Currently, socket timeout's are controlled by using the current context's time remaining (if any). This means we are tying a user's request directly to whether we kill a socket. We should come up with a different strategy so that we don't tear down sockets artificially and throw away the potentially expensive setup we've already done.