I did think of that, but I realized that there wasn't too much additional value because map-reduce can be run cheaply enough on Hadoop clusters via EC2. I wonder if there are other ideas that this would open up.