Michael
I suppose it's also possible that the B+tree implementation might even keep a counter of number of used nodes in each page to help it in its balancing operations and could then just read a 2 or 4 bytes off each page in the clustered index.
>If the table does have a
non-clustered index, SQL Server can scan the leaf-level pages of the index and count the number of keys on each page. This will give much better performance especially if one of the non-clustered indexes uses a very small key.