model-writtenrun 8756 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Mahendra Singh Thalor reproduced an autovacuum assertion after dropping a temporary schema and creating another temporary table, traced it to a NULL namespace name, and proposed resetting session state or prohibiting the drop in #1.
- Michael Paquier chose prohibition because properly resetting temporary-namespace state and cleanup callbacks was more involved, posting the first patch in #2; Kyotaro Horiguchi and Mahendra identified its missing cross-session coverage in #3 and #5, leading to isAnyTempNamespace in #7.
- Mahendra confirmed the revised patch worked in #8, while Tom Lane explained that temporary schema numbers are unstable for regression tests and that schemas are normally recycled in #9.
- Michael committed #7 as a052f6cbb8, changing only the error terminology to "temporary schema" and backpatching through 9.4, as reported in #11.
- Robert Haas objected that the prohibition removed a useful database-repair operation and masked an autovacuum robustness bug in #12; Michael responded with an allow_system_table_mods escape hatch and NULL-safe logging in #13.
- Robert proposed separate logging for missing namespaces in #14 and rejected retaining the restriction in #17, while Tom clarified that only superusers could drop these schemas and supported restoring that capability in #18.
- Michael combined a full revert with revised logging in #19; at Tom's request in #20, the revert landed separately as 65192e0244, leaving corrected logging for further review in #21.
- Mahendra's testing of the logging fix exposed another assertion during relation-descriptor construction in #23, showing that making the log message NULL-safe did not make deletion of these relations safe.
- Michael changed autovacuum to skip relations whose namespaces were missing in #24, then added the same check immediately before deletion to close a race in #25, eliminating the need for alternate logging.
- Mahendra confirmed the crash was gone but reported that newly created temporary tables remained invisible to ordinary relation listing in #26; Michael explained in #29 that dropping the schema had left the session state inconsistent.
- Tom questioned whether skipping relations could hold back datfrozenxid in #27; Michael explained the relcache and storage obstacles to deleting them in #28 and retained the conservative autovacuum-only fix in #30, omitting the separate namespace-helper adjustment.
- After a several-week gap, Mahendra confirmed that #30 fixed the assertion in #31, and Michael proposed backpatching it through PostgreSQL 11 in #32.
- Tom accepted skipping deletion but proposed expressing missing, idle, and active namespaces explicitly in #33, then supplied the three-state checkTempNamespaceStatus implementation in #34.
- Michael supported retaining the old API on stable branches in #35, and Tom committed #34 with an expanded explanatory comment as 80d76be51c, backpatching through PostgreSQL 11.
- The thread ends with Michael acknowledging the final commit in #37: the prohibition was reverted and the autovacuum crash fix landed, while repairing session state and orphaned objects after manual schema deletion remained outside the committed fix.