model-writtenrun 8571 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2020 · 10 entries
- Kirill Bychik proposed per-statement WAL record and byte counters using executor instrumentation, with pg_stat_statements as the first consumer and EXPLAIN as a possible future extension #1.
- Thomas Munro found unstable exact regression results, and reviews by Julien Rouhaud and Michael Paquier led to more robust checks, a corrected parallel DSM key, and an additional full-page-image counter #3 #6 #7 #12.
- Julien extended Kirill's patches with vacuum and autovacuum totals in pg_stat_database, but Fujii Masao preferred per-operation information, leading to replacement of that collector implementation with verbose vacuum and autovacuum output #13 #20 #21.
- Fujii identified full-page-image retry overcounting, an incorrect shared-memory pointer, inadequate byte-counter types, and pg_stat_statements naming and SQL issues; Kirill subsequently handed ongoing development to Julien because he lacked time #23 #24 #26 #27.
- Amit Kapila questioned coverage of parallel maintenance, and Julien confirmed that existing buffer statistics already omitted worker activity during parallel VACUUM and CREATE INDEX #28 #30.
- Julien explained that even read-only parallel queries can generate WAL through hint-bit updates, preserving parallel-query instrumentation and motivating a separate EXPLAIN and auto_explain consumer in the expanded series #35 #38 #40.
- Masahiko Sawada supplied separate parallel-maintenance buffer fixes using worker DSM arrays and leader accumulation, while deferring a common maintenance-worker framework as a larger project #42.
- Amit showed that the proposed FPI calculation mishandled forced images and records containing multiple images, so Julien changed it to count images during record assembly and accumulate them independently of the total record count #45 #46 #57.
- Review and testing corrected requested-versus-launched worker counts and stale vacuum statistics when a later phase launched no workers; Dilip Kumar verified multi-pass vacuum, and Amit polished the resulting fix #47 #48 #58 #64 #65.
- The completed thread left database-level vacuum WAL aggregates replaced by per-operation reporting, while manual VACUUM VERBOSE enhancements and EXPLAIN planning-phase WAL remained deferred proposals rather than unfinished parts of the landed series #21 #102 #109.
Apr 2020 · 16 entries
- Julien bundled the maintenance fixes ahead of the WAL features and extended their accounting to WAL, while subsequent review simplified regression tests and retained numeric WAL-byte output to avoid signed-overflow presentation #67 #71 #72.
- The parallel-vacuum buffer component last bundled in #72 landed unchanged as 3a5e22138a, independently of the remaining WAL and index-buffer work.
- Questions about extra parallel-index buffer accesses postponed that prerequisite; removing it also temporarily removed index WAL tracking, which Amit requested restoring independently, producing the revised infrastructure in #82 #78 #80 #81.
- Dilip investigated apparently inconsistent WAL totals and traced them to real differences in catalog full-page-image hole sizes and index page splits, resolving concern that the new instrumentation itself was inaccurate #86 #92 #95 #96.
- Amit prepared the four-part series in #97, and its unchanged infrastructure component landed as df3b181499, providing counters for execution and parallel maintenance #98.
- The pg_stat_statements component of #97 landed unchanged as 6b466bf5f2, exposing record, full-page-write, and byte totals while sharing the extension version upgrade already underway #104.
- Amit and Julien deferred EXPLAIN reporting of planning-phase WAL pending a stronger use case, and the execution-phase EXPLAIN and auto_explain component of #97 landed after rebase as 33e05f89c5 #102 #104.
- Masahiko resumed the outstanding index-buffer fix against the landed WAL APIs, then completed missing worker-side reporting and demonstrated that it restored the heap-scan buffer totals for parallel index creation #105 #120.
- Amit asked to consider manual VACUUM VERBOSE buffer and WAL reporting together in a separate proposal, so Julien reduced the remaining consumer to autovacuum logging in #111, which landed with only whitespace adjustments as b7ce6de93b #109.
- After the feature commits, Euler Taveira questioned the inconsistent wal_num_fpw name and Peter Eisentraut objected to double-spaced EXPLAIN fields, opening a separate cleanup phase #112 #115.
- Peter urged consistency with existing EXPLAIN formats, Julien proposed abbreviating the full-page-write field, and Amit favored matching buffer-statistics output #124 #126 #128.
- Masahiko traced small extra parallel-index buffer counts to each worker loading sort-support catalog entries; Amit accepted that overhead, and testing covered PostgreSQL 11 and 12 backport variants #125 #129 #130 #133.
- The completed index-buffer patch from #120 landed unchanged as 5c71362174, with the stable-branch variants representing backports of the same fix.
- Julien consolidated Euler's naming suggestion and Justin Pryzby's documentation work into a cosmetic patch, explicitly leaving disputed EXPLAIN formatting aside; Amit standardized its wording, and #141 landed unchanged as ef08ca113f #137 #143.
- Peter endorsed single spaces and an abbreviated field, Julien supplied a separate formatting patch, and Justin's review led to acronym documentation and discussion of non-text label capitalization #146 #148 #149 #150.
- Kyotaro Horiguchi and Justin challenged FPW terminology for a count of full-page images, while Amit supplied capitalized non-text labels and Peter and Michael Paquier supported the more precise FPI terminology #151 #153 #155 #159 #162.
May 2020 · 2 entries
- Julien expanded the formatting patch to rename fields, variables, documentation, and output consistently to FPI across all consumers; Amit restored the remaining capitalization changes and applied pgindent in #169 #167.
- The final cleanup in #169 landed unchanged as 69bfaf2e1d, and Amit announced it and updated the open-items page, leaving no pending attached patch in this thread #171 #173.