etcd stores every object in the cluster, so a snapshot is the backup to take before risky changes. On a kubeadm control plane, etcdctl snapshot save FILE needs the endpoint and the client certificates in /etc/kubernetes/pki/etcd (--cacert, --cert, --key); without them etcd refuses the connection.
kubeadm upgrades one minor version at a time, in a fixed order. On the control plane: install the new kubeadm package, run kubeadm upgrade plan to see the target, then kubeadm upgrade apply VERSION, which refuses a target newer than kubeadm itself. Then each node in turn: drain it, run kubeadm upgrade node, install the matching kubelet package, restart the kubelet, and uncordon it. The kubelet may lag the API server by up to three minor versions, so this can proceed node by node without downtime.
In this cluster the machines are modeled: packages install instantly and the control plane's own kubelet upgrade is left out.
Plan, apply on the control plane, then node by node.