About 10 years ago, I worked on a project where I had to develop some sort of elaborate chain of map/reduce jobs because "big data" and "Hadoop." We were processing about 10 megabytes for each run. Most of the processing was consumed in job scheduling / overhead.