model-writtenrun 5759 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Nazir Bilal Yavuz reported widespread SSL test failures on OpenBSD 7.0 with LibreSSL in #1; Andres Freund questioned whether LibreSSL tests should be skipped in #2, while Thomas Munro and Tom Lane reproduced failures on older OpenBSD releases in #3 and #4.
- Daniel Gustafsson found that explicit host=localhost settings eliminated most failures, leaving certificate-revocation errors, in #9; his first attachment added those settings and a temporary server-side sleep that made the remaining tests pass in #15.
- Thomas Munro confirmed that removing the sleep restored failures in #18, Andres Freund established that OpenSSL passed on the same operating system in #19, and Tom Lane reproduced the hostname and timing effects separately in #21.
- Tom Lane traced the hostname failures to a temporary socket-directory pathname being passed as the SNI hostname and initially proposed host='' in #22; Daniel Gustafsson confirmed LibreSSL's stricter hostname validation in #24, and Tom ultimately favored localhost so the tests would still exercise SNI in #26.
- Tom Lane located the remaining intermittent failures on the socket write path and supplied an experimental patch that ignored EPIPE and ECONNRESET in #23; Daniel Gustafsson repeatedly tested it without the server sleep and found that ignoring EPIPE alone was sufficient in #25.
- Tom Lane proposed moving write-failure postponement below SSL/GSS processing into pqsecure_raw_write() in #27, replacing the experimental suppression with a mechanism that preserved useful server errors.
- Daniel Gustafsson committed the unchanged hostname-test portion of #15 as 6d503d2a47, omitting the diagnostic sleep, and announced its application to master and PostgreSQL 14 in #28.
- Tom Lane's implementation in #29 deferred socket failures at the physical I/O layer while reporting SSL/GSS-level errors immediately, avoiding invasive changes to both encryption implementations; it also included a small my_sock_write() errno correction.
- Andres Freund reviewed the new implementation, requested clearer comments, and identified a separate PQisBusy() condition that could prevent callers from obtaining an error result in #30; he explored a possible Windows walreceiver connection in #32.
- After initially suggesting an inverted write_failed check in #33, Tom Lane reproduced the PQisBusy() bug in #34 and supplied #35, which instead checks CONNECTION_BAD so input processing can continue until a server message or read EOF is available.
- Tom Lane committed #35 together with the errno correction from #29 as 335fa5a260, backpatched through PostgreSQL 12, and committed the remaining write-postponement changes with revised commentary as faa189c932 on master only.
- Tom Lane reported successful OpenBSD 7.0 testing on physical hardware and a VM after committing in #36, and Thomas Munro confirmed OpenBSD 6.9 success in #37; OpenBSD 6.8 remained broken, with further investigation declined because that release was already unsupported.
- Tom Lane questioned whether improved error reporting justified backpatching faa189c932 in #38, and Andres Freund favored waiting for more confidence or user reports in #39, leaving that backpatch decision open.
- A later report of duplicated FATAL output from isolationtester in #40 was linked by Tom Lane to an existing issue in #41; after Tom encouraged retaining a regression test in #43, Andres Freund reported pushing that test in #44, without another SSL patch submission.