model-writtenrun 5907 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane explained in #1 that applying aliases could relabel named composite values as anonymous RECORD values, allowing tables to store data unreadable by other sessions; he proposed preserving named types and supplied a separate cleanup of unused column-name lists.
- Robert Haas questioned the loss of alias behavior in #2, arguing that aliases should rename columns consistently; Tom countered in #3 that this would conflict with functions expecting the original named composite type.
- After Robert tested the example in #4, Tom corrected its alias syntax and demonstrated that typed functions and row_to_json already observed different column names in #5; Robert acknowledged that inconsistency but noted that the proposed fix would still differ from SELECT * behavior in #6.
- Tom argued in #7 that making aliases change composite types everywhere would break longstanding SQL, complicate query deparsing and stored queries after column renames, and require substantially more work than reverting the problematic behavior.
- Returning to the issue in March, Robert accepted that preventing unreadable tables warranted proceeding despite compatibility concerns in #8, and Aleksander Alekseev confirmed the fix and successful tests while recommending it for merging in #9.
- Tom committed the executor fix from #1 as ec62cb0aac, retaining its implementation while removing the view regression cases, and backpatched it to supported branches because data corruption outweighed the behavioral compatibility concern.
- The separate cleanup from #1 landed unchanged apart from rebasing as d7b5c071dd; Tom confirmed completion in #10, leaving both submitted components committed.