This is all still controlled through etcd and right now only does round-robin load balancing but there are plans to do other strategies like master elected, least loaded, etc. You can read more about this concept over on the docs[2].
The nice thing about this whole model is that the application isn't aware or involved in things about master-election and is only configured once at startup to talk to a local port.
[1] https://github.com/GoogleCloudPlatform/kubernetes/ [2] https://github.com/GoogleCloudPlatform/kubernetes/blob/maste...