Agreed. Something like this but based on RethinkDB or FaunaDB would be nice to see. Also if you chose to leverage these newer databases, you could leverage their in-built PUB-SUB support instead of having to using long-polling or placing socket.io/redis ontop of the REST API.
I'm definitely hoping to modularize the database. Thanks for the suggestions!
MongoDB is as durable as anything else if you aren't planning to use replica sets.
Under load it seems to have stale reads on replica sets but I believe that it is being targeted for 3.1. Even rock solid databases like Cassandra suffered from the same issue.