I'm seeing the same failure as
MAPREDUCE-3462 in downstream projects running against a recent build of branch-23. MR-3462 modified the tests rather than fixing the framework. In that jira Ravi mentioned "I'm still ignorant of the change which made the tests start to fail. I should probably understand better the reasons for that change before proposing a more generalized fix." Let's figure out the general fix (rather than require all projects to set mapreduce.job.hdfs-servers in their conf we should fix this in the framework). Perhaps we should not default this config to "$fs.default.name"?