+simonhatch, any ideas why these aren't reproducing in bisect?

Note that these use the deprecated memory metrics, so we don't have traces breaking down memory usage.

One thing we've seen happen with memory metrics in cases like this is that sometimes a change shifts the timing of a memory dump slightly and the root cause might measuring before a GC vs after or something like that.

No ideas at the moment, tried kicking off a bisect with a wider range and no filter to see if that changes anything. Will wait for that to come back.

So widening the range seems to have worked, which would imply the alert is in the wrong spot? There's a bit of weirdness with the regression, the normal values for say "vm_working_set_delta_size/video.html?src_crowd.wav_type_audio" are in the 7000 range, while the regressed ones are in the 27,000 range. The alert is sitting on a value of around 20,000, kinda in-between the 2. The bisect can reproduce the same values as the dashboard, so hard to explain that stray value on the dashboard. Maybe something fired up in the background during that run?