In partially upgraded clusters, whose config server is at version 4.0.5 and shards are at pre-4.0.3 (includes 3.6), sharding an empty collection with hashed sharding or performing map/reduce with outputting to an empty collection (with any kind of sharding) will end up placing all chunks on the primary shard of the collection.

This is not a correctness issue, but it means that the balancer will have to move non-empty chunks later on and thus defeats the purpose of pre-sharding the collection.