Current implementation of scale in/out is based on fault recovery, which can cause long latency. We need to make it more efficient in the later version.