I think when you have de-normalized and big table (over 2 billions) then order of importance should change in my view. #2 (distribution) does move up. With de-normalized table you will never know for certain which predicates (WHERE) will be used most. Your calculation should consider worst case scenario which is full table scan and well distributed tables will pay off in most situations. In addition with big tables you usually end up with partitioning it with multiple levels and PI access with partitioned table is not useful.
↧