Kubernetes uses cAdvisor for monitoring pod usage of CPU, memory etc. In observation, it came out cAdvisor reporting RAM, which is not only hot and cold RAM as reported here: https://github.com/google/cadvisor/issues/913.

The issue is with kubernetes making use of the same for the purpose of implementing hard limit for quota and brings our POD as it exceeds memory consumption. This is a problem since the docker inside the POD is actually not consuming the memory reported. This can be verified using docker stats.

Is there a way to get around this? We want our pod to go down only when they are actually consuming more memory.