model-writtenrun 370 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2025 · 10 entries
- Nathan Bossart opened the thread with a small PostgreSQL 19 proposal to sort eligible autovacuum tables by transaction or multixact age, replacing their existing catalog-scan order #1.
- Sami Imseih warned that older, expensive tables could occupy workers while smaller tables waited, Álvaro Herrera suggested different worker modes, and Andres Freund raised concerns about repeatedly encountering problematic tables in a deterministic order #2 #3 #4.
- David Rowley proposed expressing each maintenance need as a ratio to its existing trigger threshold and taking the maximum, allowing transaction ages and tuple counts to share a priority scale #10 #13.
- Nathan's next prototype combined ordinary scores additively and gave anti-wraparound work a separate tier, but Robert Haas and David objected that this could overweight secondary problems and abruptly favor mild wraparound risk over severe bloat #19 #20 #22 #24.
- Nathan adopted a unified maximum score with exponential freeze-score scaling, then followed David's feedback to reserve aggressive scaling for effective failsafe ages and explain the policy in comments #26 #27 #30 #38.
- Reviews corrected reversed age subtraction, simplified maintenance eligibility and score initialization, and removed redundant checks; Sami and wenhui qiu also caught a duplicated assignment that Nathan fixed #41 #42 #46 #48 #50 #51 #53.
- Sami requested documentation and an escape hatch for workloads affected by the new order, while Nathan initially preferred to avoid another configuration setting #52 #55.
- At Robert's request for practical evidence, Sami tested mixed batch and OLTP workloads and David examined logged scores, finding workload-dependent shifts but no decisive overall improvement; subsequent tests compared sorting enabled and disabled, and Nathan added DEBUG3 score reporting #60 #63 #66 #67 #71.
- Sami explored processing quicker tables first, but David warned that this could further neglect tables already slowed by bloat; discussion of periodic reprioritization also exposed the risk of repeatedly reintroducing processed tables, so Nathan kept both ideas outside the initial change #67 #68 #76 #79 #83.
- Robert argued that a smarter order could still regress workloads that benefited from catalog order and proposed adjustable component weights; Sami supported retaining a route back to the old behavior, and Nathan agreed to implement weights #87 #90 #91 #99.
Mar 2026 · 9 entries
- After a gap of several months, Nathan demonstrated a contrived wraparound scenario in which scoring processed the oldest table first, then posted a rebased implementation with five weight settings and fixed a missing include #103 #104 #105.
- Sami supplied documentation and questioned very large failsafe scores; Nathan incorporated rewritten documentation and corrected the age types while retaining aggressive scaling because near-wraparound tables should overwhelmingly outrank ordinary maintenance #106 #107 #109 #111.
- Greg Burd supplied simulations and a full alternative patch that always placed forced vacuums ahead of routine work, but David argued that merely reaching freeze-max-age was not an emergency and instead proposed larger configurable weights with earlier scaling thresholds #116 #117.
- Nathan retained unified scoring, raised the weight limit to ten, and made weights above one lower the aggressive-scaling threshold; Greg accepted that compromise, and Sami's subsequent review produced comment and logging corrections #119 #120 #121 #124.
- Bharath Rupireddy requested more guidance, tests, and observability, while Nathan deferred per-table weight options and broad new additions because the initial policy needed field experience and a manageable scope #126 #127.
- Sami posted a separate priority-view patch, but Nathan and David chose to land scheduling first; David's request to expose individual score components nevertheless led Nathan to introduce AutoVacuumScores and component logging in the main patch #128 #129 #131 #132.
- Sami and Bharath approved the component-score version, while renewed suggestions for periodic reprioritization and automatic cost adjustments remained future work #133 #134 #135.
- David's final review prompted clearer weight documentation and an explicit bypass of sorting when all weights are zero, guaranteeing the advertised escape hatch without relying on sort implementation details; he then accepted the final patch #139 #140 #141.
- Nathan committed the final attached scheduling patch unchanged as d7965d65fc, and Sami moved the priority-view work to a separate discussion #140 #142 #143.
Aug - Sep 2026 · 3 entries
- Months later, Nathan reported a corner case noticed by Claude: sufficiently high freeze weights could make exponentiation reduce a score instead of increasing it, and he posted guards requiring the base score to exceed one #144.
- Bharath approved the fix and requested tests, while Greg checked the arithmetic and reproduced the corrected monotonic scores; Nathan committed the same logic with shorter explanatory text as 1bda0b23bc, including a backport to PostgreSQL 19 #145 #147 #148.
- The thread ends with both the scheduling feature and scaling fix committed, while shihao zhong's new TAP test remains proposed; it fabricates catalog ages to test both freeze components quickly and deterministically #149.