model-writtenrun 9420 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Fabien COELHO opened the thread with a replacement for the removed pgbench progress test, requiring only one progress line and skipping detailed log checks when runtime was abnormal to accommodate slow buildfarm machines #1.
- Tom Lane questioned whether conditional checks would conceal failures and proposed guaranteeing a progress report in pgbench itself; Fabien responded with a reporting helper, a closing-report rule, and an environment-controlled minimum runtime for TAP tests #3 #5.
- After rebases, Heikki Linnakangas challenged both the half-second closing-report rule and behavior enabled only during testing, favoring a fallback report when none had appeared and ordinary duration compliance instead #6 #7 #8 #9 #12.
- Fabien removed the environment switch and added a trailing wait-and-report loop, then restricted that loop to throttled runs with non-aborted clients after Heikki demonstrated unwanted waiting when every client failed #13 #14 #15.
- Heikki proposed fixing early termination inside the client state machine; after Fabien explained how a late transaction could otherwise start beyond the run deadline, Heikki supplied an alternative patch returning control to threadRun() #16 #19 #20.
- Fabien preferred keeping all state transitions inside doCustom() and supplied a separate proof of concept that also simplified statistics collection and script interruption rules, but these changes were not incorporated into the later progress-test revisions #21 #23.
- Following several inactive commitfests, Fabien rebased the earlier trailing-wait implementation, while Tom again questioned the value of timing-sensitive tests and asked whether the C changes fixed independently worthwhile behavior #22 #23 #25.
- Fabien explained that the C changes guaranteed testable progress output under scheduling uncertainty and rebased the combined proposal again in March 2019 #26 #27.
- Heikki committed only the reporting-function extraction from the latest combined patch as 9f75e37723, with naming, interface, placement, and buffer adjustments, explaining that the readability improvement stood independently of the remaining work #27 #28.
- Fabien reposted the uncommitted duration behavior and TAP tests using the newly committed helper #29.
- Heikki then criticized the test's cryptic positional arguments and limited output validation; Fabien added argument comments rather than adopting named arguments, and subsequently rebased after reindentation #30 #31 #32 #33.
- The remainder stalled, and Tom reiterated concerns about timing-dependent failures and changing benchmark behavior primarily to support tests; Fabien defended consistent duration and progress behavior under low transaction rates #35 #37 #38.
- David Steele announced that the remainder was returned with feedback, recommending tests without special exceptions in pgbench.c, and Daniel Gustafsson later completed the status change after finding it had not been recorded; only the reporting refactor had landed #39 #40.