Interesting, suggests a quick fix might be for the client to not use keepalive, or to selectively close connections that are very slow so as to start new ones. Potentially a much easier solution than writing a new transport.