I've heard of pushdown techniques including function, predicate and aggregate pushdown in distributed relational engines before.
Another interesting idea I read about (I can't find it anywhere online) was called "join zippering". Basically you first request the cluster to solve a join by querying and streaming the key columns from a join predicate back into the cluster itself to identify which nodes have matches and then streaming the results from each node in parallel, and doing the join in the stream.
This is hard stuff but so cool too :)