Group by keys may be a mix of unique (or primary) keys and regular columns. In such cases presence of regular column won't alter cardinality of groups. So, if regular columns are not referenced later, they can be dropped from group by keys. Depending on operator tree may result in those columns not being read at all from disk in best case. In worst case, we will avoid shuffling and sorting regular columns from mapper to reducer, which still could be substantial CPU and network savings.