model-writtenrun 6021 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sadhuprasad Patro reported that logical replication failed when the publisher used SQL, MDY and the subscriber used SQL, DMY in #1; Dilip Kumar proposed making the publisher use the subscriber's DateStyle in #2.
- Japin Li implemented that proposal using a SQL command after connecting in #3 and added a regression test in #4; Sadhuprasad reported successful testing and whitespace issues in #5.
- Masahiko Sawada demonstrated that differing IntervalStyle settings could silently change interval values in #6, while Tom Lane proposed unambiguous output settings without an extra network round trip and pointed to postgres_fdw in #7 and #8.
- Michael Paquier suggested avoiding duplicated GUC-handling logic in #9, but Tom Lane explained that the components needed different save-and-restore behavior in #13; Japin instead moved fixed DateStyle and IntervalStyle settings into walsender startup in #14, which Dilip endorsed in #16.
- Japin's further testing found floating-point precision loss, leading to extra_float_digits=3 and an expanded regression test in #17.
- Masahiko warned that changing every database walsender could affect other decoding plugins in #18 and proposed passing connection options without another round trip in #21; Tom also explained that changing subscriber settings could surprise user-defined triggers in #23.
- After compatibility concerns were illustrated in #27, Japin moved the settings into connection options in #30, initially copying subscriber values; Tom rejected that complexity and its cross-version assumptions in #31, prompting the fixed, curated settings in #32.
- Tom committed #32 as f3d4019da5, refining comments and simplifying DateStyle to ISO to match pg_dump; the commit was limited to the development branch because subscribers storing replicated values as text could observe incompatible output changes.
- The thread ended with Tom's confirmation in #35 that the fix had landed and the regression test was deliberately omitted because it increased the test script's runtime by roughly a third while offering little expected future coverage; no further patch remained pending.