When you MapRed job yelling "Too many fetch-failures"

That happens when too many fetch failures happen on a specific reducer task node.

Three attributes could be check for this issure:

- mapred.reduce.slowstart.completed.maps = 0.80

allows reducers from other jobs to run while a big job waits on mappers

- tasktracker.http.threads = 80

specifies number of threads used by the reducer task node to serve output from mapper

- mapred.reduce.parallel.copies = sqrt(#of nodes) with a floor of 10

number of parallel copies used by reducers to fetch map output