled dead line pointers, and Sawada agreed while questioning which statistics would help users; no revised reporting patch or landing is shown afterward #103 #108.
- Sawada revived the TOAST follow-up with a rebase, arguing that the omission was small enough to address for PostgreSQL 12; Freund requested regression coverage and identified a completion typo, which Sawada fixed in the next submission #109 #110 #112.
- Paquier took over the TOAST patch's final preparation and requested broader parent/TOAST option tests; Peter Geoghegan suggested EXTERNAL storage and a text index, which Paquier incorporated while deferring direct pageinspect verification #114 #115 #116.
- Paquier committed the TOAST implementation and expanded tests with a small final test correction as ce59b75d44, then closed the open item in the thread's last message; the reporting redesign and the concern about indefinitely accumulating dead line pointers remained without a recorded resolution here #89 #108 #118.