Experiment Compare values could adjust for TTFB variance, at least optionally

[scottjehl]

Often, there's network variance between experiment and control runs that shows up in TTFB. Being additive, the TTFB diff can end up rippling into the diffs of later metrics that can have much smaller diffs if it weren't for the TTFB variance.

One approach to fixing this will be the longstanding idea of normalizing TTFB between compare runs, which we should still aim to offer.

But another, and perhaps additional thing we can do in Experiment results would be to detect TTFB variance and offer a way to view the experiment results with TTFB removed from chronological metrics before comparing them. That'd mean we could more reliably see diffs in the metric changes that we offer on that page.

here's an experiment with ttfb variance: https://www.webpagetest.org/video/compare.php?tests=220906_AiDcYK_EPH,220906_AiDcX7_EPG

and its results:

Here's those same results adjusted for TTFB:

Offering this via query string and a link, non-default, would be a nice way to start. A message that says:

We detected some variance in response time between your experiment and control runs, which might make other metric changes seem more significant than they really are. View this comparison with TTFB differences removed