On Feb 27, 2013, at 3:51 PM, jimtronic <jimtro...@gmail.com> wrote: > This seems to violate the strong consistency model doesn't it? If a write > doesn't succeed at a replica, it shouldn't succeed anywhere.
We don't provide a strong consistency model - we provide an eventual consistency model overall - though you can do some things to get other consistency levels for certain things. If a write doesn't succeed on a replica, I told you, that replica is either dead or asked to recover. The update stays in the cluster, so there is no need to roll it back. You need to figure out why the updates failed on the replicas, but the replicas were responding to other requests. That doesn't just happen out of the blue with no logs about how generally. - Mark