Normalizing a Data Model in SQL Server 2005 and 2008

This chapter is from the book

Developer's Guide to Data Modeling for SQL Server, A: Covering SQL Server 2005 and 2008

Learn More Buy

This chapter is from the book

This chapter is from the book 

Developer's Guide to Data Modeling for SQL Server, A: Covering SQL Server 2005 and 2008

Learn More Buy

Data normalization is probably one of the most talked-about aspects of database modeling. Before building your data model, you must answer a few questions about normalization. These questions include whether or not to use the formal normalization forms, which of these forms to use, and when to denormalize.

To explain normalization, we share a little bit of history and outline the most commonly used normal forms. We don't dive very deeply into each normal form; there are plenty of other texts that describe and examine every detail of normalization. Instead, our purpose is to give you the tools necessary to identify the current state of your data, set your goals, and normalize (and denormalize) your data as needed.

What Is Normalization?

At its most basic level, normalization is the process of simplifying your data into its most efficient form by eliminating redundant data. Understanding the definition of the word efficient in relation to normalization is the key concept. Efficiency, in this case, refers to reducing complexity from a logical standpoint. Efficiency does not necessarily equal better performance, nor does it necessarily equate to efficient query processing. This may seem to contradict what you've heard about design, so first let's walk through the concepts in normalization, and then we'll talk about some of the performance considerations.

Normal Forms

E. F. Codd, who was the IBM researcher credited with the creation and evolution of the relational database, set forth a set of rules that define how data should be organized in a relational database. Initially, he proposed three sequential forms to classify data in a database: first normal form (1NF), second normal form (2NF), and third normal form (3NF). After these initial normal forms were developed, research indicated that they could result in update anomalies, so three additional forms were developed to deal with these issues: fourth normal form (4NF), fifth normal form (5NF), and the Boyce-Codd normal form (BCNF). There has been research into a sixth normal form (6NF); this normal form has to do with temporal databases and is outside the scope of this book.

It's important to note that the normal forms are nested. For example, if a database meets 3NF, by definition it also meets 1NF and 2NF. Let's take a brief look at each of the normal forms and explain how to identify them.

First Normal Form (1NF)

In first normal form, every entity in the database has a primary key attribute (or set of attributes). Each attribute must have only one value, and not a set of values. For a database to be in 1NF it must not have any repeating groups. A repeating group is data in which a single instance may have multiple values for a given attribute.

For example, consider a recording studio that stores data about all its artists and their albums. Table 4.1 outlines an entity that stores some basic data about the artists signed to the recording studio.

Table 4.1. Artists and Albums: Repeating Groups of Data

Notice that for the first artist, there is only one album and therefore one release date. However, for the fourth and fifth artists, there are two albums and two release dates. In practice, we cannot guarantee which release date belongs to which album. Sure, it'd be easy to assume that the first release date belongs to the first album name, but how can we be sure that album names and dates are always entered in order and not changed afterward?

There are two ways to eliminate the problem of the repeating group. First, we could add new attributes to handle the additional albums, as in Table 4.2.

Table 4.2. Artists and Albums: Eliminate the Repeating Group, but at What Cost?

We've solved the problem of the repeating group, and because no attribute contains more than one value, this table is in 1NF. However, we've introduced a much bigger problem: what if an artist has more than two albums? Do we keep adding two attributes for each album that any artist releases? In addition to the obvious problem of adding attributes to the entity, in the physical implementation we are wasting a great deal of space for each artist who has only one album. Also, querying the resultant table for album names would require searching every album name column, something that is very inefficient.

If this is the wrong way, what's the right way? Take a look at Tables 4.3 and 4.4.

Table 4.3. The Artists

Table 4.4. The Albums

We've solved the problem by adding another entity that stores album names as well the attribute that represents the relationship to the artist entity. Neither of these entities has a repeating group, each attribute in both entities holds a single value, and all of the previously mentioned query problems have been eliminated. This database is now in 1NF and ready to be deployed, right? Considering there are several other normal forms, we think you know the answer.

Second Normal Form (2NF)

Second normal form (2NF) specifies that, in addition to meeting 1NF, all non-key attributes have a functional dependency on the entire primary key. A functional dependency is a one-way relationship between the primary key attribute (or attributes) and all other non-key attributes in the same entity. Referring again to Table 4.3, if ArtistName is the primary key, then all other attributes in the entity must be identified by ArtistName. So we can say, "ArtistName determines ReleaseDate" for each instance in the entity. Notice that the relationship does not necessarily hold in the reverse direction; any genre may appear multiple times throughout this entity. Nonetheless, for any given artist, there is one genre. But what if an artist crosses over to another genre?

To answer that question, let's compare 1NF to 2NF. In 1NF, we have no repeating groups, and all attributes have a single value. However, in 1NF, if we have a composite primary key, it is possible that there are attributes that rely on only one of the primary key attributes, and that can lead to strange data manipulation anomalies. Take a look at Table 4.5, in which we have solved the multiple genre problem. But we have added new attributes, and that presents a new problem.

Table 4.5. Artists: 1NF Is Met, but with Problems

In this case, we have two attributes in the primary key: Artist Name and Genre. If the studio decides to sell the Juices of Brazil albums in multiple genres to increase the band's exposure, we end up with multiple instances of the group in the entity, because one of the primary key attributes has a different value. Also, we've started storing the name of each band's agent. The problem here is that the Agent attribute is an attribute of the artist but not of the genre. So the Agent attribute is only partially dependent on the entity's primary key. If we need to update the Agent attribute for a band that has multiple entries, we must update multiple records or else risk having two different agent names listed for the same band. This practice is inefficient and risky from a data integrity standpoint. It is this type of problem that 2NF eliminates.

