The Power of Randomization: Distributed Submodular Maximization on Massive Datasets