In my view, Zookeeper's biggest shortcoming is that it doesn't support dynamic membership. That is to say that the Zookeeper nodes themselves are not self-organizing, you have to specify the ip addresses of all Zookeeper cluster members. Issue 107 [1] tracks this deficiency.
I don't mean to be totally negative. I really respect what the zk team has done and it's a fantastic tool that is often and unfortunately overshadowed by Hadoop. I think a small, reliable, coordination cluster providing atomic distributed primitives is an excellent distributed system architecture. Obviously Google and Yahoo have had success with it. I just recently learned that Ceph takes a similar approach.
Zookeeper feels like they spent a lot of time re-implementing Erlang's node and communication system in Java. In their design doc, they talk at length about how messages are guaranteed to be delivered in-order (and they seem very proud of this fact).
I think Zookeeper could be implemented in Erlang with 1/5th the code and 5x the reliability of what they have now.