Database normalization

Normalization is grounded in the theory of functional dependencies, where a column's value is determined by another column or set of columns. The first normal form (1NF) requires that each table cell contain a single value and that each record be unique. Second normal form (2NF) eliminates partial dependencies on a composite key, and third normal form (3NF) removes transitive dependencies. The Boyce–Codd normal form (BCNF) is a stricter version of 3NF that handles certain edge cases where every determinant must be a candidate key. Higher normal forms, such as 4NF and 5NF, address multi-valued and join dependencies, but are rarely used in practice. The process of normalization is often taught as a step-by-step decomposition, but in real design, it is applied iteratively and often combined with denormalization for performance.

Edgar F. Codd introduced the relational model in his seminal 1970 paper, and he later defined the first three normal forms in 1971–1972. Raymond F. Boyce and Codd jointly developed the Boyce–Codd normal form in 1974 to address anomalies not covered by 3NF. Ronald Fagin contributed the fourth and fifth normal forms in the late 1970s, and later the domain-key normal form (DKNF) as a theoretical ultimate. Chris Date and Hugh Darwen have been influential in promoting and refining normalization theory in textbooks and standards. The concept of normalization has also been extended to object-oriented and NoSQL databases, though with different constraints and goals.

In practice, normalization is a key step in database design, often performed after entity-relationship modeling. It helps ensure data consistency and reduces storage waste, but it can lead to increased join operations, which may degrade query performance. Database administrators often denormalize selectively—for example, by adding redundant columns or precomputed summaries—to meet performance requirements in read-heavy systems. Normalization also plays a role in data warehousing, where star schemas and snowflake schemas represent different degrees of normalization. Tools like database design editors and automated normalizers exist, but human judgment remains essential. The trade-off between normalization and performance is a classic topic in database literature, and modern systems often use indexing and materialized views to mitigate the costs.

Beyond the standard forms, there are several niche concepts. The domain-key normal form (DKNF) is a theoretical ideal that is rarely achievable but serves as a benchmark. The concept of 'normalization' also applies to database schemas in temporal databases, where time-varying data introduces additional dependencies. In the context of NoSQL, 'normalization' is often reversed, with denormalization being the norm for document stores. A lesser-known fact is that Codd originally proposed a 'relational normalization' process that included a 'synthesis' approach, which was later formalized as the 'Bernstein synthesis algorithm' for achieving 3NF. Another edge case is the 'lossless join' property, which ensures that decompositions do not lose information—a critical requirement for any normalization step. Finally, some database systems, like PostgreSQL, offer tools to analyze and suggest normalizations, but they are not fully automated.

Functional dependency

A constraint between two sets of attributes in a relation, where one set determines the other.

Normal form

A level of database schema quality that reduces redundancy and anomalies.

Denormalization

The process of intentionally introducing redundancy to improve query performance.