The policy seems sane. * They know when they can serve 0RTT from their cache safely because they can be reasonably certain if handling a cached request is side

The policy seems sane. * They know when they can serve 0RTT from their cache safely because they can be reasonably certain if handling a cached request is side effect free. * If connections to backend origins are reasonably persistent, there's not much latency reduction benefit from 0RTT compared to connections from consumer user agents.