zookeeper needs 3 nodes to support 1 failure

hi all

a quick question that you would find stupid but...

Looking at Zookeeper, you need 3 nodes to support 1 node failure.
Looking at in memory data grid (coherence co...), with 2 replicas, you
can still continue working.

So "basically", you need less replica to support the same amount of
failure.

I am not saying one is better than the other ;), cause we are using both,
but does some of you know why the difference?
I assume this is related to the Zab protocol but if you have more details.
What are the guarantees zookeeper provide and a data grid does not provide?
the lock? a sequential consistency?

Many thx for your answers

oliv/