Native R serialization/deserialization, which implies that all R objects are supported as key or value, without any conversion boilerplate code.This is the new default.
csv reader/writer -- somewhat experimental
Multiple backends (hadoop and local); local backend is useful for debugging at small scale; having two backends enforces modular design, opens up further possibilities (EMR, CUDA have been suggested), forces to clarify semantics.
Multiple tests of backend equivalence
Simpler interface for profiler.
Equijoins (rough equivalent of merge for mapreduce) part of the API