There are approaches to aggregation that give much of the value here.
If you can look at median vs 99th percentile, median will pretty much always be steady state, and 99th percentile will pretty much always be loading.

It's not as good as splitting by loading vs not loading though.
Ideally, TTI would give you a clear point to split on. We're currently investigating how effective this is for Expected Queueing Time. Once we've done that investigation, I think it would make sense to dig in here.

I'll ping this bug in a few weeks, when we've started getting some data.

Deep, any sense of how well the EQT post TTI metric is holding up on the bots since we fixed TTI?

I don't know where we have the EQT post TTI metric on the perf dashboard, but I looked at TTI in system_health.common_mobile and it still looks way too bimodal to me.

Here's what I was looking at: https://chromeperf.appspot.com/report?sid=079eee19e3a0990acf6438edcfdde650972bdb56a96b0aa03d185cc3ed1b1ad1&start_rev=561847&end_rev=562217

Maybe you'll have different intuitions about these graphs. 

The EQT specific split is issue 830882. Are there any other metrics where this split would be  useful?