model-writtenrun 7316 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier opened the thread with #1, showing that pg_tablespace_location() failed on in-place tablespace directories and proposing a relative-path result with regression coverage; Kyotaro Horiguchi supported the fix while raising the absolute-versus-relative-path question in #2.
- Kyotaro Horiguchi reported a related pg_basebackup readlink() warning in #3 and initially suspected misplaced backup files in #5; Thomas Munro demonstrated correct directory copying and supplied a separate warning fix in #6, after which Horiguchi traced the misplaced files to his own experimental change in #8.
- Michael Paquier identified that Windows junctions would be misclassified by the first backup patch in #10, leading Thomas Munro to add explicit junction detection in #11; Horiguchi confirmed the revision on CentOS8 and Windows11 in #13 and showed that Windows previously suffered a backup failure rather than merely a warning.
- Thomas Munro committed the revised backup patch as c6f2f01611, retaining its logic and adjusting its explanation to cover the Windows failure, then announced the landing in #16.
- Thomas Munro also proposed returning an empty string from pg_tablespace_location() in #16, matching the original LOCATION argument, but Michael Paquier argued in #18 that tests needed a usable filesystem path; Paquier preferred a relative path to distinguish in-place tablespaces in #20, and Munro agreed in #21.
- After Kyotaro Horiguchi requested documentation in #22, Michael Paquier combined the simpler platform-specific checks with relative-path semantics, the original tests, and a documentation paragraph in #23; Horiguchi tested that revision on Rocky8 and Windows11 and challenged its wording in #24.
- Michael Paquier landed #23 as f6f0db4d62, preserving the implementation and tests while shortening the documentation, and announced it in #25.
- Thomas Munro pointed out that the new documentation still omitted the empty LOCATION requirement in #26, then proposed removing the developer-only detail entirely in #28; Michael Paquier and Kyotaro Horiguchi agreed in #29 and #30, producing the documentation-only rollback 7a7cd84893.
- Revisiting the Windows workaround in #31, Thomas Munro identified that directory attributes were checked before junction attributes, while noting that pgwin32_is_junction() error handling was being discussed separately; Michael Paquier supported reversing the checks in #32.
- Thomas Munro supplied a two-patch follow-up in #33 to fix Windows junction classification, remove the backup workaround, and add a TAP test, but also reported that tar-format backups appeared to omit in-place tablespaces; Michael Paquier reviewed the series in #34 and initially thought those tablespaces were included in the main tarball.
- After a gap until July, Thomas Munro landed the junction fix as 9d3444dcce with an expanded comment and the cleanup/test patch as 5344723755 with a small test-variable adjustment, announcing both in #35.
- The thread ends with Michael Paquier confirming in #37 that plain-format backups include in-place tablespaces but tar-format backups omit them from both the main archive and separate tablespace archives, leaving that bug unresolved despite all submitted fixes having landed.