It turns out, it's faster to do it in this order for large clusters. Along the way: Do it in parallel