model-writtenrun 7168 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund reports that interrupting DROP DATABASE can roll back its catalog deletion after irreversible damage, allowing connections that panic, and explains that simply holding interrupts would deadlock the process-signal barrier #1.
- Thomas Munro tests an externally posted invalidation patch, confirms that interrupted drops prevent subsequent connections, recommends a hint and skipping invalid databases in autovacuum, and endorses the approach; Andres agrees that vacuum horizons also need protection while deferring automatic cleanup by autovacuum #2 #3 #4.
- Andres resumes work by extending invalid-database handling across vacuum, database templates, and client tools, then devises a reproducible interruption test using a pg_tablespace lock #5 #6.
- The first attached series marks databases invalid through a durable in-place datconnlimit update before irreversible operations, prevents their use, permits retrying DROP DATABASE, and adds tests alongside a separate vacuum lock-release fix #7.
- Daniel Gustafsson reports successful manual and automated testing, requests named constants, corrected hints and debug logging, and notices the missing pg_upgrade test; Andres incorporates those changes while retaining datconnlimit as the invalidity marker and leaving psql's database listing unchanged #8 #9.
- Daniel endorses the revised approach and resends Andres's two patches with an additional pg_upgrade preflight diagnostic, proposed as a separate change to simplify backpatching #10.
- Andres lands the main interruption fix as c66a7d75e6 from the second patch of #10, adding documentation and light code and test cleanup, and reports that backpatching the fixes required substantial test adaptation #11.
- The added pg_upgrade diagnostic is excluded from that landing, and Daniel says he will submit it in a separate thread #12.
- Peter Eisentraut later finds that the new pg_dump test uses an unsupported not_like key and recommends an explicit empty like condition; Andres agrees, and 849d367ff9 implements the discussion's correction with a comment explaining that invalid databases must never be dumped #13 #14.
- Alexander Lakhin subsequently reports that the interruption test hangs with cache-clobbering settings because its pg_tablespace lock blocks new backend initialization; Thomas acknowledges the problem, but the thread ends without a proposed fix for that test hang #15 #16.