Or your CNI implementation is made of rolled up turds and you lose a node or two from the cluster control plane every day.

(Large EKS cluster)