The column index uses "Row IDs" which are always 64 bit integers. Since we're guaranteed that Row IDs are 64 bit integers, we do not have to pay the cost of copying/passing around RecordIds on the read path.

Eventually when we support clustered collections in column indexes, we will (very likely) map these Row IDs to the RecordId in a special part of the index.

We will benchmark to get a rough idea of the performance improvement from this change.