The Jepsen YARN tests sporadically fail because TM containers are exceeding their virtual memory limits:
By default YARN enforces a virtual memory limit of 2.1 times the requested physical memory. However, in my experiments, the virtual memory of a JVM process running the ClusterEntryPoint (without submitting job) is already in the region of 3.3 GB. Hence, the virtual memory enforcement should be disabled.
- yarn.nodemanager.vmem-check-enabled is false in yarn-site.xml