model-writtenrun 6160 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andrei Lepikhov encountered the 65,000-entry rangetable limit with heavily partitioned tables and proposed moving the special varno constants near INT_MAX in #1.
- David Rowley argued that allocation limits would intervene before INT_MAX and suggested one million in #2; Amit Langote and Julien Rouhaud expressed support in #3 and #4, and Andrei supplied that revision in #6.
- Tom Lane favored Peter Eisentraut's earlier idea of small negative special values in #5, while Tomas Vondra warned that larger positive constants could inflate bitmapsets in #7; Tom questioned whether those values belonged in bitmapsets at all in #8.
- Tom posted a negative-sentinel proof of concept that passed check-world in #11, changing Var.varno to int and removing the explicit rangetable limit; Peter reviewed the type choices in #12, and Tom argued that varnosyn should remain an unsigned original rangetable reference in #13.
- Tom preferred introducing the change early in the PostgreSQL 15 cycle rather than rushing it in #15; Andrei then extended the patch to ROWID_VAR and guarded two planner operations that failed with negative values in #17.
- After the gap, Tom supplied a broader type audit in #18, covering executor parameters, Var constructors and deparser variables, while concluding that CurrentOfExpr.cvarno never holds special values and could remain unsigned.
- David raised extension compatibility and runtime-cost concerns in #19; Tom defended the negative-value design and proposed casting inside IS_SPECIAL_VARNO to accommodate unsigned callers in #20.
- Aleksander Alekseev reported successful installcheck-world testing and no additional problems found by inspection in #22; Tom then rebased the patch and added the compatibility cast in #23, retaining negative sentinels despite the extension concerns.
- Tom explicitly left comprehensive int-versus-Index cleanup outside the patch's scope in #24; Andrei reported successful SELECT and JOIN tests with two tables containing 100,000 partitions each in #25.
- At the end of the discussion, Andrei reported extreme planning time and memory consumption for the large synthetic join in #28, and Aleksander asked that further performance investigation focus on realistic workloads in #29; that performance work remained separate.
- Later that day, e3ec3c00d8 landed the full patch from #23, with only a redundant assertion cast removed and rebase adjustments, completing the rangetable-limit change.