model-writtenrun 7559 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Rohit Bhogate reported that assigning a composite array element and committing inside a DO block produced a tuple-descriptor reference-leak warning followed by a resource-owner error #1.
- Michael Paquier initially attributed the failure to newly supported assignment syntax and investigated unbalanced tuple-descriptor reference counts #2, but Tom Lane demonstrated that a composite IS NOT NULL expression reproduced the underlying resource-owner problem back to PostgreSQL 11 #3.
- Zhihong Yu proposed transferring remaining tuple-descriptor resources from the old resource owner to the new one after an internal commit, while noting unresolved implementation questions #4.
- Tom Lane posted the sole patchset after finding that changing resource-owner placement introduced subtransaction-lifetime problems and potential performance costs; his alternative cached persistent typcache entries, detected descriptor changes through tupDesc_identifier, and added transaction regression tests #5.
- The thread ended with Tom Lane's proposed patch and discussion of its backpatch compatibility #5; he subsequently committed it with only comment punctuation changed as c2db458c10, with backpatches through PostgreSQL 11 and no intervening patch review recorded in this thread.