model-writtenrun 7382 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Amit Kapila reported intermittent hoverfly recovery-test failures caused by IPC::Run encountering a broken pipe while sending SELECT 1 during server restart in #1.
- Michael Paquier proposed polling with pg_isready in #2, while Tom Lane favored passing an undefined query through the existing poll_query_until helper in #3.
- Tom Lane's patch in #4 used undefined queries for restart probes and required empty stderr to prevent connection failures with empty stdout from being mistaken for success; he rejected checking exit status because existing tests sometimes depended on accepting status 1.
- Tom Lane committed #4 unchanged as f452aaf7d4, with backpatches through PostgreSQL 10, and reported in #5 that the fix worked across branches and that switching query input to stdin had introduced the failures.
- After the commit, Andrew Dunstan questioned the fragility of requiring undefined queries and proposed temporary-file input in #6; Tom Lane raised efficiency concerns and noted that SIGPIPE could expose unread input, while asking whether the failures could be made reproducible in #7.
- Andrew Dunstan suggested creating the file outside the polling loop and noted that psql and safe_psql also used stdin in #8; Tom Lane agreed that all helpers needed consistent behavior in #9, leaving that broader redesign unresolved while the immediate fix remained committed.