Based on rtt, downlink, hardwareConcurrency, deviceMemory and event the microbenchmark score, we should be able to create a "RUM profile" like "low-end device with a slow network"and surface perf metrics for that bucket as a graph. This would ensure that we are looking at our part of the responsibility in the RUM mix, by reducing the effect of the device and the network on the metrics.

We already collect all that data, it's an open question about whether or not this is something that could be done in the navtiming+prometheus paradigm with the right queries, or if we need a new pipeline for something like this, extracting the data periodically from hadoop.

Of course there's always the possibility of baking this into the collection point, but that makes updating the criteria (which we will have to do when the landscape evolves) a destructive process. Whereas applying the filters after the fact can allow us to still track the old buckets for reference when we update them.