The minicluster runs Kudu without an explicit memory limit, and Kudu defaults to assuming it can use 80% of system memory. We should set a limit to avoid overcommitting system memory.
We need to keep the value low because aws test nodes (jenkins.impala.io) are small, see
IMPALA-5297. I tested a limit of 1gb but that wasn't enough – TestKuduMemLimits failed. We may try something a bit larger, or perhaps find other ways to reduce the requirements of minicluster tests.