It gets sent to the cluster (http://www.datastax.com/dev/blog/atomic-batches-in-cassandra-1-2)
- a batch is accepted
- a batchlog is written locally and to two other replicas
- If the batch does not complete for some reason then it’s retried on the other two replicas
This means that a batch log will almost certainly succeed, but using a circuit breaker for retries (not rollbacks) is still best practice regardless of if a batch log is used or not.