The current implementation performs the location of unsplitable collections in two steps. First, it gets all the chunks, and then it validates which chunk belongs to an unsplitable database. Since the current balancer issues unsplitable migrations one by one, the system must do a linear operation on the chunks to get each migration. 

It would be better to do an aggregation that is aware of the limit, and does not perform a full scan of the chunks before starting the join with collections (i.e. not hash joined based)