The key/value model embraced by NoSQL databases (Scalaris, Voldemort, Tokyo Cabinet, etc) is the simplest and easiest to implement but inefficient when you are only interested in querying or updating part of a value.

This article (http://seattleweb.intel-research.net/people/lamarca/pubs/pap...) coming out of Intel argues that it is also difficult to implement more sophisticated structures on top of a distributed key/value. The author's main point is that a few specialized applications can and have been built on a plain distributed key/value store, but most applications have ended up having to customize the key/value store's internals to achieve their functional or performance goals.

From the little bit I have read about Membase it looks well positioned to bring simple distributed key/values stores to the next level and back into the lime light.