Loading…
Presubmit CI health for the ROCm and NV lanes of openxla/xla and
jax-ml/jax, day by day. Every figure is the one
the nightly Markdown report
defines — this view only adds the time axis.
How fast the code CI builds actually runs is a separate dashboard,
perf-metrics.
Commit pushed → last job of the lane reports.
How long the critical job sat between being created and starting.
Mean seconds per commit, split into the phases every second belongs to. Means, not medians — medians of the parts would not add up to the whole. One panel per lane.
Distinct head commits per day. Weekends are genuinely quiet, not missing data.
Every job in the lane succeeded. Lane-wide and therefore volatile — a single broken job takes the whole commit out, so low-volume days swing hard.
How many tests one test job of the lane runs, from bazel's
Executed X out of Y tests: the largest Y of the day's sampled
jobs, cached results included. The largest, not the median, because a job that
fails early reports only what it reached — a median would track how the day's
jobs ended rather than the size of the suite.
openxla/xla runs two suites in one workflow, so its lanes are drawn one
line per job family — solid for the XLA jobs, dashed for the JAX ones.
A sample of each day's jobs, not all of them (sample_logs.py caps how
many logs it downloads), so this is what one job runs — never the lane's total
for the day, and a day with few samples can fall short of the real suite. Every
tooltip carries the number of jobs behind the figure; the table adds the median
beside it.
A percentile is a position in a sorted list, so the commits at or above it are what it is: move them and the tail moves. Per day and lane, the slowest three at or above that lane's p90 time to signal, newest day first, with the phase that took most of the wait. A PR listed twice is two of its head commits, each in the tail on its own.
Both lanes of a repo run on the same head commits, so ROCm only marks one this lane put in its tail while the other lane did not: the two disagree about the same commit, which points at the lane rather than at the PR. A row reading shared with NV is in both tails — a heavy PR, not a lane problem. Outside the other lane's tail does not mean fast there, only not among its own slowest tenth. The comparison is against the whole other lane, so filtering it off the charts does not change it.
Fork and export PRs run without a PR number attached, so those rows link to the workflow run, which names the pull request on its own page.
Every plotted value, as text.