One issue with this is that the world is not at a standstill. The 90th percentile is a moving target. The average internet connection speed in India went up by 23% in 2015:

The work we invest in characterizing and optimizing for the bottom 10% will be very quickly obsolete. For a small team such as ours, this is a serious matter.

We should also remove our filter for the maximum of 60000 for Navigation Timing properties:

https://github.com/wikimedia/operations-puppet/blob/77d4eb8715/modules/webperf/files/navtiming.py#L114

This was added way way back. This shouldn't be needed with our percentiles and various other sanity filters (e.g. properties in order etc.)

That data source (stateoftheinternet) lacks distribution and a mobile/desktop breakdown. Typically we could be seeing a huge increase in bandwidth for desktop in metro areas because fiber is getting rolled out, while for rural areas mobile access isn't seeing any improvement.

To me the main questions are:

• if we focused on things from this angle, how would the highest priority tasks be different than if we didn't?
• how can energy spent on this not yield returns for higher bandwidth situations too?

I don't think we can really go wrong here, I would be surprised if there's anything we work on that makes a huge difference for the worst performing clients that wouldn't make performance better for everyone.

Cf. T125414: Investigate how connection speed varies by country