When using spicedb in dispatch mode, my understanding is that each subproblem is assigned to a specific node, based on consistent hashing. What happens during a rolling upgrade, when nodes gets replaced? Or if they become unavailable for some other reason

Each SpiceDB node is consuming endpoint updates from the Kube API Server. During a rolling upgrade or a node gets replaced, the hashring gets reconfigured.

thanks. It doesn't use a sytem of backup nodes, if a node becomes temporarily unresponsive ?

if a node disappears from the ring, the corresponding subproblems get distributed across the other nodes. SpiceDB nodes don't store data, this is delegated to the underlying datastore.