this is too broad - it's mixing the country-code to country-of-location
distance, plus a bit of the latency heuristic.

we're randomizing the last step.

it can be useful for getting a random sample of candidates (with low
load) so that the client can then measure their real-world latency.

for a real world usage, I think picking one single location, filtering
by lower-load, plus a region-based assignation, will work better.

(i.e., provider already knows what are the gws reserved for, and can
tweak manually).