I know that k8s has default Hard Eviction Threshold memory.available<100Mi. So k8s should evict pods if thresholds exceed. In these conditions can pod provoke SYSTEM OOM? When I talk about SYSTEM OOM I mean such situation when Linux starts to kill processes randomly (or not almost randomly, doesn't matter). Lets suppose that other processes on node consume constant amount of memory. I hope that k8s watches pods and kills them before the threshold exceeded. Am I right?
Yes, very yes. Eviction takes time. If the kernel has no memory, oomkiller activates immediately. Also if you set a resources.limits.memory then if you exceed that you get an OOM.