model-writtenrun 2025 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas opened the discussion by questioning inconsistent fabricated relation names and the difficulty of distinguishing them from user-supplied names #1; Peter Eisentraut explained that the recursive-CTE names came from SQL-standard terminology #2.
- Robert tested replacing fabricated eref names with NULL and posted a broad proof of concept that generated fallback names for display #3 #4.
- Tom Lane objected that nullable eref names could crash existing code, warned that generated names such as *VALUES* can be referenced by working SQL, and identified rte->alias as the intended record of user input #5 #7 #8; Robert accepted the compatibility concern, while Andrei Lepikhov confirmed extension reliance on non-NULL eref names #9 #10.
- After the discussion stalled, Tom reopened it to resolve its implications for separate VALUES-display work #11; the ensuing uniqueness discussion led Robert to explain his real goal of identifying plan components for planner-control extensions and acknowledge that removing generated names would not solve that broader problem #12 #13 #14.
- Tom found a narrower actionable bug: generated SELECT names were incorrectly stored as user aliases, so he posted a replacement patch that leaves rte->alias unset and preserves generated eref names #15; Robert supported correcting that distinction #16.
- Robert returned in April with a rebase of Tom's patch and two analogous fixes for ANY_subquery and recursive CTEs, favoring consistent unnamed_subquery output over an expanded parser API; he targeted PostgreSQL 19 and separately questioned inheritance-child alias handling #17.
- Once PostgreSQL 19 development opened, Tom accepted all three patches, fixed an additional memoize expected-output failure and polished their commit messages without changing backend logic; he argued that inheritance-child aliases should remain because ruleutils uses them to preserve column aliases #18.
- Robert committed the final SELECT and ANY_subquery patches unchanged apart from rebasing as 585e31fcb6 and 6f79024df3, using the series posted in #18.
- Robert also committed the recursive-CTE patch from #18 as 5a170e992a, preserving its code and expanding its commit message, then confirmed that all three patches had landed #19.
- The thread ended with the submitted series committed and inheritance-child alias behavior unchanged: Robert still questioned why copying alias and eref separately would fail, but said leaving it alone would not immediately obstruct his work #19.