Riak is probably a better fit in circumstances where capital spending is not a limiting factor, but it was interesting to see some of the behaviours under pressure in a resource-limited environment, some of which I personally found eye-opening, especially the regressions related to leveldb.
Problems with leveldb with riak in this case can probably be related directly to the number of vnodes. 256 pretty much guarantees you will need a lot of resources with the leveldb backend.
If you have a high number of vnodes so you will be able to scale out a lot, the upshot is you will need to start “scaled up” on your smaller number of nodes.
The other option is the keep your number of vnodes low which will limit your ability to scale out.