model-writtenrun 8604 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - May 2017 · 4 entries
- Alexander Korotkov revived the former Partial Sort work as Incremental Sort, describing prefix-aware sorting with separate path and plan nodes, tuple-slot storage and an enable GUC in #1.
- Robert Haas’s review of naming, executor separation and foreign-table tests in #2 led Korotkov to create a dedicated executor node and use full Incremental names in #3.
- Heikki Linnakangas exposed slow performance for tiny groups in #6; initial costing penalties in #9 gave way to minimum-size batches in #18 and a single saved pivot tuple per group in #25.
- Peter Geoghegan questioned perfectly presorted benchmarks in #21, Robert raised group-skew concerns in #26, and Korotkov returned after a long rebase with broader measurements showing substantial wins but remaining skew and text-comparison regressions in #28.
Sep 2017 - Jan 2018 · 2 entries
- Antonin Houska’s costing and test review #35 and Thomas Munro’s comparison questions #36 tightened the implementation, while Korotkov deferred minmax and broader planner enhancements in #37.
- After further maintenance, Tels identified node-deserialization and wording defects in #48, fixed in #49; Tomas Vondra then found a tuplesort memory-lifetime crash in #53, addressed by #54.
Mar - Apr 2018 · 4 entries
- Alexander Kuzmenkov supplied a cleaned-up version in #73, and Tomas demonstrated in #74 that automatically preferring incremental sort could discard a cheaper ordinary sort.
- Korotkov first restricted cost-based selection to create_sort_path in #78, then Kuzmenkov changed the planner to generate distinct competing paths in create_ordered_paths in #80.
- The next revisions separated plan builders, restored abbreviated-key support and bounded minimum batches in #81, then recovered useful partially ordered non-cheapest inputs in #89.
- Tom Lane opposed a last-minute PostgreSQL 11 landing because group-size estimates could still cause regressions in #86 and #90, so the feature was deferred rather than landed and reverted.
May - Jun 2018 · 1 entry
- James Coleman took over maintenance with #95, supplied a practical pagination use case in #97, and revived the patch after another long gap in #105.
May - Sep 2019 · 6 entries
- Coleman found that missing ExecSetTupleBound support prevented top-N heapsort and bounded batching in #112, then proposed a hybrid executor to improve behavior for both small and large groups in #113.
- Tomas’s parallel prototype #128 placed incremental sort below Gather Merge and exposed double discounting of LIMIT costs; Coleman’s major rewrite #129 introduced separate full-key and prefix-only sort states with a heuristic mode switch.
- Testing exposed attempts to sort aggregate expressions below aggregation in #134 and #135; Tomas introduced relation-appropriate useful-pathkey discovery in #143, while Coleman independently explored caller-owned sorting in #144.
- Tomas combined those planner ideas in generate_useful_gather_paths and added a temporary forcing GUC in #145, while discussion in #148 and #149 favored independent enable controls and postponing invasive nonparallel merge-join support.
- Tomas discovered that add_partial_path discarded useful low-startup-cost paths in #154; Coleman folded core fixes into #155 and later carried the startup-cost prerequisite separately while temporarily omitting the broader planner extension in #175.
- Benchmark-driven review narrowed planner expansion to demonstrably useful query shapes in #168, and Tomas’s December rebase #179 reported six useful insertion sites across 4,960 tested combinations.
Jan - Mar 2020 · 7 entries
- March development resumed with conditional instrumentation and richer EXPLAIN statistics in #192; #193 corrected instrumentation to record each completed batch once, and Tomas restored separately reviewable planner extensions in #194.
- Alvaro Herrera’s cleanup patches #195 and Justin Pryzby’s review #196 were folded into the series, while Tom Lane recommended masking unstable memory figures in #199 and Andres Freund proposed reproducible EXPLAIN output in #200.
- Tomas traced memory leaks and assertions to a dangling tuplecontext and supplied #215; Coleman introduced shared batch initialization in #219 and reversed a cleanup that skipped clearing the final tuple slot, diagnosed in #217.
- Andreas Karlsson tested the revised series and identified a const warning in #221, then Alvaro clarified rescan requirements in #231; Coleman implemented complete rescan behavior and tests in #236.
- Coleman refined sort-state reuse, fixed JSON resource reporting, masked unstable EXPLAIN values and shared an equivalence-class helper in #237, before consolidating the series in #239.
- The attempt to replace useful-pathkey discovery with truncate_useless_pathkeys in #241 was rejected in #244; the revised series #247 retained relation-aware discovery and addressed shared-memory instrumentation and worker reporting.
- Tomas measured planner overhead in #255, prompting pointer-identity fast paths, single-key filters and list reuse through #261, #277 and #278, with final benchmark differences near noise by #286.
Apr - Jul 2020 · 11 entries
- Tomas committed the core implementation from #286 as d2d8a229bc, adding parallel regression coverage while keeping the prerequisite and additional planner paths separate, as announced in #288.
- Immediate force_parallel_mode failures led Tomas to diagnose a premature EXPLAIN return in #294 and commit 7d6d82a524; Coleman’s independently matching attachment #297 arrived after that commit.
- Tom reported worker-count-dependent expectations in #300, and Tomas fixed the regression setup with 23ba3b5ee2, announced in #301, although Tom later preferred masking the count in #303.
- Unlaunched workers exposed reliance on a zero sort-method sentinel: Tom rejected the NULL-check proposal in #309, explained the invariant in #312, and committed the enum and loop-bound repair c7654f6a37.
- Tomas landed the additional-path component of #286 with only comment and test-output adjustments as ba3e76cc57; his follow-up #328 deferred startup-cost handling to PostgreSQL 14 because precheck and parameterization consistency needed more work.
- James’s group-invariant comments from #309 and Justin’s Maximum-to-Peak naming corrections from #320 were combined with regression updates in d22782a539.
- CLOBBER_CACHE_ALWAYS failures exposed autoanalyze-dependent plans in #329 and #331; Tomas enlarged test tables in #334, James verified transition coverage and requested ANALYZE t in #335, and cea09246e5 stabilized the tests.
- Post-landing scrutiny of operator-class function permissions and hooks in #342, #343 and #344 found no additional checks necessary.
- Justin resent remaining comment fixes and proposed consistent EXPLAIN spacing, colons and separate group lines in #345; James consolidated the patches, simplified indentation and standardized Pre-sorted naming in #346.
- Tomas committed the comment subset as 1a40d37a9f and the EXPLAIN-format subset as 6a918c3ac8, announcing both in #352.
- The thread remained open over machine-readable resource statistics: Justin wanted zero-valued memory and disk fields in #354, James eventually supported that in #356, and the final message #359 proposed negative values or JSON null for unknown memory usage without a resolution recorded here.