We should set yarn.nodemanager.disk-health-checker.max-disk-utilization-per-disk-percentage to a higher value in the mini cluster during tests (default is 90.0). This property determines the maximum percentage of disk space that can be utilized before the NodeManager goes to unhealthy and won't run any tasks. As-is, this means that if you have < 10% of your disk free, unit tests that submit Yarn jobs will fail because they can't run the job; and it's not always clear that this is the problem.
We should set this to a higher value such as 99.0. Our unit tests don't need that much disk space.