Bugzilla

I summarized my findings in this post:
https://groups.google.com/a/mozilla.com/d/msg/perfteam/Zk9GN6kx2Fc/NU6QQby8DwAJ

I described my findings in this post:
https://groups.google.com/a/mozilla.com/d/msg/perfteam/Zk9GN6kx2Fc/NU6QQby8DwAJ

Summarized here:
In Bug 1564569 it was shown that the injection of new resources via `setTimeout()` (which may or may not occur before onload) is the source of the bimodal load performance in many, likely most, cases (the first 3 of 3 in tp6m).

You can see the difference between our current behaviour (top) and a prototype in which we run the `setTimeout()`s after load here:
https://imgur.com/NBOKnLe

Given that the prototype solution is not viable (hurts visual metrics, poor behaviour on some sites), two alternatives were discussed:

• For completeness, Markus raised the idea of creating our own metric, e.g. `onunbrokendependencyload` which would then exhibit less noise.
• Given the preference for visual metrics, the idea of focusing noise reduction on a visual metric (which may be more stable) was discussed and was favorably received