What a developer waits for, what got slower, and what to fix next. CI numbers are GitHub's own timings of every finished run of CI and Simulator CI, on main and on pull requests; local numbers are measured every day on a GitHub macOS runner.
Loading build timings...
Per head commit of a pull request, the longest job of its CI, Simulator CI and nightly dry-run runs, which is what the author waits for; a re-run counts the whole time from its first start. The second line is the same for commits pushed to main. The axis fits the lines; a single run above it is an arrow on the top edge, and the tooltip has its time. Queueing is not in GitHub's job timings, so real waits are a little longer.
Local builds, measured daily on a GitHub macOS runner (a shared virtual machine) at the tip of main (make build-bench): for the trend, not for the seconds on your Mac, which has more and faster cores. A point is the median of that day's runs and the band is fastest to slowest (a day with one run has no band). A day without a record breaks the line. The no-op build is a health check, not a wait: it is in the list at the top when it climbs above a few seconds. An orange ring round a point marks a row whose runs differed by more than half its median: the runner is shared, so read that point as a guess. A red cross marks a day a step failed on the runner.
The gate runs its cargo and python branches side by side, so its wall time is the slower branch plus a little set-up and the two lines meet on the days one branch is the long pole. The round dot is the whole gate; a hollow diamond or square round it names the branch that set it.
One line per job that finished green on main, taken from the data as it comes: a job that is new starts a new line, a job that was removed ends its line, and a run without a job leaves a gap, not a zero.
The same, for the simulator workflow.
The sum of every green job's minutes, which several optimisations traded wall time against: a split into parallel jobs lowers the wait and raises this. The total counts only commits that ran both workflows.
The share of finished CI and Simulator CI runs on main that failed, and that passed only after a re-run. Cancelled runs (a newer push superseding an older one) are not timed or counted in the share; the tooltip says how many there were.