Resisting Normalization

If the four foreign keys of the fact table are tightly administered consecutive integers, you could reserve as little as 16 bytes for all four keys (4 bytes each for time, product, customer, and geography) of the fact table. If the four measures in the fact table were each 4-byte integer columns, you would need to reserve only another 16 bytes. Thus, each record of the fact table would be only 32 bytes. Even a billion-row fact table would require only about 32 gigabytes of primary data space.

With its compact keys and data, such a storage-lean fact table is typical for dimensional databases. The fact table in a dimensional model is by nature highly normalized. You cannot further normalize the extremely complex many-to-many relationships among the four keys in the fact table because no correlation exists between the four dimension tables; virtually every product is sold every day to all customers in every region.

The fact table is the largest table in a dimensional database. Because the dimension tables are usually much smaller than the fact table, you can ignore the dimension tables when you calculate the disk space for your database. Efforts to normalize any of the tables in a dimensional database solely to save disk space are pointless. Furthermore, normalized dimension tables undermine the ability of users to explore a single dimension table to set constraints and choose useful row headers.