This has been observed with the SportResults, when launching the mapreduce job with a runtime argument "league=foo" (a partition that does not exist in the input fileset).
The correct status should be "FAILED".
Worse, when the mapreduce successfully completes, the apps/<app>/mapreduce/<mapreduce>/runs endpoint shows a "KILLED" status, instead of "COMPLETED".
This does not happen in standalone.
reproduceable with the PurchaseHistory example as well
COMPLETED runs were incorrectly getting marked as KILLED. Fixed this issue in the above PR. There is already another jira for failed mapreduce programs getting marked as COMPLETED : https://issues.cask.co/browse/CDAP-1745.
Marking this issue as resolved fixed.