I’ve just done this for a single node install, but I would assume the process would be the same under the hood for each individual node. This is a heavy assumption, but it would make sense that rancher would still write to /var/lib/rancher on each node, even though it’s part of a cluster. This would allow every HA node to fail at the same time, but the cluster to come back online after restoring from the persisted directory when they come back up.
I would look into etcd specifics (outside of Rancher specifically) as that’s the actual mechanism for storing the configs I think, or at least what I’ve seen in other Kubernetes configurations.
My reply for a single node install: