model-writtenrun 4691 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund reported intermittent postgres_fdw connection-termination failures and tenk1 query-plan changes when regression tests ran against an existing server, suspecting cache resets for the former and insufficient VACUUM diagnostics for the latter #1.
- After an initially suspected related problem was fixed, failures persisted; Andres reproduced the postgres_fdw failure and confirmed that a shared-invalidation reset discarded the remote connection #2 #3 #4.
- Andres found that enlarging the invalidation queue did not eliminate resets and suggested checking whether foreign-server definitions actually changed before dropping connections, but no such implementation followed in this thread #5.
- To investigate the separate tenk1 instability, Andres attached a work-in-progress patch enabling DEBUG1 VACUUM diagnostics without VERBOSE, while leaving instrumentation overhead and additional VERBOSE output as open questions #6.
- Peter Geoghegan questioned the usefulness of per-index deletion counts; Andres proposed phase-progress messages, less repeated statement and context output, and greater verbosity for multiple vacuum passes, while Peter supported some changes but wanted consistent summaries and careful treatment of cleanup-only index scans #7 #8 #9 #10 #11.
- When postgres_fdw failures continued, Andres proposed removing the unstable test; Tom Lane instead proposed suppressing the termination query's variable row count, and Andres agreed and suggested removing the debug_discard_caches workaround #12 #13 #14.
- Tom's first postgres_fdw patch passed the connection-recovery test with cache discarding enabled or disabled, but removed the application_name tests because suppressing their termination results would eliminate their assertions #16.
- Tom considered retaining only escape-sequence code coverage, while Andres suggested using a transaction and questioned why the tests terminated connections; Tom explained the reconnection requirement and remaining race, then proposed having the remote session read its own application name #17 #18 #19 #20.
- Tom implemented the remote-view approach, preserving application_name assertions and using ALTER SERVER to trigger reconnection; Andres approved the revised patch and valued it as an example for future postgres_fdw tests #21 #22.
- The final thread message announced Tom's intent to back-patch the fix; he subsequently committed 53fe7e6cb8 from #21, changing only the initial remote-table probe and expected output to report a count #23.
- The postgres_fdw test hardening landed, while the separate VACUUM logging proposal and tenk1 instability had no demonstrated resolution in the supplied thread #6 #11 53fe7e6cb8.