We were looking at a use case of using remote server mode of JanusGraph with Cassandra. While exploring various pros and cons we cam across this https://groups.google.com/forum/#!msg/gremlin-users/6GYiHn3QR8g/81TFiW7yoTsJ .
Quoting the link " When you execute those locally against a distributed graph database, you are essentially pulling all those edges over the wire to conduct the traversal locally. That is very costly. In a way, that is what is happening when you run Titan locally and connect to a remote cluster (cassandra/hbase). With this setup, you are very limited in the kinds of traversals you can do (<200ms).". How valid is it in today's time and if it is valid what is the correct model to go for?