There were some subtle bugs in calculating certain Hive on Spark tuning parameters, which have pretty complicated logic. I saw this exception during development of the next release, but not in exactly your scenario. Still, I believe this should be fixed in the next minor release of Cloudera Manager (5.7). The bug is probably related to having hosts with different hardware or role assignments, which cause CM to generate multiple role config groups by default. Removing that host probably caused CM to generate identical configs for all NodeManagers, leading to them getting merged into a single role config group (the default group) and avoiding the bug. You can also probably work around this by simply not selecting the Spark service initially, and just adding it later. If using CM < 5.7 (ie all released versions today, 4/6/2016), then you'll need to manually apply the performance tuning if you want to use Hive on Spark (a beta feature in current releases) and see reasonable performance.
... View more