We saw an issue where a custom InputSplit was returning invalid hostnames for the splits that were then causing the JobTracker to attempt to excessively resolve host names. This caused a major slowdown for the JobTracker. We should prevent invalid InputSplit hostnames from affecting everyone else.
I propose we implement some verification for the hostnames to try to ensure that we only do DNS lookups on valid hostnames (and fail otherwise). We could also fail the job after a certain number of failures in the resolve.