model-writtenrun 8261 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed numeric infinities encoded in unused NaN bits, aiming for upgrade compatibility and POSIX arithmetic behavior, while flagging unresolved sorting, aggregate, and JSON-path interactions #1.
- Andrew Gierth reviewed the sorting implementation and recommended separating positive infinity from NaN in abbreviated keys, with a test mixing very large finite values, infinity, and NaN; Tom accepted those changes #3 #4 #8.
- Robert Haas raised application-compatibility concerns, leading Tom to propose allowing infinities only in unconstrained numeric columns while preserving existing NaN behavior, which Robert supported #5 #6 #7.
- A side discussion considered negative numeric scales: Tom concluded they could be implemented through typmods without changing stored values, but client compatibility remained a concern and no implementation entered this series #10 #11.
- Tom's next revision added documentation, incorporated the sorting recommendations and precision restriction, and rebased over nearby commits #12.
- Dean Rasheed argued that gcd and lcm with infinite inputs should return NaN, while Vik Fearing clarified that SQL permits omitted precision and scale; Tom acknowledged that PostgreSQL's unconstrained numeric behavior differs through its treatment of scale #13 #14 #15.
- Tom adopted Dean's gcd/lcm semantics and fixed missing precision enforcement in text and binary input by sharing special-value typmod checking #16.
- After a month, Dean's detailed review exposed incorrect infinite in_range() boundaries and infinity-to-pg_lsn conversion, and proposed helper and comment cleanups; he also noted an existing float-versus-numeric discrepancy for NaN divided by zero #17.
- Tom reported that the JSON-path concern had been resolved separately, agreed to fix floating-point in_range() behavior before adapting numeric, and left SQL isfinite()/isnan() functions as follow-up work #18.
- The final revision synchronized numeric range behavior with the corrected floating-point behavior and incorporated Dean's remaining code comments, conversion fixes, and additional window tests #19.
- After Dean's final acknowledgment, Tom committed the final patch unchanged as a57d312a77 and confirmed the push, completing this series; negative scales, SQL predicates, and the separately noted NaN-division discrepancy were outside the landed patch #20 #21.