model-writtenrun 9103 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier opened the thread by announcing a second landing of session start and end hooks, with a simplified test module after an earlier attempt encountered missing MSVC support for NO_INSTALLCHECK #1.
- Buildfarm failures under force_parallel_mode=regress exposed the test module attempting DML from parallel workers, prompting Michael Paquier to attach a workaround that disabled parallelism through configuration overrides #2.
- Tom Lane challenged the underlying safety of executing queries after proc_exit had begun and recommended reverting the feature if its useful cases required that behavior; Michael Paquier explained that logging session-end activity motivated the test #3 #5.
- Michael Paquier changed the parallel-worker fix to skip logging in parallel workers, landing it as 002962dc72 instead of the attached configuration overrides, and reported that crake was green again #4 #6.
- A remaining Windows failure was traced by Andrew Dunstan to SSPI authentication and missing matching authentication configuration; Michael Paquier posted an independent test patch using --create-role and revised role setup and cleanup #7 #8 #9.
- Andres Freund demonstrated a crash with forced JIT compilation and explained that the session-end hook ran after before_shmem_exit callbacks had already shut down subsystems it could access #10.
- Michael Paquier acknowledged that ensuring transactions were aborted did not make the hook's location safe and chose to revert; 9555cc8d2b removed both hooks and their test module, leaving the proposed Windows test fix unapplied #11 #9.
- After the revert, Fujii Masao proposed restoring only the session-start hook, Michael Paquier was open to splitting the pair, and Alvaro Herrera supported the start hook while suggesting that an end hook would need to run before subsystem shutdown and might not run in every case #12 #13 #14.
- The thread ended without a replacement landing: Fujii Masao questioned whether existing exit callbacks made a dedicated session-end hook necessary, while Andres Freund called for substantially more design and code review #15 #16.