model-writtenrun 3767 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed automatic backtraces for internal errors in #1, explaining that function-name filtering and manually adding errbacktrace() were awkward for common low-level failures.
- Robert Haas advocated unconditional backtraces in #3, while Matthias van de Meent and Nathan Bossart raised concerns about development logging, extensions and excessive output in #6 and #8.
- Peter clarified in #10 that selecting SQLSTATE XX000 already excludes ordinary lower-level elog messages and preserves equivalence between elog() and ereport(); Tom Lane endorsed that design without additional severity controls in #11.
- Peter submitted the default-off setting and a separate variable-name cleanup in #12, leaving the default open for discussion; Tom argued for default-on or unconditional behavior in #13.
- Andres Freund's production-log analysis in #14 showed that automatic backtraces could materially increase output, especially for extensions, and exposed misleading SSL failures reported as Success; he reproduced the disconnect case in #16 and traced it to zero errno in #19.
- Andres supplied an SSL handshake fix and regression test in #20, while Tom expanded the fix to initialization and established-connection reads in #21; the test was left out after platform differences weakened its assertion and Tom questioned its value in #22.
- After Andres questioned write handling in #23, Tom recognized that SSL_write() can itself read and expanded the series in #27 to cover writes, frontend libpq and defensive handling outside SSL.
- Andres identified incorrect error queuing for ordinary EOF and requested explanatory comments in #28; Tom incorporated both changes in #29, making pqsecure_raw_read() return zero for the errno-zero case.
- Tom committed the complete final SSL series from #29 unchanged on the development branch as 0a5c46a7a4 on December 11; the separately supplied candidate 271d24f31d is its PostgreSQL 12 backport.
- The backtrace-policy discussion continued with Robert stressing production debugging benefits in #31, Tom supporting XX000-triggered backtraces while questioning their detail in #32, and Robert agreeing that richer output belonged to a separate project in #33.
- Jelte Fennema-Nio proposed an explicit EOF return value for compatibility with his cancellation work in #34; after Andres defended consistency with existing APIs in #36, Jelte accepted the already-committed approach in #37, leaving questions about other pqReadData() failures to his separate work.
- Jelte also documented widespread extension use of unspecified error codes and requested control over automatic backtraces in #35, reinforcing the concerns about unconditional logging.
- Peter landed the unchanged cleanup portion of #12 as 541e8f14a1 on December 28 and the feature portion as a740b213d4 on December 30, retaining default-off behavior and removing only the TODO about alternative defaults.
- Peter closed the thread by confirming his original patch was committed in #38; both backtrace patches and the SSL fix had landed, while unconditional backtraces, richer diagnostic output and the discarded SSL test were not included.