It took me a week's worth of work to write a distributed ranker that will work on billions of ids in hundreds of sets; the algorithm is not obvious. On a single node you'd write a for loop that did a sort.
Those are the sort of obstacles that the hadoop programming model introduces. Again, if you haven't used it for real work, you'll be surprised at how few tasks are truly parallelizable without communication. The model is very simple, but accomplishing tasks that don't quite fit in the model is difficult.
Another example is a matrix mult. There is no way to segment the data to eliminate internode communication. It can be accomplished in an iterative fashion, but it's decidedly nontrivial. Or most optimization problems -- I don't think there even are known algorithms for no communication parallelization.