Using clustered indexes

Although indexes can dramatically improve the performance of statements that search for a specific range of key values, two rows appearing sequentially in the index do not necessarily appear on the same table page in the database.

You can further improve a large index scan by declaring that the index is clustered. Using a clustered index increases the chance that two rows from adjacent index entries will appear on the same page in the database. This can lead to performance benefits by reducing the number of times a table page needs to be read into the buffer pool.

The existence of an index with a clustering property causes the database server to attempt to store table rows in approximately the same order as they appear in the clustered index. However, while the database server attempts to preserve the key order, clustering is approximate and total clustering is not guaranteed. So, the database server cannot sequentially scan the table and retrieve all the rows in a clustered index key sequence. Ensuring that the rows of the table are returned in sorted order requires an access plan that either accesses the rows through the index, or performs a physical sort.

The optimizer exploits an index with a clustering property by modifying the expected cost of indexed retrieval to take into account the expected physical adjacency of table rows with matching or adjacent index key values.

The amount of clustering for a given table may degrade over time, as more and more rows are inserted or updated. The database server automatically keeps track of the amount of clustering for each clustered index in the ISYSPHYSIDX system table. If the database server detects that the rows in a table have become significantly unclustered, the optimizer will adjust its expected index retrieval costs.

If you decide to make one of the indexes on a table clustered, you need to consider the expected query workload. Some experimentation is usually required. Generally, the database server can use a clustered index to improve performance when the following conditions hold for a specified query:

Many of the table pages required for answering the query are not already in memory. When the table pages are already in memory, the server does not need to read these pages and such clustering is irrelevant.
The query can be answered by performing an index retrieval that is expected to return a non-trivial number of rows. As an example, clustering is usually irrelevant for simple primary key searches.
The database server actually needs to read table pages, as opposed to performing an index-only retrieval.

Discuss this page in DocCommentXchange.