Forked from crbug.com/766556 . In a nutshell, regressions look like progressions and vice versa, because slower runs mean that the video (2.7s long) will loop before the end of the test, and while it's restarting it generates 1000x faster numbers than the common case, causing a heavy skew.
Slowing down the video 10x largely alleviates the problem (the test "should" run within 20s by construction), but possibly better would be to find a longer video asset.
Comment 1 by junov@chromium.org
, Sep 26 2017Owner: xidac...@chromium.org
Status: Assigned (was: Untriaged)