model-writtenrun 1281 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund reported that the server's restricted ECDH curve forced an extra TLS handshake roundtrip when clients offered X25519, measuring almost 30% more connections per second when that roundtrip was avoided #1.
- Daniel Gustafsson initially suspected protocol-version negotiation, but Andres showed that requiring TLS 1.3 did not help and that the server's curve restriction caused the retry #2 #3.
- Jacob Champion cautioned that OpenSSL's defaults were broader than recommended because of denial-of-service concerns, while Daniel connected the solution to separate work allowing a list of server curves #4 #5.
- Andres and Daniel discussed accepting both curves and possibly fixing stable branches, but recognized the compatibility risks of changing defaults and left backpatching unresolved #6 #7 #8 #9.
- Ahead of PostgreSQL 18, Jacob renewed the proposal and Daniel submitted a four-file patch adding x25519 alongside prime256v1 to the default configuration and SSL tests #10 #11 #12.
- Jacob discovered that LibreSSL required the uppercase spelling X25519, then forwarded Daniel's patch with a capitalization fix tested across LibreSSL and OpenSSL versions #13 #16.
- Daniel committed the combined submission without code changes as daa02c6bd9 and explicitly left stable-branch treatment for another discussion #19.
- After the initial landing, Andres asked for better diagnostics to uncover unnecessary TLS roundtrips, and Daniel noted that useful tracepoints might lie inside OpenSSL rather than libpq #18 #19.
- Nearly a year later, Tom Lane reported that the new default broke SSL connections under FIPS because X25519 was forbidden, requesting a documentation warning and compatible regression tests while also suggesting earlier configuration validation #20.
- Daniel and Tom discussed validating groups through a temporary SSL context, but prioritized the simpler fixes; Daniel said related validation was being developed in the SNI work and submitted separate test and documentation patches #21 #22 #23.
- The two components of #23 landed unchanged as 07e90c6913, removing X25519 from the SSL tests, and db93988ab0, documenting the FIPS configuration requirement.
- Nazir Bilal Yavuz subsequently found two more affected test suites, prompting Daniel to submit explicit FIPS-compatible curve settings for the libpq and ssl_passphrase_callback tests #24.
- Tom objected to progressively overriding test configurations and proposed a compile-time FIPS default, while Daniel favored tests that worked under FIPS plus a dedicated CI job; Tom stressed preserving coverage of the normal default configuration #25 #26 #27.
- The thread ended with tentative agreement to detect FIPS in the test framework and adjust settings only when necessary, plus Tom's request to backpatch the fixes to PostgreSQL 18 after its re-release; the additional test patch and detection approach had no confirmed landing in the supplied record #28 #29.