model-writtenrun 6335 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Jul 2021 · 18 entries
- Gregory Smith reported that pgbench logs contained timestamps near 1970 instead of current Unix time, demonstrated the failure on Ubuntu and macOS, and advocated reverting the time refactoring until logging had adequate tests #1.
- Yugo Nagata traced the regression to using a monotonic clock for absolute timestamps and proposed restoring gettimeofday() at logging and progress-reporting sites #2, while Fabien COELHO proposed changing pg_time_now() centrally #4.
- Yugo argued that monotonic timing should remain for interval measurements despite system-clock adjustments and extended his targeted restoration to aggregate logging #8; Michael Paquier also identified missing seconds-to-microseconds conversion in aggregate accounting and urged combining the fixes #7.
- Kyotaro Horiguchi proposed calculating a single offset between the timer's epoch and Unix time #9, and Fabien adopted that approach with comments and simpler timestamp conversion #12.
- Fabien combined epoch translation with aggregate-interval scaling and forced final-aggregate output #15, after correcting an empty attachment #14; Yugo then fixed progress timestamps and aggregate timestamps expressed in the wrong units #16.
- Thomas Munro supplied a fallback revert that preserved thread synchronization #18, but Fabien objected to losing the improved TPS reporting #23, and Michael favored fixing the identified problems in this thread #24.
- Alvaro Herrera questioned landing immediately before the beta release #28, prompting Thomas to defer the planned fix #30; Andrew Dunstan and Michael subsequently supported allowing more time for a targeted correction #32 #33.
- Fabien added TAP tests for timed execution, progress, and aggregate logs #34; Michael tested the arithmetic successfully but challenged the reliability and documentation of the timing assertions #35 #37.
- Fabien first commented out the elapsed-time assertion #38, then adopted Alvaro's suggestion to run it inside a TODO block so failures could be observed without failing the suite #39 #40.
- Michael discovered that the existing log checker treated a regex object as a truth value instead of matching each line, and that another test incorrectly allowed only client zero #41; Fabien incorporated those fixes and corrected the proposed client-ID regex #42, while Andrew explained the Perl semantics #43.
- Michael extracted the existing-test fixes from #42 into c13585fe9e, added comments and stricter matching for the single-client case, and reported backpatching the correction through PostgreSQL 11 #44.
- Fabien resubmitted the remaining timestamp fixes and timing tests #45, then ensured that calculating the epoch offset used 64-bit arithmetic #49.
- Thomas identified that aggregate buckets no longer began at the whole Unix seconds printed in the logs #50; Fabien's shared-start-time adjustment #51 did not fully address that distinction, so Thomas supplied rounding against Unix time and revised final-aggregate handling #52.
- Fabien made aggregate accounting consistently use Unix-epoch time, restored the ignored timing test, and reiterated that Kyotaro and Yugo had contributed to the patch #54.
- Thomas demonstrated that PostgreSQL 13 discards the final partial aggregate and removed the proposed behavior change to keep the release fix narrowly compatible #56; Fabien reported successful checks and accepted the result despite preferring to retain that final bucket #57.
- Thomas then showed that delayed log writing could exceed the test's expected one-to-three aggregate lines #58; Fabien expanded the TODO block, but left that particular assertion outside it #60.
- Thomas committed the C portion of #60 as 0e39a608ed, with comment edits and without the new timing tests, closing the timestamp and aggregation bugs while preserving the monotonic timer #61.
- Fabien resubmitted the remaining timing tests with the aggregate-count assertion also inside TODO #62, leaving deterministic aggregate output and reliable timing coverage for further work.
Nov - Dec 2021 · 1 entry
- The test-only remainder subsequently failed to apply, and after requesting a rebase #63, Daniel Gustafsson marked it Returned with Feedback #64; the two fixes had landed, the fallback revert had not been used, and the new timing tests remained uncommitted.