The count aggregation that the coordinator sends to the donor doesn't respond back until it finishes scanning the entire collection. This means that it has to start over from the beginning if it gets interrupted. If we split the count into a series of smaller ranges, then we only need to start from the last range that got interrupted instead of scanning from the beginning of the collection.

The most simple way to decide the splits is to use the chunk boundaries (useless if shard only has a single big chunk). One possible alternative is for a way to get the top level keys from the shard key index btree.