model-writtenrun 2402 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2015 - Mar 2016 · 3 entries
- Teodor Sigaev opened the thread with native OR-expression support for index ScanKeys, aiming to avoid expensive BitmapOr scans while preserving index ordering and KNN opportunities; the prototype supported BRIN, GIN and GiST, with B-tree work deferred #1.
- David Rowley and Tomas Vondra exposed documentation, stack-safety and planner issues, leading Teodor to revise the implementation and split planner/executor infrastructure from index-method changes and tests #4 #7 #9.
- The original implementation did not land in this thread, and outstanding redundant-qualification and assertion problems remained in its final review round #10 #11 #13.
Dec 2022 - Jan 2023 · 1 entry
- Andrei Lepikhov revived the topic years later with OR-equality-to-IN conversion, and Alena Rybakina documented the severe memory and runtime costs of tens of thousands of ORM-generated predicates before posting a new parser-based OR-to-ANY implementation #14 #15 #17.
Jun - Dec 2023 · 9 entries
- Peter Geoghegan supported ScalarArrayOpExpr as a useful representation for index access, while Ranier Vilela supplied cleanup patches and test feedback that Alena incorporated selectively #19 #23 #22 #26.
- Alena also tried conversion during index planning, but selectivity and expression-placement problems kept the parser implementation as the main development branch #28 #32.
- Andrei supplied successive corrections for constant-side detection, operator checks, expression ownership and tests, and Alena integrated them and explicitly credited him as coauthor #36 #37 #39 #40 #47.
- Prepared-query support and performance experiments led to further debate over the transformation threshold and whether conversion belonged in parser processing or planner normalization #56 #59 #60.
- Alena explored conversion before and after selectivity calculation, packaged those alternatives alongside the parser version, and fixed RestrictInfo-related crashes without resolving all nested-OR cases #62 #63 #70 #74 #75.
- Alexander Korotkov demonstrated quadratic grouping costs and lost partial-index plans, prompting a hash-based parser implementation while preserving the need to recover bitmap alternatives #76 #78.
- Robert Haas argued that a configuration switch could not substitute for sound planning choices, while Peter supported a restricted normalization approach; the disagreement concerned both transformation placement and protection against regressions #79 #81 #82 #87.
- Andrei replaced fragile nodeToString hashing with expression jumbling and collision checks, then adopted standard hash callbacks; Robert's subsequent review drove corrections to operator identity and type handling #95 #99 #102 #103 #110.
- Alena experimented with reversing ANY back into OR to recover partial-index paths and compared that with later conversion, while Andrei continued the parser branch and fixed its regression expectations #108 #113 #116 #117 #118.
Jan - Apr 2024 · 6 entries
- Jian He found volatility and type issues, and the next development phase added a separate SAOP-decomposition patch that could recover partial-index bitmap plans #120 #122 #125 #126 #127.
- Documentation, hash combining, clause-order preservation and duplicate handling were refined, but review continued to question array limits and the consequences of modifying expressions during parsing #151 #159 #160 #166.
- Korotkov explained that parser transformation made stored definitions depend on a runtime setting, so Andrei moved conversion into prepqual.c and then restored planning-time partition pruning by evaluating the resulting SAOP expressions #167 #170 #172.
- The companion bitmap patch expanded into general SAOP-to-OR recovery, but its potential planning cost led to a threshold-controlled main transformation and continued disagreement over how much reverse exploration to perform #173 #174 #175.
- Korotkov committed the first component of the series as 72bd38cc99, with threshold, documentation and test changes, while leaving out reverse conversion because it could slow planning and undo the intended simplification #175 #176.
- Justin Pryzby immediately noticed that the sample configuration retained the wrong threshold default, prompting the discussion-derived correction 2af75e1174 #177.
Jun 2024 - Feb 2025 · 15 entries
- The first implementation was subsequently reverted from PostgreSQL 17, and Korotkov restarted it for PostgreSQL 18 with conversion during base RestrictInfo insertion, no GUC, no expression-jumbling hash and support for Params, citing Tom Lane's design feedback #178.
- Andrei warned about mutating RestrictInfo metadata and supplied a fix, Alena demonstrated a remaining planning regression, and Robert pressed for constructing transformed clauses while generating index paths #179 #180 #185 #188 #199.
- Nikolay Shaplov's reviews improved naming, factoring, tests and explanations while Alena corrected mixed BoolExpr and RestrictInfo handling in the intermediate redesign #196 #198 #202 #204.
- Korotkov then moved conversion directly into match_clause_to_indexcol(), constructing arrays without general expression evaluation and abandoning quadratic whole-list grouping; Alena's alternative of grouping before the index-column loop was not adopted #206 #213 #214.
- A separate bitmap patch restored broader benefits by grouping similar OR arguments while considering both the grouped form and individual paths, establishing the two-component design that eventually landed #216 #217.
- Unusual-operator and type tests exposed an operator-family bug, and subsequent review added collation checks, preserved original OR clauses for rechecks and corrected clause/orclause confusion #220 #222 #223 #224.
- The developers deliberately limited bitmap exploration to whole groups or individual arguments rather than arbitrary subgroups, and replaced direct RestrictInfo copying with constructors that initialize its metadata correctly #226 #232 #239.
- Peter's mixed-type examples and discussion with Robert and Tom reinforced the decision to preserve operator semantics and defer broader coercion support rather than extend the patch's scope #245 #250 #263 #264 #267.
- Andrei supplied explanatory comment patches, and Korotkov added bitmap-path improvements plus an exploratory join extension, then deferred the join component until the main pair had landed #272 #274 #277.
- Alena proposed storing the transformed qualification for rechecks, but the discussion retained the original-clause representation; the final main submission consequently followed the existing two-patch branch with light cleanup #280 #281 #282 #283.
- The final pair became d4378c0005 for index matching and ae4569161a for bitmap grouping, each taking its corresponding component with only a comment correction after the final review #283 #284.
- Richard Guo then reported that a valid zero constant violated an assertion, leading directly to assertion removal and a regression test in d4d11940df without an attached patch submission #287 #288.
- Alexander Lakhin's hash-index failure exposed missing SAOP-capability checks; Alena supplied the fix, Ranier moved the guard earlier, and Korotkov added the corresponding grouping check before committing 5bba0546ee from the final revision #290 #291 #293 #295.
- Korotkov resumed the deferred join extension using relation-relative constant detection, and Andrei found duplicate join qualifications and a wrong-RestrictInfo crash that were corrected in the next full revision #299 #303 #304 #305.
- Reviews by Pavel, Alena and Andrei refined join tests and explanations, while the discussion accepted that broader path choices could change short-circuit behavior; the final join component landed as 627d63419e #308 #310 #311 #315 #317.
Mar 2025 · 3 entries
- Andrei next demonstrated unnecessary OR-clause permutations and proposed preserving the original list when no groups formed, but Korotkov replaced that approach with group-position tracking that also minimizes movement when grouping succeeds #320 #323.
- Pavel, Andrei and Alena reviewed the ordering rework, and 775a06d44c landed it with corrected group-key conjunctions and additional tests #324 #327 #329.
- At the thread's end, the supplied commits show the index, bitmap, join and follow-up fixes landed, with Alena's final concern resolved; broader mixed-type conversion and arbitrary bitmap subgroup exploration remained deferred enhancements #329 #263 #226.