Under the following conditions:
1. A request arrives for a session with an unknown route
2. A request arrives for a session with a known route in an error state, or if the intended worker is unresponsive and needs to failover
... the failover target must be chosen in a deterministic way. This is meant to mitigate the situation where concurrent requests failover to different nodes and cause cluster-wide contention for the session. Given a consistent ordering of workers (e.g. lexographically), the failover target should be chosen by something like: hash code of session id % active worker count.