While killing running MR job, hive doesn't clean up scratch directory (mapred.cache.files). So that, afterwards, scratch directory is left there in hdfs. HDFS name node doesn't know it and try to do lease recovery. while such instances happen more, it will eventually crash namenode.
The fix is to leverage hdfs clean up functionality. While creating scratch dirs, hive registers it to hdfs cleanup hook. While killing happens, hdfs will clean them up.