Weigh gilknocker Prometheus metric by duration #8558
Conversation
| raw_interval = dask.config.get( | ||
| "distributed.admin.system-monitor.gil.interval", | ||
| ) | ||
| interval = parse_timedelta(raw_interval, default="us") * 1e6 | ||
|
|
||
| interval = parse_timedelta(raw_interval) * 1e6 |
There was a problem hiding this comment.
Don't interpret a raw 1 in the config (which violates the jsonschema) as 1us. Crash instead.
| self._gilknocker.reset_contention_metric() | ||
| result["gil_contention"] = self._last_gil_contention = gil_contention | ||
| self.cumulative_gil_contention += duration * gil_contention |
There was a problem hiding this comment.
This is the key change.
Unit Test ResultsSee test report for an extended history of previous test failures. This is useful for diagnosing flaky tests. 27 files + 26 27 suites +26 10h 2m 13s ⏱️ + 10h 1m 49s For more details on these failures, see this check. Results for commit f919669. ± Comparison against base commit e16a7af. ♻️ This comment has been updated with latest results. |
|
Thanks, @crusaderky! It would be nice to be able to get more accurate GIL data via Prometheus. |
There was a problem hiding this comment.
Thanks, @crusaderky! Changes look good, assuming CI turns green.
stress test from #8557:
BEFORE
AFTER
CC @ntabris @milesgranger