model-writtenrun 1357 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas opened the thread by reporting that the newly extended checkpoint/logical-slot test broke Windows CI because standby startup could not rename backup_label #2.
- Tom Lane established that successful Windows buildfarm runs did not exercise this injection-point test, and Greg Burd subsequently enabled injection points on unicorn to improve coverage #4 #26.
- The investigation focused on Windows file sharing: Robert Haas found that delaying standby startup did not help, and Amit Kapila reproduced the failure and identified ERROR_SHARING_VIOLATION from MoveFileEx #13 #17.
- Andres Freund questioned the unnecessary archive-restoration configuration, which Amit Kapila acknowledged was a copy-and-paste oversight and later removed #15 #21 #24.
- Amit Kapila's diagnostic attachment disabled logical decoding and its injection-point wait in the preceding test, making the combined test pass and implicating reuse of the earlier primary node #20.
- Amit Kapila then preserved the test and explicitly quit its checkpoint background session, reporting consistent success; Robert Haas supported promptly committing the fix to restore CI #22 #23.
- The final attachment finished all three background IPC::Run sessions after the earlier test, avoided sending another quit command to already closed sessions, and removed has_restoring; Amit Kapila reported successful Linux and Windows tests #24.
- Andres Freund supported the cleanup while disputing the proposed parent-directory locking explanation and recommended inspecting open handles separately #25.
- Andrey Borodin withdrew his earlier suggestion that a longer rename timeout helped, explaining that an apparently improved CI job had not actually run and that long timeouts still failed #19 #27.
- Amit Kapila committed #24 unchanged as 9649f1adfd, with backports through PostgreSQL 17.
- After the fix landed, Amit Kapila reported that handle.exe showed backup_label handles in Perl and standby-related processes before the fix and none afterward; the cleanup had landed, but the mechanism connecting unfinished primary sessions to those handles remained unexplained #28.