The two are linked. Need to change the configuration of a fleet? That's going to be restarting every instance of the process. Update an extension model? Same. Load balance? You'll need one that understand RADIUS or clients will suffer because of incorrect session affinity. Client with dramatically different loads? Better put them on different clusters. Somebody had a power outage? Better have 10x capacity on hot standby for the load.

And on and on.

A stateless compute model with separation between the packet handling and the authentication logic solves pretty-much all of it.