model-writtenrun 7311 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro proposed a Windows fdatasync() wrapper and a companion pg_test_fsync buffered/direct comparison in #1, initially treating the Windows implementation as an experiment restricted to newer Windows and NTFS.
- Munro's measurements exposed suspiciously fast open_datasync results with device write caching enabled in #3; Robert Haas supported safer defaults in #5, but Munro kept fdatasync opt-in because non-NTFS support was uncertain in #6 and #8.
- The revision in #6 fixed CI problems and enabled the wrapper through Windows build configuration; Munro subsequently reported roughly 25% more TPS than fsync in his virtual-machine tests in #10.
- Munro deferred the work to the PostgreSQL 16 cycle to avoid implementing old-Windows detection in #11; Michael Paquier, Tom Lane and Dave Page discussed compiler requirements and the practicality of retiring older Windows versions in #12, #13, #14 and #15.
- After the Windows support baseline advanced, Munro refreshed the series and documentation in #16, also proposing eventual removal of the fdatasync availability probe once the remaining older buildfarm systems were retired or upgraded.
- Tom Lane challenged assumptions about macOS durability but accepted the SUSv2 availability argument in #17 and #19; Munro clarified that platform defaults would remain unchanged in #18, and Michael Paquier's declaration-check question in #20 established why HAVE_DECL_FDATASYNC still needed to remain in #21.
- The Windows component of #16 landed unchanged apart from rebasing as 9430fb407b, while the pg_test_fsync comparison remained uncommitted.
- Announcing that landing in #23, Munro introduced a separate cleanup series removing obsolete O_FSYNC compatibility macros and the fdatasync availability probe, with the latter initially waiting on the remaining old macOS buildfarm host.
- David Rowley reported HAVE_FDATASYNC redefinition warnings and supplied a successfully compiled guard fix in #24; it landed verbatim as bde60daa0e, and Munro rebased the pending cleanup over it in #26.
- The O_FSYNC component of #26 landed as a1b56090eb with an explicit undefined-O_SYNC condition and explanatory comment; Munro posted the remaining probe cleanup in #27.
- The remaining cleanup continued in thread #46216, whose later patchset #512500 superseded #27 and supplied the version committed as d2e150831a, retaining the Solaris library search and renaming the Windows wrapper to win32fdatasync.c.
- In the final update #28, Munro relayed David Rowley's hardware tests showing fdatasync performance benefits and possible open_datasync cache-safety problems even on NVMe; the Windows implementation and portability cleanups had landed, with the probe cleanup completed through thread #46216, while changing the Windows default, checking non-NTFS behavior and completing the pg_test_fsync comparison remained open.