model-writtenrun 8823 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug 2018 · 1 entry
- Masahiko Sawada opened this thread's patch history with a worker-state-machine design that parallelized heap scanning and cleanup at block level, processed separate indexes concurrently, and supported autovacuum, but depended on resolving relation-extension locking #1.
Oct 2018 - Apr 2019 · 11 entries
- After an alternative extension-lock approach hurt partitioned-table bulk loading, Sawada returned to a simpler group-lock conflict prerequisite and rebased the implementation #2 #3.
- Amit Kapila proposed separating independent-index parallelism, parallelism within one index, and heap parallelism into successive projects; Sawada accepted the narrower first step to avoid the extension-lock obstacle #7 #8.
- Sawada's replacement implementation launched workers for individual index phases while the leader retained heap work, shared dead tuples through DSM, and updated catalog statistics after parallel mode; it dropped autovacuum support and added a vacuumdb companion #10.
- Haribabu Kommi's review improved option representation, verbose worker reporting, statistics handling, and tests, while Robert Haas questioned redundant state and reuse of the parallel-scan reloption; Sawada removed that reloption and split the options prerequisite #14 #18 #31 #35.
- Robert replaced the separate VacuumOptions structure with fields in VacuumParams and proposed a DefElem parse representation, which Sawada implemented alongside a corrected forwarding of Robert's patch #40 #41.
- Robert committed the VacuumParams component from #41 with a local vacuum_open_relation rewrite as f41551f61f, then committed its grammar companion as 6776142a07, leaving parallel execution for further review #44 #45.
- Kyotaro Horiguchi explored overlapping heap and index processing, but discussion favored retaining full dead-tuple batches because dividing finite tuple storage could require additional index passes #43 #51 #57.
- The new parser enabled Boolean arguments for existing VACUUM options; Haribabu corrected misplaced documentation, and Robert required Boolean locals instead of progressively OR-ing flags so false values would work correctly #59 #60 #65.
- The corrected Boolean prerequisite in #66 landed as 41b54ba78e, after which Sawada resent only parallel vacuum and vacuumdb #67 #69.
- Further review consolidated leader and worker helpers, improved translation and statistics handling, and led to a June update; after another gap, Amit resumed intensive review of memory budgeting and worker behavior #63 #75 #77 #79 #82.
- The completed feature covers parallel processing of separate indexes and its vacuumdb interface; heap-scan parallelism, parallel autovacuum, FAST, and the leader-disable facility remained outside the landed scope #10 #335 #382 #408.
Sep - Oct 2019 · 4 entries
- Amit identified that GIN could consume maintenance_work_mem independently of the shared dead-tuple array, opening the related memory-budget discussion thread #41328 and forcing the design to account for index-AM memory usage #85 #94.
- Mahendra Singh Thalor's testing exposed invalid frees of copied statistics; Amit supplied individual allocations, Sawada incorporated them, and subsequent testing confirmed the correction #105 #106 #107 #108.
- Amit then found that GiST's extended statistics did not fit fixed IndexBulkDeleteResult slots, leading first to an AM capability flag excluding GiST and then to an estimator callback, memory-use metadata, and variable-length shared statistics #109 #118 #141.
- Independent worker cost limits could exceed the intended aggregate I/O budget, so Dilip Kumar prototyped divided limits while Sawada tried shared atomic accounting; both added instrumentation and compared equal-sized and uneven index workloads #119 #132 #141 #143 #150.
Nov 2019 · 5 entries
- Mahendra supplied force-parallel testing deltas and reported check-world success, revising the test mechanism to exercise a single worker without leader participation #167 #174.
- Review established that index AMs needed different behavior for bulk deletion, unconditional cleanup, and cleanup only when bulk deletion had not occurred; the resulting series integrated per-phase flags, memory budgeting, and the cost-delay work developed in thread #41476 #173 #211 #214.
- Dilip requested that the leader process unsupported indexes while workers were active and that temporary tables fall back to serial vacuum, while benchmarks motivated excluding tiny indexes whose worker overhead outweighed useful work #198 #200 #186 #203.
- Amit supplied successive API and execution review deltas, including fixes for shared-statistics indexing and handoff; Sawada incorporated them, introduced phase-specific worker counts, and restored accidentally omitted companion patches together with Dilip's GiST prerequisite #229 #230 #233 #234 #238.
- Sergei Kornilov and Robert favored having the leader process nonparallel indexes before joining the shared work pool; Robert also rejected an unused estimator API and preferred worker-count changes during parallel-context reinitialization #246 #269 #268 #275.
Dec 2019 · 4 entries
- Amit's next review fixed cost-accounting retries and a race caused by initializing shared counters after workers launched, and extracted clearer delay and index-processing helpers #259.
- The resulting rework limited parallel work by min_parallel_index_scan_size, simplified the statistics contract, separated leader-disable testing, and carried FAST as an independent proposal; follow-up revisions corrected worker counts and index eligibility #285 #293 #301.
- With separate GiST work removing the immediate need for private shared statistics, Amit and Sawada deferred the generalized estimator interface and retained a no-parallel capability for unsupported AMs, while Amit began preparing the remaining API and execution patches for commit #298 #300 #310.
- Tomas Vondra reviewed memory sizing and the AM interface and argued that manual vacuum should use parallelism by default, with cost throttling and the index-size threshold controlling resource use; Mahendra supplied that behavior and Amit integrated it #318 #320 #333 #334.
Jan 2020 · 9 entries
- FAST lacked consensus, and the leader-disable facility remained a development aid, so Amit and Sawada excluded both to focus on the main feature #334 #335.
- Sergei found that a temporary table could disable parallelism for subsequent tables; Mahendra supplied a fix-up, and Amit adopted per-table handling instead of mutating shared command options #341 #342 #345.
- Default parallelism made a bare PARALLEL option redundant, so the final syntax required an explicit degree, with zero disabling parallel workers #348 #357.
- The IndexAM component of #357 landed unchanged as 4d8a8d0c73; Sawada's cost-state initialization and PARALLEL 0 tests were folded into the remaining execution patch #360 #362.
- Final testing covered functional and error paths, and Dilip verified that aggregate cost delays tracked serial vacuum while supplying worker-count and pointer-initialization corrections #371 #375 #378.
- Amit committed the final execution patch from #379 with comment and typedef cleanup as 40d964ec99, completing parallel index vacuum for manual VACUUM #380 #383.
- Andres Freund immediately reported and diagnosed oversized dead-tuple allocations because the capacity calculation omitted the structure header; Sawada supplied the fix, Amit revised and tested it, and Dilip corrected macro parentheses before it landed as 79a3efb84d #385 #387 #390 #391 #393.
- Sawada then resumed the remaining vacuumdb interface with zero-degree and PostgreSQL 13 handling; Mahendra tested older-server compatibility and combined connection-level and worker-level parallelism, while Amit added validation and documentation refinements #395 #397 #398.
- The final vacuumdb attachment #405 landed unchanged as 47bc9ced0d, and Amit marked the work committed at the end of the thread #407 #408.