model-writtenrun 6153 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sait Talha Nisanci reported that PostgreSQL 14beta3 rejected a previously working SELECT DISTINCT ROW query because bit lacked a hash function #1; David Rowley reduced the reproduction and suggested adding a bit hash operator class #2.
- Tom Lane rejected requiring every type to support hashing and proposed removing the type cache's optimistic assumption that anonymous records are hashable #3 #4; after Peter listed other affected types #5, Tom argued that smarter planner analysis should wait beyond PostgreSQL 14 and that named composite casts could provide a workaround #6.
- Peter explained that recursive path tracking and SEARCH/CYCLE still needed record hashing, then offered two variants: default to unhashable with caller overrides, or retain hashing by default and suppress it where alternatives exist #7.
- Tom preferred the conservative first variant #8, and Peter posted a polished patch that disabled general anonymous-record hashing while preserving explicit overrides for recursive unions and array hashing, with updated union regression expectations #9.
- Peter committed #9 as a3d2b1bbe9, moving synthetic type-cache entry creation inside hash_array's cache-refresh conditional before landing, and announced completion #10.
- Tom subsequently traced buildfarm Valgrind failures to the synthetic entry's uninitialized type_id and identified a permanent memory leak from allocating it in CacheMemoryContext, proposing the function memory context instead #11.
- Peter supplied both corrective changes #12; Tom reviewed them and suggested zero-initializing the synthetic entry with palloc0 #13.
- Peter incorporated palloc0 and landed the follow-up as 851ff93357, confirming the push in the final message #14; both fixes were complete, while broader anonymous-record planner improvements remained deferred.