It’s a non-goal (currently) for us to support generalized map reduce. We consider that a feature of a compute substrate (like Spark, distributed Dask, Hive, etc) rather than the orchestration substrate.
We may look at supporting some very coarse-grained parallelism (e.g. firing N runs off simultaneously that don’t need to coordinate in a real way) but I would suggest pushing any sort of map reduce operation down to a layer like spark as @tseader suggests