As you can see, the groups are obtained by scanning the index on the groups table, and the aggregate is obtained by applying a seek in the index on the main table. The higher the density of the grouping set, the more optimal this plan is compared to the default strategy for the grouped query.
Just like we did earlier for the default scan strategy, let’s estimate the number of logical reads and plan cost for the seeks strategy. The estimated number of logical reads is the number of reads for the single execution of the Index Scan operator that retrieves the groups, plus the reads for all of the executions of the Index Seek operator.
The estimated number of logical reads for the Index Scan operator is negligible compared to the seeks; still, it’s CEILING(1e0 * @numgroups / @rowsperpage). Take Query 4 as an example; say the index idx_sid fits about 600 rows per leaf page (actual number depends on actual shipperid values since the datatype is VARCHAR(5)). With 5 groups, all rows fit in a single leaf page. If you had 5,000 groups, they would fit in 9 pages.
Plus some love for the APPLY operator. Read the whole thing.