I have multiple Compute Engine VM's with roughly 30GB of Memory... Stackdriver Monitoring shows me with a sum aggregation of all Kubernetes Containers, that the overall memory usage is way more than that (see screenshot). Can someone explain how to read that metric? How can it be, that more memory is used than allocatable?
As I can't comment, I will leave an answer for you.
If you are sure it isn't related to auto-scaling, I would suggest making sure your filters are not to broad. This could be capturing other GKE nodes, or even GCE instances you have running within the sum.
Additionally, you could look at adding the Google Cloud Kubernetes Logging. The feature is still in Beta, but may be a bit more helpful.