But w/o clustering on a denormalized grandchild.grandparent FK column, as Peter showed,
and cluster only on grandchild.parent, that's not going to access a mostly continuous range
of pages to fetch those all grandchild rows for that one parent. But probably 10 to 50 "row-clusters",
given the fan-out I mentioned earlier at the child-table level. Or am I missing something?
Are you using HDD (spinning) or SSD discs? Is you world strictly three levels: grand,parent,child? What tests have you done so far to compare clustered to non-clustered?