model-writtenrun 9861 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut opened the thread by announcing the integration of recovery.conf settings into postgresql.conf, with recovery.signal and standby.signal controlling recovery modes #1.
- Michael Paquier reported reproducible recovery-target test failures under EXEC_BACKEND #2, and Sergei Kornilov traced them to GUC serialization losing configuration order, breaking the previous last-target-wins behavior #3.
- Michael initially favored preserving last-target-wins semantics #5, while Peter questioned that requirement #4 and Stephen Frost advocated rejecting conflicting targets during recovery startup #7; Sergei implemented that validation in the first patch #10.
- Stephen requested a boolean immediate-target setting and an expected-failure test #11, which Sergei added in #19; Stephen then identified leftover declarations and questioned the target-count check #20, before acknowledging that recovery to the end could justify allowing zero targets #22.
- Alvaro Herrera proposed stopping when any of several recovery targets became true #12, but Stephen Frost and David Steele argued that this would be a separate feature rather than the immediate fix, leaving combined-target semantics outside this work #14 #18.
- Peter rejected renaming recovery_target as unnecessary scope for a parsing fix #25 and supplied a smaller implementation that rejected conflicts directly in GUC assignment hooks #26; Sergei completed its documentation, error wording, and tests in #27.
- Stephen continued to favor the boolean interface cleanup while accepting the assignment-hook approach #28, but the committed fix retained the existing parameter names.
- Peter committed f2cbffc7a6 from #27 with additional logic allowing repeated assignments to the same target and clearing one target before selecting another, along with revised diagnostics and tests.
- The thread ended with Peter confirming the commit and explaining its extra handling for repeated and cleared settings, which he expected to resolve the EXEC_BACKEND failures #29; the conflict-rejection fix had landed, while boolean renaming and combined-target behavior remained outside it.