We've seen a fair number of instances where naive users process huge data-sets (>10TB) with badly mis-configured #reduces e.g. 1 reduce.
This is a significant problem on large clusters since it takes each attempt of the reduce a long time to shuffle and then run into problems such as local disk-space etc. Then it takes 4 such attempts.
Proposal: Come up with heuristics/configs to fail such jobs early.