model-writtenrun 3825 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut demonstrated that differing partition-key and unique-index collations could let equivalent values occupy different partitions, and proposed rejecting such index definitions with a dedicated collation error #1.
- Laurenz Albe supported the patch but requested clearer error wording and raised backpatching and existing-index concerns; Peter initially considered the case too exotic to backpatch, while Tom Lane favored backpatching and identified that immediate rejection could prevent finding another valid index-column match #2 #3 #4.
- Jeff Davis suggested allowing different deterministic collations because equality would agree, but Tom argued that ordering requirements and subtle-bug risks justified exact matching, and Jeff withdrew the suggestion #5 #6 #7.
- Peter proposed continuing past collation mismatches and relying on the existing missing-partition-key error; Tom explained his concern about misleading errors for incompatible types and accepted that continuing could resolve it #8 #9.
- Peter's revised patch implemented that continuation behavior, retaining the existing, less-specific error message for definitions without a matching column #10.
- After Peter announced his intention to commit, Tom explicitly accepted the revised patch in the thread's final message, leaving improved error diagnostics until field reports demonstrated a need #11 #12.
- Peter subsequently committed the revised patch unchanged as a11c9c42ea, with Tom credited as reviewer and backports recorded for five stable branches, completing the proposed fix #10.