model-writtenrun 6861 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sergei Kornilov reported a roughly six-minute shutdown delay in #1, traced it to hash-index vacuum holding interrupts across bucket cleanup, and suggested checking interrupts before acquiring each bucket's lock.
- Mostafa Nabil supplied that fix in #2 and reported cancellation improving from about 2.08 seconds to 0.014 seconds; he avoided a timing-dependent automated test and deferred interruptibility within a single long bucket because it would require changing lock chaining.
- Kirill Reshke endorsed the fix and requested removal of the ineffective delay call inside hashbucketcleanup() in #3, which Mostafa implemented in #4; Kirill considered an injection-point test possible but disproportionate.
- Heikki Linnakangas expanded the patch in #6 after assertion-assisted regression testing exposed similar calls in GIN and ANALYZE, and added a defensive early return in vacuum_delay_point(); he omitted the assertion because unexamined callers, including extensions, might still trigger it.
- Shihao Zhong reported no additional assertion failures and described a caller audit in #7, then pointed Heikki to Kevin Rocker's parallel work covering the same fixes.
- Heikki announced the landing in #8: 96c99955b6 preserves all code changes from #6, the latest matching patchset across this thread and thread #253278, and adds removal of a misleading, ineffective interrupt check in statistics writing after considering that parallel discussion; the commit specifies backpatching through PostgreSQL 14, while interruptibility within a single long hash bucket remains deferred.