model-writtenrun 7628 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Pavel Borisov proposed SP-GiST INCLUDE columns for index-only scans in #1, initially preserving existing key placement, appending included attributes with custom alignment, and using spare nextOffset bits for null and variable-width flags.
- Andrey Borodin questioned enforcement of SP-GiST's single-key restriction in #2, prompting explicit error handling in #3; Pavel then refined formatting, documentation, and attribute flags through #8.
- Andrey challenged the tuple-header flags and suggested considering a different covering-index header in #9; Pavel defended the compact layout and uniform access macros, documenting that choice in #10.
- Anastasia Lubennikova tested and reviewed the patch in #11, leading to clearer names and comments, stronger scan-plan tests, and VACUUM ANALYZE additions in #12, including a separate change for the existing index_including test.
- Andrey's further reviews in #13 and #15 led to tuple-pointer access macros, named key offsets, the t_info field name, and centralized column-count checks by #16, when the auxiliary VACUUM patch was also resent.
- After the autumn gap, Tom Lane rejected the custom attribute-alignment code and unsafe tuple-descriptor caching in #17; Pavel explored alternatives in #18 and #19, and Tom endorsed putting the bitmap after the header in #20.
- Pavel implemented that redesign in #21, reusing heap_fill_tuple, sharing header-independent index deformation, and embedding the descriptor in the cache allocation; #22 fixed MSVC compatibility, and #24 restored an omitted test file.
- On returning to review in March, Tom identified incompatibility between standard tuple storage and SP-GiST's full-Datum pass-by-value convention in #26; he chose normal attribute storage in #27 because the potential compatibility break was limited to unlikely existing opclasses on big-endian machines.
- Pavel incorporated fetch_att, smaller picksplit work arrays, the restored SpGistState API, and formatting cleanup in #28, then restored the missing regression files in #29.
- Tom committed the pass-by-value prerequisite separately as 1ebdec8c03, extracting it from #29 and adding dedicated leaf storage helpers and null-key guards while preserving inner-tuple storage.
- Tom then committed INCLUDE support from #29 as 09c1c6ab4b, with corrected input-versus-leaf datatype handling, revised descriptor management, and focused KNN and operator-class tests.
- In #30, Tom explained that he dropped index_including_spgist.sql because its GiST-derived cases added no measured coverage; the separate VACUUM-only attachment from #16 is absent from both candidate commits and was not carried in the final submission.
- The feature was committed by the thread's end, but additional test coverage remained open: Pavel supplied a larger radix-tree test in #31 and asked whether to adapt it for regression testing.