Alert on perf dashboard showing up weeks after occurrence |
||
Issue descriptionThis alert just showed up on the perf dashboard untriaged list today (July 25) It says: Time added: 2018-07-14T11:37:48.000Z https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICQ7d79qAsM Why did it take so long for the alert to show up on the perf dashboard? By July 17 10 more data points had been added.
,
Jul 26
The segment_size_after for that alert (from the Datastore Entities tab in pantheon) is 43, so it required 43 more points to detect the step change. We can play with the options here to try to lower that number: https://chromeperf.appspot.com/debug_alert?test_path=ChromiumPerf%2FWin+7+Perf%2Frendering.desktop%2Fpercentage_smooth%2Fbouncing_gradient_circles&rev=575108&num_before=1000&num_after=3000&config= I tried a few options, but I'm not sure which might help find_change_points require fewer points. I wish debug_alert displayed the default values of the options.
,
Jul 26
Thanks, Ben! Adding Dave since he knows about the alerting algorithm. Also adding benchmark owners Sadrul and Victor, in case they want to tweak the settings. Note that this was the only alert I know of that fired so late.
,
Jul 27
The default min_steppiness is 0.5. Reducing it to 0.4 about halved the number of points required to detect this regression. https://chromeperf.appspot.com/debug_alert?test_path=ChromiumPerf%2FWin+7+Perf%2Frendering.desktop%2Fpercentage_smooth%2Fbouncing_gradient_circles&rev=575108&num_before=1000&num_after=3000&config=%7B%0D%0A++%22min_steppiness%22%3A+0.4%0D%0A%7D |
||
►
Sign in to add a comment |
||
Comment 1 by sullivan@chromium.org
, Jul 25