model-writtenrun 7521 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Petr Fedorov reported lost displayed microseconds when extracting epoch in #1; Tom Lane explained the float-output behavior in #2, while Thomas Munro identified the deeper SQL requirement for an exact numeric result in #3 and #6.
- Peter Eisentraut identified numeric's lack of infinity support as an obstacle in #7, and Tom proposed adding numeric infinities in #8.
- Vik Fearing proposed specialized extraction functions in #9, and the compatibility discussion in #10 and #11 established the need to retain existing date_part calls in stored views while introducing new functions for EXTRACT.
- Tom proposed preserving SQL spelling during deparsing by adding COERCE_SQL_SYNTAX to existing function-expression nodes in #13, avoiding a new expression-node hierarchy.
- After numeric infinity support became available, Peter posted the initial numeric-return patch in #15, raising questions about fractional scale, dedicated date handling, and undocumented Julian extraction behavior.
- Peter expanded the work into date tests, an integer-to-numeric helper, and the EXTRACT implementation in #17, reporting roughly 20 percent overhead in his initial performance tests.
- Tom reviewed the helper, fixed an int8_sum bug, and added missed conversions in #18, while reiterating that changing existing date_part return types would break stored-view compatibility.
- Peter committed Tom's revised helper attachment #18 unchanged as 0aa8f76408, then landed the unchanged date-test portion of #17 separately as 540612fa46.
- The remaining EXTRACT work stalled and Michael Paquier marked it returned with feedback in #24; Tom resumed the supporting deparser work with a separate patch in #25, deliberately leaving EXTRACT rendering until the new numeric functions existed.
- Tom completed function-OID coverage and tests in #26, then refined ordinary SUBSTRING and OVERLAY calls and excluded misleading SQL rendering of regex substring in #27.
- Tom landed the deparser patch from #27 as 40c24bfef9, additionally removing empty EXTRACT syntax and bumping the catalog version; renaming CoercionForm and replacing SQLValueFunction remained optional future work.
- Peter returned with the redesigned EXTRACT patch in #28, preserving date_part, adding parallel numeric-returning extract functions, and using the new deparser infrastructure to reproduce EXTRACT syntax.
- Tom questioned the duplicated implementations in #30, and Peter's retesting in #31 exposed expensive numeric division that initially appeared likely to postpone the work.
- Peter introduced fast power-of-ten division in #32; Tom proposed an exponent-based interface and safer overflow handling in #33, then independently benchmarked the patch in #35 and found most cases about 10 to 15 percent slower.
- Peter's final attachment #36 combined float8 and numeric implementations into shared functions, incorporated the helper changes and overflow fallbacks, and expanded edge-case tests, while leaving Julian extraction performance and semantics for follow-up.
- Peter landed #36 with only an added catalog-version bump as a2da77cdb4 and confirmed completion in #37, delivering numeric EXTRACT results while preserving legacy date_part behavior.
- Tom investigated Julian extraction in #38 and concluded that its local-midnight behavior matched PostgreSQL's existing convention, so he proposed documentation fixes instead of changing the calculation and attached them in #39.
- Tom committed the Julian documentation from #39 with introductory wording changes and an additional example as 79a5928ebc, backpatching the clarification to supported branches.
- Tom subsequently found that the astronomical conversion should use UTC+12 in #40; Thomas Munro verified that correction against independent examples in #41, and Tom fixed the documentation and outputs in c93f8f3b8d.
- Noah Misch identified inconsistent FuncCall.funcformat equality handling in #43; Tom established that two parser checks were redundant in #46 and committed their removal together with the equality-macro correction as a65e9f3f14.
- At the final message #47, Noah expressed no concerns about the pushed correction, all submitted work had landed, and the thread contained no outstanding replacement patch, although Julian performance optimization and optional parser cleanups had not been pursued.