I agree with Sanne, Cassandra is very good at that use case due to the way they store things. OTOH, that's their best use case by far :)

We could think about a way to mix together:

- fine grained locking
- AtomicMap
- partial load

basically we could imagine some new primitive features for AtomicMap like:

- get subkeys between x and y
- add subkey (not requiring to load the AtomicMap)

That would get us a long way towards some of what MongoDB does wrt partial document update.