In a multi-tenant cluster like the Phacility cluster, it would be much easier to administrate a small number of notification servers than one-per-instance, and the capacity of a single server should greatly exceed the average demand of an instance.

We can support this by including a mode where servers partition client and administrative requests by "Host" header. Specifically, if you connect to the server and send instancea.phacility.com as the Host, you join a pool with other users on instancea. Likewise, administrative commands sent to that Host affect only the subset of users on that partition.

We can then put some dumb load balancing in front of it for now and some smarter load balancing later and get notification support relatively cheaply without needing to manage additional services per instance.