Generally, most online transactional processing (OLTP) systems will perform well if they've been normalized to either 3NF or BCNF. However, certain conditions may require that data be intentionally duplicated or that unrelated attributes be combined into single entities to expedite certain operations. Additionally, online analytical processing (OLAP) systems, because of the way they are used, quite often require that data be denormalized to increase performance. Denormalization, as the term implies, is the process of reversing the steps taken to achieve a normal form. Often, it becomes necessary to violate certain normalization rules to satisfy the real-world requirements of specific queries. Let's look at some examples.
In data models that have a completely normalized structure, there tend to be a great many entities and relationships. To retrieve logical sets of data, you often need a great many joins to retrieve all the pertinent information about a given object. Logically this is not a problem, but in the physical implementation of a database, joins tend to incur overhead in query processing time. For every table that is joined, there is usually a cost to scan the indexes on that table and then retrieve the matching data from each object, combine the resulting data, and deliver it to the end user (for more on indexes and query optimization, see Chapter 10).
When millions of rows are being scanned and tens or hundreds of rows are being returned, it is costly. In these situations, creating a denormalized entity may offer a performance benefit, at the cost of violating one of the normal forms. The trade-off is usually a matter of having redundant data, because you are storing an additional physical table that duplicates data being stored in other tables. To mitigate the storage effects of this technique, you can often store subsets of data in the duplicate table, clearing it out and repopulating it based on the queries you know are running against it. Additionally, this means that you have additional physical objects to maintain if there are schema changes in the original tables. In this case, accurate documentation and a managed change control process are the only practices that can ensure that all the relevant denormalized objects stay in sync.
Denormalization also can help when you're working on reporting applications. In larger environments, it is often necessary to generate reports based on application data. Reporting queries often return large historical data sets, and when you join various types of data in a single report it incurs a lot of overhead on standard OLTP systems. Running these queries on exactly the same databases that the applications are trying to use can result in an overloaded system, creating blocking situations and causing end users to wait an unacceptable amount of time for the data. Additionally, it means storing large amounts of historical data in the OLTP system, something that may have other adverse effects, both internally to the database management system and to the physical server resources.
Denormalizing the data in the database to a set of tables (or even to a different physical database) specifically used for reporting can alleviate the pressure on the primary OLTP system while ensuring that the reporting needs are being met. It allows you to customize the tables being used by the reporting system to combine the data sets, thereby satisfying the queries being run in the most efficient way possible. Again, this means incurring overhead to store data that is already being stored, but often the trade-off is worthwhile in terms of performance both on the OLTP system and the reporting system.
Now let's look at OLAP systems, which are used primarily for decision support and reporting. These types of systems are based on the concept of providing a cube of data, whereby the dimensions of the cube are based on fact tables provided by an OLTP system. These fact tables are derived from the OLTP versions of data being stored in the relational database. These tables are often denormalized versions, however, and they are optimized for the OLAP system to retrieve the data that eventually is loaded into the cube. Because OLAP is outside the scope of this book, it's enough for now to know that if you're working on a system in which OLAP will be used, you will probably go through the exercise of building fact tables that are, in some respects, denormalized versions of your normalized tables.
When identifying entities that should be denormalized, you should rely heavily on the actual queries that are being used to retrieve data from these entities. You should evaluate all the existing join conditions and search arguments, and you should look closely at the data retrieval needs of the end users. Only after performing adequate analysis on these queries will you be able to correctly identify the entities that need to be denormalized, as well as the attributes that will be combined into the new entities. You'll also want to be very aware of the overhead the system will incur when you denormalize these objects. Remember that you will have to store not only the rows of data but also (potentially) index data, and keep in mind that the size of the data being backed up will increase.
Overall, denormalization could be considered the final step of the normalization process. Some OLTP systems have denormalized entities to improve the performance of very specific queries, but more than likely you will be responsible for developing an additional data model outside the actual application, which may be used for reporting, or even OLAP. Either way, understanding the normal forms, denormalization, and their implications for data storage and manipulation will help you design an efficient, logical, and scalable data model.