Data Warehouse does not need a large number of non-clustered indexes?

It is a fact that non-clustered indexes improve the performance of data retrieval and we tend to add them mostly based on predicates used. As Data warehouses mainly design for data retrieval, many think that a data warehouse must contain non-clustered indexes as much as possible. In a way, it is true and required, however, we need to thoroughly analyze the requirements on indexing in data warehousing because generally non-clustered indexes do not help much with dimension and fact tables in data warehouses.

Non-clustered indexes generally give better performance on high selective queries. An example for a high selective query would be a query written for getting all customers for a given last name. Mostly attributes in a dimension table are used for pivoting in reports and typically hold few distinct values. Because of that, filters used in such reports are not much selective and will not get any benefits from indexes. However, parameterized reports looking for a particular value from an attribute would benefit as it is very selective. Therefore, when determining an index on an attribute, consider the reporting requirement similar to section example and then add.

It is common practice that adding non-clustered indexes on foreign keys. A typical data warehouse created based on Star-Schema has fact tables and fact tables have foreign keys linked to dimension tables. Do we need to create non-clustered indexes on these keys? Answer is “No”. The reason for that is, “star join query optimizations” feature in SQL Server (Enterprise only). This feature recognizes star join patterns and it uses bitmap filtered hash joins for joining tables in efficient way. Therefore indexes on foreign keys are not required unless a different edition is used. However, if the required dataset is smaller, indexes would be created on foreign keys as merge and nested loops joins give better performance than hash joins. Read more on that here: http://technet.microsoft.com/en-us/library/bb522541(v=sql.105).aspx

Considering above facts, we rarely add non-clustered indexes on both fact and dimension tables. In addition to that, the volume of records involved with the query (IO), the high cost for maintaining indexes, rebuilding every time data is loaded are other factors for not using non-clustered indexes in DWs.