fail tolerance

• k8s wait for 5 minutes to mark a node as ‘dead’ in default
• if a node is marked as ‘dead’, the pods on the node will be rescheduled to other nodes
• drain: remove all the pods from a node and reschedule them to other nodes
• cordon: mark a node as ‘unschedulable’ so that no new pods will be scheduled to the node
• uncordon: mark a node as ‘schedulable’ so that new pods can be scheduled to the node but the original pods will not be rescheduled

cluster upgrade process

- k8s supports up to recent 3 minor versions ### kubeadm upgrade 1. upgrade kubeadm 2. command: kubeadm upgrade apply 3. upgrade kubelet and kubectl