Will the second method not be better to equaly distribute edges over multiple machines in a dgraph cluster? With the first approach you will always have a type predicate for every node, which will probably also increase the index size / slowing down lookup performance of that “large” type predicate index?