Is round-robining between each TLS conn the optimal scheduling approach?
Right now when a relay has n conns that want to write, we write fairly between each of them.
But we could imagine other scheduling approaches, like weighting the attention we give each conn by its capacity in the consensus, or weighting the attention we give it by how many circuits we have on that conn.
We can even imagine more complex approaches, like adding up the ewma weights on each conn, or choosing the conn with the quietest circuit (ewma-wise), or others.
We should make patches for some of these options and get simulations going.