As far as I can understand Alex was trying to avoid the scenario when user needs to bring 1Tb dataset to each node of 50 nodes cluster and then discard 49/50 of data loaded. For me this seems to be a very good catch.
However, I agree with Val that this may be implemented apart from store and user can continue using store for read/write-through and there is probably no need to alter any API. Maybe we need to outline Val's suggestion in the documentation and describe this as one of the possible scenarios. Thoughts? --Yakov
