Right now, we're not leveraging anything fancy, just using the stdlib http.Transport, which ultimately will pile all HTTP/2 streams onto one upstream connection on one IP address.
This is less than ideal for us in production capacities, and we'd want to pool connections and help spread the load around, and pool N per IP to spread across our infrastructure more friendly.
See:
Also need to flip the flag and open source go-upstream so we can use it here.