model-writtenrun 8682 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2019 · 10 entries
- Jim Finnerty opened the thread proposing faster NOT IN execution, while Tom Lane favored semantically safe anti-join conversion and treating costing defects separately; they subsequently agreed to target PostgreSQL 13 rather than the closing PostgreSQL 12 cycle #1 #4 #15 #18.
- David Rowley revived his older planner-only patch and then repaired its outer-expression nullability proofs, using constraints and appropriately scoped strict predicates to allow conversion only when both sides cannot contain NULLs #6 #7.
- Zheng Li posted an independent implementation with Jim Finnerty that handled nullable inputs by adding NULL tests to anti-join predicates and running outer-join reduction early #8.
- Rowley's testing showed that Li's rewrite incorrectly discarded outer NULLs when the inner relation was empty, and Richard Guo independently identified that correctness defect and the loss of hash and merge join choices caused by OR predicates #10 #21.
- Richard Guo described an executor-level NOT IN join alternative, but Li deferred that direction after timing and patent concerns were raised, instead proposing collaboration on the planner implementations #11 #13 #14.
- Andres Freund requested permanent coverage of the repeatedly rediscovered empty-inner case; Rowley supplied a standalone SQL test, and Lane committed it as 3396138a6d, without either optimization patch #23 #25 #26.
- Lane challenged claims that nullable anti-join rewrites always improved performance, and Rowley demonstrated an approximately eightyfold slowdown from forcing a nested loop; Finnerty agreed that existing hashed subplans should be retained when the inner result fits in memory #28 #31 #33.
- Lane identified another correctness requirement: non-null inputs alone do not guarantee non-null operator results, so Rowley added equality-opfamily checks, first for btree and then for hash operators #35 #36 #38.
- The competing implementations and duplicate commitfest entries caused review and CI confusion, so Rowley moved his non-null-only approach to a separate thread #44 #47 #49.
- Li substantially revised his implementation to preserve hashable subplans, test inner-result existence before filtering outer NULLs, handle nullable multi-column comparisons, and incorporate Rowley's strict-predicate nullability analysis #50.
Jun 2019 · 1 entry
- After a three-month gap, Li rebased the patch, added cached planner state to SubLink to avoid planning rejected conversions twice, and fixed access to upper-level CTEs while expanding CTE regression coverage #52 #53 #54.
Dec 2019 - Jan 2020 · 2 entries
- Following another long gap and a build failure caused by list-API changes, Li refreshed the patch; the attachment also introduced an enable_not_in_transform setting and reorganized conversion eligibility checks #55 #56.
- Andrei Lepikhov's testing found that correlated benchmark queries remained unimproved; Li explained that the existing ANY-sublink conversion rejects parent-query references, leaving broader correlated-subquery support unresolved #57 #58 #59.
Mar - Apr 2020 · 3 entries
- Lane supplied a conflict-only rebase but objected to temporary planner fields in SubLink, premature and repeated outer-join reduction, excessive regression-test runtime, and the undocumented planner setting #60.
- Li accepted that early outer-join reduction could likely be removed but defended early planning to preserve hashed subplans; Lane explained why copying SubLink could lose or invalidate its cached planner state, leaving the planning strategy unsettled #61 #62 #63.
- Lepikhov continued reviewing nullability logic, naming, and comments, but no further patch followed; David Steele marked the optimization Returned with Feedback pending implementation consensus, leaving only the standalone regression test committed #64 #65.