model-writtenrun 1738 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul 2018 · 2 entries
- Oliver Ford proposed RESPECT/IGNORE NULLS and FROM FIRST/LAST in #1, using special boolean wrapper types and overloaded window functions so calls with different options could share a window definition.
- David Fetter refreshed catalog OIDs in #2, and Krasiyan Andreev reported passing tests and successful production use in #3.
Sep 2018 · 1 entry
- Andrew Gierth rejected the special-type and precedence approach in #4; discussion with Tom Lane explored keyword reservation, lexical lookahead, and generic window APIs, and Andrew returned the patch with feedback in #13, explaining that reversing sort order could violate peer-row ordering requirements.
Apr 2020 · 1 entry
- After Stephen Frost revived interest in #15, Krasiyan Andreev supplied #16, dropping FROM FIRST/LAST syntax in favor of negative nth_value positions while retaining the special ignorenulls type.
Apr - May 2023 · 3 entries
- Tatsuo Ishii restarted implementation in #17 with a proof of concept that carried null-treatment state through parser and executor nodes without changing function signatures, initially supporting only nth_value and leaving FROM FIRST/LAST out.
- Vik Fearing questioned mark handling and requested extensibility in #19, while Tom Lane supported keeping the new keywords unreserved in #20; Tatsuo explained the mark-position failure and proposed function-level checks in #22.
- Oliver Ford moved null skipping into the window API and added cached non-null positions and tests in #23; Tatsuo supplied an explicit rejection delta in #25, leaving the treatment of unsupported and user-defined functions under discussion.
Sep 2024 · 2 entries
- Oliver revived the generic implementation with documentation in #28, and Vik's aggregate counterexample in #29 led to distinct omitted, RESPECT, and IGNORE states in #30.
- Tatsuo's review in #31 and David G. Johnston's semantic clarification in #32 led to built-in-function restrictions in #34, but Tom Lane objected to centralized restrictions in #36 and #38.
Jan - Mar 2025 · 4 entries
- The design settled on per-function opt-in or rejection through WinCheckAndInitializeNullTreatment in #43, following the API proposal in #41, so existing extensions could continue operating without adopting the option.
- Tatsuo found that cached positions lived in a per-tuple memory context and survived incorrectly across partitions in #45 and #47, with Krasiyan reproducing failures in #46; Oliver corrected allocation and reset handling and added partition tests in #48 and #50.
- Exclusion-frame counterexamples in #54 caused Oliver to bypass the position cache for exclusions in #55, and the missing mark update identified in #56 was added in #57.
- Tatsuo's benchmark in #58 exposed poor scaling, so Oliver removed the position cache in #59; header and rebase fixes followed through #62, which Krasiyan reported testing successfully in #63.
Jun - Oct 2025 · 15 entries
- Tatsuo took over optimization work in #65, introducing a compact two-bit null-status cache, then extended it to partition searches and eliminated duplicate search code in #66.
- Further work avoided unnecessary frame tuple fetches in #67, fixed negative-position handling in #68, and moved conditional fetching into row_is_in_frame in #71; Krasiyan reported successful testing of the preceding version in #69.
- Tatsuo explored broader regression coverage by inserting IGNORE NULLS into existing tests in #72, but the attachment was an output-comparison experiment and its proposed integration remained unanswered.
- Tatsuo split the feature into six review patches in #73 and rebased the documentation in #74, which Oliver tested and endorsed in #75.
- Chao Li's review in #78 led to small code improvements and expanded rejection tests in #80; Oliver supplied a separate uint8/uint16 node-generator fix in #81, which Tatsuo explicitly deferred when resending the feature in #83.
- The feature landed as 25a30bbd42 from #83, crediting Oliver Ford and Tatsuo Ishii, with light comment editing and restoration of omitted syntax documentation and SQL-feature declarations confirmed in #85.
- Tom Lane reported unchecked tuple fetching and nullable isout dereferences in #86, Álvaro Herrera reported an uninitialized-datum warning in #87, and Tom requested a clearer partition loop in #89.
- Tatsuo withdrew the initial guard-only repair #90 after finding that it skipped required marks, then supplied local output-state handling and a simpler loop in #94; Paul Ramsey confirmed that it restored PostGIS operation in #95, and it landed unchanged as 2273fa32bc.
- In #98, Tom Lane disputed his original reviewer credit and, after inspecting the landed code, identified user-facing error-reporting problems, unclear extension opt-in semantics, and an incorrect assumption that only one function argument needed a null cache.
- Tatsuo proposed separate fixes for per-argument caches in #99 and error reporting in #100, and explored automatic acceptance of null treatment in #103; that last proposal lacked documentation changes and has no demonstrated landing in this record.
- Chao Li's const suggestion in #101 was incorporated alongside an assertion and null-result guard in #107, which landed unchanged as 5f3808646f.
- Michael Paquier reported an overwritten helper result and inadequate comments in #111; Tatsuo's explicit-discard cleanup #112 landed unchanged as dd766a441d, while Chao questioned the repeated call in #113 and Tatsuo defended retaining it for readability in #115.
- The multi-argument repair gained conservative mark handling in #105 and lazy per-argument allocation, independent capacities, and query-lifetime memory in #116, which landed unchanged as 2d7b247cb4.
- Chao Li identified an awkward else block in #117, and David Rowley explained the applicable brace convention in #120; Tatsuo's correction #121 landed unchanged as 20628b62e4.
- The thread ends with the final style-fix push confirmed in #122: null treatment and the five identified follow-up fixes have landed, while FROM FIRST/LAST, the deferred generator support from #81, and the automatic-acceptance proposal from #103 remain outside those landings.