design.txt

Using the 0mq router and dealer pattern this can allow
rusix to have a frontend network and a backend replication
network.  

Frontend connects REQ -> Router on a port
On the backend Dealer -> REP on a different port and possibly network

Dealer should be where the consistent hashing happens

Clients would connect to any server in the cluster and that
Router socket would then route the clients request to a Rep
via a Dealer socket?

The upside to having clients connect to any server and
then using a dealer to route requests to their final location
is the client sees full bandwidth being used because storage
servers usually are pretty well outfitted with network bandwidth.
The downside is extra network hops.  If the client had knowledge
of the hashing then it could make 1 hop directly to the final
data location unless that data has moved.

Figure 38 in the 0mq book has a good example picture:
http://zguide.zeromq.org/page:all#The-Asynchronous-Client-Server-Pattern

Dealer -> Rep socket combination.  Dealer is an async socket 
that can talk to any number of Rep sockets.

Router is an async server that can talk to any number of Req
sockets.