I have a multi-master setup:
1x loadbalancer (haproxy)
3x master
3x worker

With the update to kubelet 1.17.0 and still on the control-plane 1.16.3, the
cluster broke, all machines with kubelet 1.17.0 couldn't join the cluster
anymore.
On the other side, an upgrade was only possible with kubeadm 1.17.0

So I had to:
rollback whole cluster to kubelet 1.16.3
update to kubeadm 1.17.0 on all machines
update control-plane and all nodes to 1.17.0
update kubelet to 1.17.0

So as described in the kubernetes upgrade documentation.

Looks like we need to install the old and new kubelet in parallel until we
migrated the cluster.