I have an elastic cluster with two client nodes. When I look at the /es/_nodes/stats/os metric, I notice the used_percent for each of them is around 95% even though elastic and java are the only thing running on the server. Is this expected?
What version of Elasticsearch are you using? How much total memory is available for each machine, and what percentage of that have you allocated to Java?
You may want to consider Daniel's comment on this older forum topic: Client node crash with OOM exception