Rebalance Inefficiency increasing the number of vbuckets on a node that should be decreasing

I noticed the problem at a customer today and reproduced it myself on 1.8.1.

Take a 6 node cluster, add 12 more but stop rebalance almost right away (moving only a few vbuckets). Let things stabilize and restart the rebalance. If you watch the vbucket movement on all of the nodes, you'll see at least one of the original 6 nodes actually start to increase its vbucket count. And then it starts to go down as expected.

I reproduced this with a completely empty cluster (one bucket). But it makes a much bigger deal in a customers environment when we're trying to grow the cluster...and now vbuckets are going back on nodes that are already under pressure.