Seems like a leaky abstraction. Did their interfaces have a way to account for differences in latency and failure modes, or was that hidden?