Tables 4.6 and 4.7 show one possible solution to our problem. In this case, we can break the entity into two different entities. The original entity still contains only information about our artists; the new entity contains information about agents and the bands they represent. This technique removes the partial dependency of the Agent attribute from the original entity, and it lets us store more information that is specific to the agent.

Table 4.6. Artists: 2NF Version of This Entity

Table 4.7. Agents: An Additional Entity to Solve the Problem

Third Normal Form (3NF)

Third normal form is the form that most well-designed databases meet. 3NF extends 2NF to include the elimination of transitive dependencies. Transitive dependencies are dependencies that arise from a non-key attribute relying on another non-key attribute that relies on the primary key. In other words, if there is an attribute that doesn't rely on the primary key but does rely on another attribute, then the first attribute has a transitive dependency. As with 2NF, to resolve this issue we might simply move the offending attribute to a new entity. Coincidentally, in solving the 2NF problem in Table 4.7, we also created a 3NF entity. In this particular case, AgentPrimaryPhone and AgentSecondaryPhone are not actually attributes of an artist; they are attributes of an agent. Storing them in the Artists entity created a transitive dependency, violating 3NF.

The differences between 2NF and 3NF are very subtle. 2NF deals with partial dependency, and 3NF with transitive dependency. Basically, a partial dependency means that attributes in the entity don't rely entirely on the primary key. Transitive dependency means that attributes in the entity don't rely on the primary key at all, but they do rely on another non-key attribute in the table. In either case, removing the offending attribute (and related attributes, in the 3NF case) to another entity solves the problem.

One of the simplest ways to remember the basics of 3NF is the popular phrase, "The key, the whole key, and nothing but the key." Because the normal forms are nested, the phrase means that 1NF is met because there is a primary key ("the key"), 2NF is met because all attributes in the table rely on all the attributes in the primary key ("the whole key"), and 3NF is met because none of the non-key attributes in the entity relies on any other non-key attributes ("nothing but the key"). Often, people append the phrase, "So help me Codd." Whatever helps you keep it straight.

Boyce-Codd Normal Form (BCNF)

In certain situations, you may discover that an entity has more than one potential, or candidate, primary key (single or composite). Boyce-Codd normal form simply adds a requirement, on top of 3NF, that states that if any entity has more than one possible primary key, then the entity should be split into multiple entities to separate the primary key attributes. For the vast majority of databases, solving the problem of 3NF actually solves this problem as well, because identifying the attribute that has a transitive dependency also tends to reveal the candidate key for the new entity being created. However, strictly speaking, the original 3NF definition did not specify this requirement, so BCNF was added to the list of normal forms to ensure that this was covered.

Fourth Normal Form (4NF) and Fifth Normal Form (5NF)

You've seen that 3NF generally solves most logical problems within databases. However, there are more-complicated relationships that often benefit from 4NF and 5NF. Consider Table 4.8, which describes an alternative, expanded version of the Agents entity.

Table 4.8. Agents: More Agent Information

Specifically, this entity stores information that creates redundancy, because there is a multivalued dependency within the primary key. A multivalued dependency is a relationship in which a primary key attribute, because of its relationship to another primary key attribute, creates multiple tuples within an entity. In this case, John Doe represents multiple artists. The primary key requires that the Agent Name, Agency, and Artist Name uniquely define an agent; if you don't know which agency an agent works for and if an agent quits or moves to another agency, updating this table will require multiple updates to the primary key attributes.

There's a secondary problem as well: we have no way of knowing whether the phone numbers are tied to the agent or tied to the agency. As with 2NF and 3NF, the solution here is to break Agency out into its own entity. 4NF specifies that there be no multivalued dependencies in an entity. Consider Tables 4.9 and 4.10, which show a 4NF of these entities.

Table 4.9. Agent-Only Information

Table 4.10. Agency Information

Now we have a pair of entities that have relevant, unique attributes that rely on their primary keys. We've also eliminated the confusion about the phone numbers.

Often, databases that are being normalized with the target of 3NF end up in 4NF, because this multivalued dependency problem is inherently obvious when you properly identify primary keys. However, the 3NF version of these entities would have worked, although it isn't necessarily the most efficient form.

Now that we have a number of 3NF and 4NF entities, we must relate these entities to one another. The final normal form that we discuss is fifth normal form (5NF). 5NF specifically deals with relationships among three or more entities, often referred to as tertiary relationships. In 5NF, the entities that have specified relationships must be able to stand alone as individual entities without dependence on the other relationships. However, because the entities relate to one another, 5NF usually requires a physical entity that acts as a resolution entity to relate the other entities to one another. This additional entity has three or more foreign keys (based on the number of entities in the relationship) that specify how the entities relate to one another. This is how many-to-many relationships (as defined in Chapter 2) are actually implemented. Thus, if a many-to-many relationship is properly implemented, the database is in 5NF.

Frequently, you can avoid the complexity of 5NF by properly implementing foreign keys in the entities that relate to one another, so 4NF plus these keys generally avoids the physical implementation of a 5NF data model. However, because this alternative is not always realistic, 5NF is defined to help formalize this scenario.