We use map jvm opts for root vertices and reduce jvm opts for intermediate and leaf vertices. For a simple order by or skewed join the Sampler will be a root vertex and Partitioner will be a intermediate vertex. If the map opts and reduce opts are different containers will not be reused and there will be lot of resource wastage when the reduce opts specify higher memory than map which is usually the case. The partitioner does not do much and just does identity mapping. So we should be good using map opts there.