Clients should pool both "real" (TCP) and application-level streams. Servers should respond to connections at the stream level. The fact that streams are mapped M:N to TCP connections should be invisible to the end-user.

Both clients and servers should be aware of service clusters. Clients should round-robin requests to identical services, and services should be able to proxy connections to other services transparently if they are experiencing failure or in the process of shutting down.