model-writtenrun 7651 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Brown reported that recursively synchronizing roughly 700,000 database files delayed startup by seven to ninety minutes and requested an opt-in alternative in #1.
- Thomas Munro rejected portable sync() because of its completion and error-reporting limitations, proposed configurable Linux syncfs(), and described a longer-term WAL/checkpoint-based alternative in #2 and #3.
- Tom Lane required coverage of separate tablespace filesystems and warned about missing syncfs error reporting before Linux 5.8 in #4 and #8; Thomas confirmed the existing prototype covered tablespaces and linked WAL directories in #5.
- Thomas brought the earlier prototype into this thread with an fsync-default configuration setting and documentation in #10, then fixed Windows configuration generation and typos in #11.
- Paul Guo and Fujii Masao argued that already-synchronized backups should be able to skip startup synchronization in #12 and #15, while Thomas questioned the safety of signal files and a persistent none setting in #13, ultimately proposing that syncfs proceed independently in #18.
- Paul reviewed symlink handling and suggested a clearer setting name in #19 and #22, while Fujii requested documentation covering backup startup in #14; Thomas incorporated these changes in #23 and demonstrated synchronization through a cross-filesystem symlink in #24.
- Fujii's review in #25 prompted transient-file APIs, nonfatal directory-read errors and clearer error messages, while Bruce Momjian and Fujii proposed recovery-oriented naming in #27 and #28; Thomas adopted recovery_init_sync_method and demonstrated successful and injected-error behavior in #29.
- Thomas expanded the submission in #29 with a none mode and an experimental WAL mode that synchronizes WAL first and schedules synchronization for data pages skipped during replay, explicitly reserving the WAL work for PostgreSQL 15.
- After correcting the mistakenly attached WAL revision in #30, Thomas incorporated Fujii's error-reporting and none-mode documentation feedback from #31 and #32 into #33, whose WAL patch still questioned checkpoint durability assumptions and coverage of other files.
- Justin Pryzby and David Steele requested clearer documentation of tablespaces, WAL and unsupported symlinks in #35 and #36; Thomas applied those edits and committed only the syncfs component of #33 as 61752afb26.
- Thomas put the none mode on hold because safe use required remembering to undo it, announcing that decision alongside the syncfs landing in #37; David continued the discussion of supported symlinks and missing documentation in #38.
- In May, Justin questioned the restart-only restriction in #40 and posted an inline patch making recovery_init_sync_method reloadable in #42, so administrators could prepare for a future crash without stopping the running service.
- Michael Paquier initially opposed reloadability over API and usefulness concerns in #43, but Justin explained the operational benefit in #44, Fujii identified automatic crash restart as a concrete use case in #46, and Michael accepted that reasoning in #47.
- Thomas tested the reloadable behavior in #48 and committed Justin's inline patch as 34a8b64b4e, announcing completion in the final message #49; syncfs and reloadability had landed, while none remained on hold and the WAL alternative remained future work.