The key of its efficiency is an architecture represented by a set of isolated nodes that don’t know aware of each other. The client makes decisions on data management. This provides very high scale-out factor.

On the other hand, in regular cluster-based systems there is a master server and slave nodes. Slave nodes are very helpful as they help to distribute and speed up read requests. But as soon as write activities occur in a cluster, it’s a pain to broadcast all changes to all nodes. The deployment schema becomes inefficient.

As always, it’s a chain of trade-offs to be considered…

Enjoy – http://www.linuxjournal.com/article/7451