I am currently conducting some tests that look at the result of uncompressed vs. page vs. row. The results are not final yet, but only a handful of tables compressed less than 50% out of my 54 table sample. I am impressed overall with the results at this level. It is important to point out two things about this preliminary test: 1. the overall savings at the database level was not as much as I expected, and, 2. when I compared before and after, I was careful only to compare data pages consumed [not reserved]. So my numbers looked great at the table level, but the overall savings was not what I had hoped for.

A few other observations...

My earlier testing indicates that the proc for estimating compression is not terribly reliable.

I believe that compression need only be applied to: 1) data warehouses, 2) tables larger than one sector read [typically 64 to 128 8k pages].