The Difference Between Logical and Physical Modeling
The term “physical” has two meanings that cause a lot of confusion when DBA’s are talking to each other. The first encompasses SQL objects: tables, columns, constraints, etc. The second uses the term as referring to data is stored in the hardware. The latter usage fits better with the original meaning that was coined in Codd’s Twelve Rules for a Relational Database Management System in Rule 8: Physical Data Independence. It basically states that the RDBMS should be allowed to store the data in any way or anywhere it pleases without changing the meaning or access.
While my RDBMS experience is specifically with Microsoft SQL Server, from what I have gathered, nearly all of the popular RDBMS tools allow data to be arranged physically in different manners, like sorted on disk, partitioned across different disk drives, non-unique indexes applied, etc. Realistically, no matter what term is used, there are three distinct models that you ought to consider:
• Logical: Implementation nonspecific requirements for data storage
• Implementation: RDBMS specific details needed to create a database that can actually store data
• Physical: Hardware model has no client meaning, only where/how data is stored
Logical The initial phase of implementation nonspecific requirements are commonly called the logical model. The goal is to get the requirements from the user/documentation, and end up with a relational database design that will represent near perfection. More or less the logical model is a highly structured document of what the database is attempting to do for the client. Ideally the design will be normalized, be void of implementation specific knowledge such as how SQL Server or Oracle would perform some task. The logical model is more or less documentation to the meaning of the client’s data structure needs.
Implementation The logical model directs programmers to create an RDBMS specific model. In this model, you adapt the logical model for implementation in the host RDBMS without dropping expected meaning. Exact datatypes, constraints, triggers, and whatever else may be needed to implement the logical model. This model should be reasonably void of performance details, like indexes to tune a query. This will be the job of the physical model.
Physical Once you have the implementation model complete, data is being stored correctly, you will have a set of objects sitting in an RDBMS with little or no data. The physical model is all about adapting the implementation to a combination of users and hardware. This model will start to form by developers as they tune their programs, but in production, the DBA can change the physical model and no client will be affected in any manner other than performance. In an ideal situation, even moving a database to a new server, os, platform, could be a completely transparent action as long as the RDBMS stays the same.
In the end, having three distinct models for the “meaning”, the “implementation”, and the "physical” structures make the reality of the data architect, the programmers, and the job of the DBA much easier.