Huh, I'd have guessed that distributed synchronization of etcd (which goes over the network and has multiple events, and uses raft, which has unbounded worst case latency) would be the limiting step.