model-writtenrun 4774 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund reported that postgres_fdw could block inside libpq during connection establishment, preventing termination and leaving other backends waiting for ProcSignalBarrier acknowledgments during parallel regression tests #1.
- Andres supplied a postgres_fdw-local interruptible connection prototype, reported successful repeated test runs, and identified the same problem in dblink; he proposed sharing the polling and descriptor-accounting code through either a static library or inline helpers #2.
- Thomas Munro supported exploring a shared header and possible future direct backend linkage to libpq, but Tom Lane opposed direct linkage because of software-update concerns; Andres also raised library-version conflicts and proposed a BUILDING_DLL guard against backend use #4 #5 #6.
- Thomas suggested a separate header-library location, while Andres retained src/include/libpq and submitted shared connection and disconnection helpers used by postgres_fdw, dblink and libpqwalreceiver, adding descriptor accounting to the latter #7 #8.
- Andres deferred a more invasive connection-object wrapper and recommended against backpatching the connection changes because of their size and the absence of field reports; Thomas endorsed proceeding despite reservations about the header location #8 #11.
- Andres split the work into a prerequisite walreceiver error-handling fix, the helper header, the dblink/postgres_fdw conversion and the walreceiver conversion, with comments and function ordering polished; the walreceiver conversion was considered cleanup rather than the interrupt-handling bug fix #12.
- After separately landing the resource-leak prerequisite, Andres announced the main landing: 28a591711d, e4602483e9 and 728f86fec6 exactly match their respective components of #12 #13.
- Nathan Bossart relayed crashes found and traced by Robins Tharakan to e4602483e9, and identified that dblink_connect() disconnected the newly created connection instead of the previous unnamed connection #14.
- At Andres's request for regression coverage, Nathan submitted the variable correction with a test that connects twice and then disconnects; Robins subsequently reported several successful hours of testing against his previously crash-prone workload #15 #16 #17.
- Andres committed Nathan's fix and tests unchanged as fbed4bc59e and confirmed the push in the final message, leaving the submitted connection work and its reported regression resolved #18.