model-writtenrun 7695 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread after Windows recovery tests exposed gaps in buildfarm coverage, and requested parallel testing support through PROVE_FLAGS and a more effective test runner #1; Andrew Dunstan explained that instability or unexplained errors had prompted disabling miscellaneous tests #2.
- Andrew corrected the buildfarm client and enabled almost all TAP tests on drongo, but recovery tests remained excluded because 021_row_visibility.pl crashed the client #4; Michael Paquier confirmed a SIGBREAK failure and requested PROVE_TESTS support #6.
- Andres separately traced recovery-target failures in Windows containers to service detection redirecting server diagnostics away from stderr, confirmed that forcing non-service behavior bypassed that failure, and reported an additional initdb permissions problem #8 #10.
- Michael supplied failure logs #11, while Andrew identified IPC::Run's kill_kill() as a known Windows hazard already responsible for skipping crash recovery and logical decoding tests #12; Andrew also found that Git's bundled Perl lacked a module required to run prove #13.
- Andrew demonstrated an inline fix that sent '\q' to both psql sessions, pumped their input, and slept briefly, eliminating both the hang and signal failure #14.
- Andres proposed waiting for process completion instead of sleeping and suggested applying the same technique to crash recovery #15; Andrew simplified the implementation to finish() alone and reported successful testing #16 #17.
- Andrew posted the sole attached patchset, replacing forced termination in both recovery tests and removing the Windows crash recovery skip, while explicitly leaving pg_recvlogical unchanged because he found no convenient equivalent solution #18.
- After Andres requested commitment #19, Andrew committed the attached patch unchanged as d3676a2e9f and confirmed that kill_kill() itself appeared responsible for the severe crash #20.
- The thread ended with the graceful psql shutdown fix landed; the pg_recvlogical limitation remained, and no completed fixes were shown for the container logging problem, initdb permissions, or the broader test-runner requests #18 #20 #8 #1 #6.