model-writtenrun 6389 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Euler Taveira proposed automatic temporary-file cleanup after backend crashes in #1, motivated by orphaned files consuming disk space until a restart, with a reloadable, default-on GUC preserving the debugging option.
- Tomas Vondra initially approved the approach in #2, while Euler questioned whether a GUC was necessary in #3 and Michael Paquier emphasized users' disk-space problems over the rarely used debugging behavior in #4.
- Anastasia Lubennikova supplied an unconditional-cleanup alternative in #5, dropping both the GUC and tests because she doubted the tests would be stable.
- After several months, Tomas revived the original configurable design in #7, retaining the debugging escape hatch, renaming the setting to remove_temp_files_after_crash and revising its documentation.
- Michael Paquier and Thomas Munro requested documentation and formatting corrections in #8 and #10; Euler incorporated them in #11 and added short sleeps intended to stabilize temporary-file creation in the test.
- Tomas committed #11 as cd91de0d17 with documentation and comment wording edits, announcing the landing in #12.
- Buildfarm failures reported in #13 exposed that the test could inspect a nonexistent temporary directory; Tomas proposed a blocking UNIQUE-constraint test in #14 and confirmed platform-dependent spill thresholds in #16.
- Euler implemented the two-session test in #18, then increased its row count and added a pre-crash file assertion in #20 after Tomas's 32-bit testing in #19.
- Tomas instead committed #18 as a16b2b960f, raising the row count to 5000 and adjusting a comment; he explained in #21 that the extra assertion was unnecessary because the cleanup-disabled case already checked file creation.
- Further buildfarm failures appeared in #22, and Tom Lane identified in #25 that the test's output markers preceded the useful commands and did not establish that either insertion or blocking had occurred.
- Euler proposed polling for temporary-file creation in #26, but Tomas developed command reordering and lock inspection in #27, then supplied a working PL/pgSQL polling loop in #29 after recognizing that the Perl pump functions did not rerun SQL.
- The lock-based version in #29 landed unchanged as e589c4890b; Tomas reported testing it across his machines in #30, and Euler reported apparently successful buildfarm results in #31.
- Justin Pryzby reopened the configuration-placement issue in #32, proposing Developer Options classification and removal from the sample configuration; Euler supplied those changes in #33.
- Michael Paquier accepted the relocation in #34 and committed #33 as 797b0fc0b0, adding the missing GUC_NOT_IN_SAMPLE flag; his final announcement in #35 leaves no proposed work outstanding.