model-writtenrun 5237 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2021 - Apr 2022 · 9 entries
- Bharath Rupireddy proposed logging and continuing after logical-rewrite filename parsing or deletion failures in #1, later explaining that manual deletion during recovery had caused a checkpoint failure and repeated recovery work in #6.
- Nathan Bossart initially supported the patch in #2, while Andres Freund warned that weakening errors could conceal persistent problems in #5; Tom Lane argued that repeated checkpoint failure could itself exhaust disk space in #13.
- Tom Lane identified a missing directory fsync while reviewing the code in #9 and clarified that file contents still needed syncing in #10; Andres Freund agreed that both file and directory durability were required in #11.
- Nathan Bossart combined directory fsync, better lstat checks, and the original softer errors in #14, but Andres Freund challenged the mixed error policy and requested a separate backpatchable fsync fix in #15.
- Nathan Bossart extracted that fix in #17, and Andres Freund committed it unchanged as 1fabec7d7c, confirming application to all supported branches in #18.
- Nathan Bossart then proposed stricter filesystem error handling in #19, whereas Bharath Rupireddy returned to softer errors and removed lstat calls in #20, reflecting an off-list discussion that was later clarified.
- Nathan Bossart showed that ReadDirExtended(LOG) could terminate scanning before all required files were synced in #21 and #23, leading Bharath Rupireddy to restore ReadDir() for logical mappings in #24 and refine comments in #26.
- Andres Freund clarified in #28 that the intended optimization was get_dirent_type(), preserving type checks while avoiding unnecessary system calls; Nathan Bossart implemented that approach in #29 and expanded it across backend directory scans in #32.
- Thomas Munro reviewed the expanded conversion in #33 and proposed moving reorder-buffer directory filtering into startup, which Nathan Bossart incorporated in #35 before rebasing in #36.
Jul - Sep 2022 · 8 entries
- Discussion resumed after several months with continuing disagreement over checkpoint error policy in #37 and #38, while Andres Freund considered the conversion reasonable in #39; Nathan Bossart also identified a possible transaction-ID wraparound risk from retained mapping files in #42.
- Thomas Munro prepared refreshed patches with commit messages in #43, but Tom Lane opposed the error-hardening patch and questioned behavioral changes embedded in the optimization in #45 and #47.
- Bharath Rupireddy separated optimization from error-policy changes in #53, dropping the hardening patch and adding WAL recycling; Thomas Munro caught an inverted regular-file comparison in #54, which Bharath corrected while merging the patches in #55.
- Nathan Bossart found that file-type lookup failures could now skip required fsyncs in #56, prompting explicit preservation of the old fall-through behavior in #57.
- Further review exposed unresolved reorder-buffer behavior and readability concerns in #58 and #60, so Bharath Rupireddy removed that portion in #61, which Nathan Bossart endorsed in #62.
- Michael Paquier accepted the final submission in #63, and Andres Freund withdrew an error-handling objection after clarification in #64, #65, and #66.
- Michael Paquier committed #61 as bfb9dfd937, reducing diagnostic verbosity, deriving the WAL filename from its dirent argument, and removing two API comments, as explained in #67.
- The thread ended with Michael Paquier confirming that the optimization had already landed in #70; the directory-fsync fix and reduced stat calls were complete, while the original softer-error proposal, later error-hardening proposal, and reorder-buffer conversion remained outside the committed work.