Right now, we establish a new inner HBONE tunner when sending traffic across cluster. This somewhat defeats the point of HBONE and is expensive as it requires a new TLS/H2 handshake.
Doing this correctly is a little tricky since we do split horizon wds. So a naive approach of just throwing stuff in connection pools will make it so we always hit the same backend. Not good.
I think a good place to start it to establish up to O(N) hbone inner hbone tunnels before pooling where N is the weight of the split horizon wds.