Migrating etcd from v3.4 to v3.5 without cluster downtime

We're planning an in-place etcd upgrade on a 3-node production cluster (k8s 1.28, ~40k keys). The docs say rolling upgrades are supported, but I've seen mixed reports of leader election storms during the minor version swap. What's your actual playbook? Do you bump one member at a time and wait for raft sync, or is there a value in temporarily adding a 4th node to absorb the transition? Curious about real-world migration patterns, not just the happy path.