model-writtenrun 4385 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep - Oct 2021 · 3 entries
- Thomas Habets proposed sslmode=verify-system to let libpq verify server certificates against system trust roots, motivated by reuse of publicly issued certificates, and supplied a tested implementation in #1.
- Tom Lane reproduced the existing requirement for a private root.crt and suggested automatic system-store fallback in #6, while Andrew Dunstan proposed the explicit sslrootcert=system selector in #9.
- Thomas accepted the selector and emphasized hostname verification in #15; Jacob Champion argued that automatic fallback could weaken explicitly selected trust and proposed separating system-root support from broader default changes in #18.
Oct 2022 - Apr 2023 · 16 entries
- After roughly a year, Jacob revived the work as two patches in #24: one loading system roots and another defaulting to verify-full when those roots were selected.
- Thomas argued that accepting weaker verification would be unsafe in #25, and Jacob's revision #34 rejected weaker explicit SSL modes while adding LibreSSL detection because its tests could not override roots through SSL_CERT_FILE.
- Jacob traced macOS test failures to OpenSSL's error queue and a missing certificate directory in #37, added a Homebrew cleanup workaround in #38, and adjusted its path for M1 machines in #39.
- Jelte reconsidered automatic fallback and server-side system roots in #40; Jacob defended explicit trust selection and kept client-certificate support outside this patch in #44 and #47.
- Jelte tested the feature against Azure and requested improved libpq error reporting in #45; Jacob incorporated those changes in #48, which Jelte approved in #49.
- Magnus Hagander questioned the configure checks and environment-variable API usage in #51, then agreed that adding C API plumbing to the Perl tests was unnecessary in #53.
- Following configure and Meson rebases in #55 and #58, Daniel Gustafsson's review #59 led to allocation-failure handling, clearer diagnostics, and documentation of the system filename collision in #60.
- Daniel accepted skipping the relevant tests on obsolete OpenSSL versions in #63, allowing #64 to replace new configure probes with an existing feature-macro heuristic.
- Daniel squashed #64 into 8eda731465, changing only formatting and comments, and announced the landing in #65.
- Peter Eisentraut reported that OpenSSL 3.1 produced an undefined SYSCALL error instead of the expected verification error in #67; Jacob reproduced it without the CI workaround in #75, and Tom reproduced it with MacPorts in #76.
- Peter and Tom confirmed that recreating the certificate directory restored passing tests in #77 and #78, while Daniel traced the behavior to deferred loading through OpenSSL's certificate-store API in #80.
- The investigation prompted Tom to establish a MacPorts buildfarm animal; its PostgreSQL 11 LDAP test failure led to the backpatch proposed in #82 and committed as 9517e6e1dd, with the animal reported online in #83.
- Daniel proposed extracting certificate verification errors from the SYSCALL failure path in #85; Tom's reviews #86, #88, and #90 led to portable errno preservation and the corrected error-code argument in #92.
- Jacob demonstrated that the broad fix could misreport interrupted handshakes with weaker SSL modes in #93, so Daniel restricted it to the known missing-system-issuer case in #94, leaving broader SSL error-reporting improvements for future work.
- Jacob approved the narrowed fix in #95; 0b5d1fb36a applied #94 unchanged in C and added the agreed OpenSSL 3.0 test expectation, after which Daniel closed the open item in #97.
- With both error variants handled, Jacob supplied and tested the CI workaround removal in #98; Daniel committed it unchanged as c5e4ec293e and confirmed successful CI in the final message #99, leaving the submitted feature, bugfix, and cleanup landed.