> Atomic sections must be short: long atomic sections will break liveness, i.e. may cause the system to become unresponsive. > ... > Based on disk latency, we

> Atomic sections must be short: long atomic sections will break liveness, i.e. may cause the system to become unresponsive. > ... > Based on disk latency, we may target say a millisecond as duration before which to commit the current transaction. When the timer is reached, the transaction is delayed until all current atomic sections are completed; and (possibly after a grace period) new atomic sections are blocked from even being started, until after the transaction is committed. Maybe I'm reading this wrong, but the limitations on transaction duration seem to be disqualifying for real usage? If it's not possible to run an atomic transaction for longer than a few milliseconds without bringing the system down?