Mainly because the dataset is the ZFS version of a partition. In one pool, each dataset can have its own record size, compression settings (on/off or different

Mainly because the dataset is the ZFS version of a partition. In one pool, each dataset can have its own record size, compression settings (on/off or different levels, even different algorithms), be encrypted, have deduplication turned on... Its impossible for it to not be an "expensive operation".