model-writtenrun 9756 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Alvaro Herrera found that adding a foreign key to a partitioned parent could leave a partition's cached foreign-key list empty, causing DELETE to skip action triggers; he proposed handling foreign-key pg_constraint changes centrally in CacheInvalidateHeapTuple instead of adding an explicit invalidation at the failing call site #1.
- Andres Freund raised a consistency concern based on an earlier pg_index invalidation discussion #2, while Tom Lane supported the attached fix because centralized invalidation was safer than expecting explicit calls everywhere #3.
- Alvaro committed the patch #1 unchanged as 1755440935, backpatched it through PostgreSQL 9.6, and announced the push #4.
- Andres objected to committing and backpatching before addressing his concern #5; Tom argued that the foreign-key dependency clearly belonged to the table's relcache entry, unlike the disputed index case #6.
- Alvaro defended central invalidation because cached foreign-key information extended beyond OIDs and pg_constraint updates were uncommon, mentioning a revert and localized invalidation only as an alternative he considered wrong #7.
- Andres challenged the distinction by listing index information cached on tables #8, Tom replied that supported index redefinitions could not change those properties #9, and Andres closed by calling for systematic, consistent invalidation and consensus rather than opposing this fix itself #10; the patch remained committed with that broader disagreement unresolved.