model-writtenrun 8486 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Hayato Kuroda reported that an exiting backend could remove another backend's newly attached injection point when both used the same name, and proposed tracking slot-and-generation handles alongside a reproducer; he also supplied an alternative that waited for cleanup in the affected concurrent-upsert test #1.
- Kuroda added missing typedef registrations and removed the alternative attachment to avoid confusing cfbot, then resent the series after discovering that his mailer had attached an invalid file #2 #3.
- Michael Paquier opposed complicating core injection-point facilities for this case and recommended waiting for backend exit or using different names #4; Kuroda explained that checking ownership by PID also crossed the core/module boundary and remained open to fixing the tests instead #5.
- Koshi Shibagaki suggested a module-supplied callback that would check permission to detach under the removal lock, but questioned whether the additional core API complexity was justified without another use case #6.
- Masahiko Sawada reconsidered his original tracking suggestion after seeing its complexity and proposed a reusable optional backend-exit wait in BackgroundPsql::quit for the concurrent-upsert and effective-WAL-level tests; he considered distinct names or explicit detachment sufficient for isolation tests #7.
- Paquier advocated waiting for backend cleanup by default throughout the testing infrastructure #8, while Kyotaro Horiguchi agreed that this was a test-usage issue and opposed adding machinery to core or local injection points #9.
- Kuroda switched to a three-patch Perl series that saves the backend PID during wait_connect(), polls process existence after psql exits, updates affected tests, and finally enables waiting by default with quit(0) for the temporary-object test that intentionally blocks cleanup #10.
- The thread ends with the Perl replacement submitted for review and no linked commit; Kuroda chose PID capture at connection time to avoid burdening callers or mixing asynchronous query output during quit(), leaving the core handle proposal behind #10.