With a persistent global state enabled, when a node that was previously part of a cluster rejoins it currently processes caches from the cluster state before the ones from the local state. This means that, if the cache configuration is incompatible, it will be overwritten with the one coming from the cluster.
When joining the node should perform compatibility checks between caches in the cluster state and the local state before proceeding with creating them. If a mismatch is found, it should fail fast.