model-writtenrun 2673 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Aug 2024 · 12 entries
- Nathan Bossart proposed separating reserved autovacuum capacity from a reloadable concurrency limit so busy installations could increase workers without downtime, while allowing existing workers to finish when the limit falls #1.
- Sami Imseih suggested sharing the background-worker pool, but Nathan explained that other workers could exhaust it and that reserving slots would broaden the work; they settled on a dedicated autovacuum pool #2 #4 #5 #6.
- Sami tested changing worker counts under table activity and saturated client connections, confirmed the intended behavior, and requested clearer naming and feedback when the requested count exceeded capacity #7 #9.
- Nathan avoided cross-dependent GUC check hooks and submitted a four-patch implementation retaining autovacuum_max_workers as the familiar concurrency setting while introducing a separate startup-only capacity setting #10 #12.
- Justin Pryzby proposed a separate policy setting instead, but Nathan and Sami favored preserving the established meaning of autovacuum_max_workers to simplify upgrades and administration #13 #14 #16.
- Sami verified cost balancing and reviewed the documentation; Nathan removed redundant wording and corrected the reloadability description, leaving broader cost-setting guidance outside this patch #18 #19.
- Robert Haas raised the larger problem of choosing worker counts and avoiding harmful cost-balancing or scheduling effects, while agreeing that this operational improvement need not solve automatic tuning #20 #21 #22.
- Nathan explored eliminating the extra setting entirely, first reserving 1024 slots and then lowering that fixed limit to 256 after Sami's review #24 #26 #27 #29.
- Andres Freund challenged the memory and cache costs of reserving hundreds of slots, reporting that TAP cluster shared memory would grow from roughly 5 MB to 18 MB, and proposed relating capacity to max_connections #30 #34 #36.
- Nathan tried a capacity of half max_connections capped at 64, temporarily bundling num_os_semaphores and GUC-hook cleanup prerequisites, but expressed concern about coupling unrelated settings and possible EXEC_BACKEND problems #38.
- Nathan then returned to the earlier configurable-pool design, naming the setting autovacuum_worker_slots, defaulting it to 16, and reserving only three slots for TAP clusters; the experimental prerequisites were omitted from this renewed feature patch #39.
- After rebasing, Nathan added warnings for requests exceeding reserved capacity, choosing launcher checks instead of cross-dependent GUC hooks #40 #41.
Nov 2024 - Jan 2025 · 6 entries
- Following several months and another rebase, Yogesh Sharma reported expected worker behavior and warning output; Nathan subsequently moved the initial warning to autovac_init() to avoid repetition when the launcher restarts #44 #45 #46.
- Nathan rebased once more and committed the final attachment unchanged as c758119e5b, making autovacuum_max_workers reloadable within the startup-reserved pool #47 #48.
- Tom Lane immediately reported that the additional semaphore demand broke buildfarm member sawshark under the small BSD defaults, prompting Nathan to propose lowering the global slot default from 16 to 5 #49 #51.
- Andres suggested letting initdb select a lower value only where necessary, and Tom proposed coordinating that choice with max_connections; Nathan first implemented a separate slot-probing loop #52 #53 #54.
- Tom objected to extra initdb subprocess probes, so Nathan folded slot selection into the connection tests using division by eight; Tom then suggested division by six to preserve 16 slots on ordinary systems #55 #56 #57.
- Tom accepted the revised formula, and Nathan committed #58 with only a comment typo correction as 6d01541960, retaining the normal default while reducing it on constrained systems #59 #60.
Apr 2025 · 4 entries
- In April, Peter Eisentraut objected that initdb's separate slot-selection progress report described work already completed; Tom proposed simply removing it, Peter agreed, and Nathan committed that discussion-driven deletion as 2b49492eda #61 #62 #64 #65.
- Tom initially proposed reverting both the earlier BSD semaphore accommodation and adaptive slot selection, but reconsidered because scaling autovacuum capacity down with small connection limits remained useful; Nathan agreed that the derived capacity must stay at least three #66 #67 #68 #69.
- Tom committed 810a8b1c80, restoring 16 semaphores per set and a minimum trial connection count of 20 while retaining adaptive slot selection and documenting that NetBSD/OpenBSD semaphore limits must be raised #70.
- The thread ended with Nathan accepting the final updates: the reloadable-worker feature and adaptive initdb sizing remained committed, the misleading progress report was removed, and no attached patch remained awaiting action #71.