Yep! This is exactly how it works. Each app node runs a magicd which connects to an haproxy instance on localhost (connected to every machine in the database cluster), so when a Riak node goes down we don't miss a beat.