There is a very nice post on http://codahale.com/you-cant-sacrifice-partition-tolerance/.
One complications to that argument is replication implemented with group communication, which is a reasonable approximation to a system with both availability and consistency. (Just thinking aloud, as these stuff needs lots of thinking to be sure). Now what happen in group communication is that if a node fails, it creates a new group excluding failed node, and although the system operation has to wait till new group is defined (rather repair the existing node ), after the repair, things will work fine given that you have enough replicas left. So does group communication contradict or works around above assertion?