Layering logically partitions the system into sets of subsystems with certain rules regarding how relationships can be formed between them. Layering provides a way to restrict inter-subsystem dependencies, with the result that the system is more loosely coupled and more easily maintained.

Consider the number and purpose of the layers carefully. Do not over-complicate the solution by defining more layers than are needed to meet the needs of the solution. More layers can always be added in the future to meet new requirements. Removing layers is not always as easy and may introduce risks into the project.

The criteria for grouping subsystems follow a few patterns:

• Visibility: Subsystems may depend only on subsystems in the same layer and the next-lower layer.
• Volatility:
  • In the highest layers, put elements that vary when user requirements change.
  • In the lowest layers, put elements that vary when the implementation platform changes (hardware, language, operating system, database, and so forth).
  • Sandwiched in the middle, put elements that are generally applicable across wide ranges of systems and implementation environments.
  • Add layers when additional partitions within these broad categories help to organize the model.
• Generality: Abstract model elements tend to be placed lower in the model. If not implementation-specific, they tend to gravitate toward the middle layers.
• Number of layers. For a small system, three layers are typically sufficient. A three-layer architecture of Presentation, Business, and Data layers is very common in information systems.  For a complex system, five to seven layers could be appropriate. For any degree of complexity, more than 10 layers should be viewed with suspicion that increases with the number of layers.

Failure to restrict dependencies according to Visibility criteria mentioned above can cause architectural degradation and make the system difficult to extend and maintain.

Allowable exceptions to the visibility rule include cases where subsystems need direct access to lower-layer services beyond the next-lower layer. Make a decision about how to handle primitive services that are needed throughout the system, such as printing, sending messages, and so forth. There is little value in restricting messages to lower layers if the solution is to effectively implement call pass-throughs in the intermediate layers.  This usage of a less strict rule on dependencies down through the layers is sometimes called a Relaxed Layered Architecture ([BUS96]).

Partitioning patterns

Within the top layers of the system, additional partitioning may help organize the model. The following guidelines for partitioning present different issues to consider:

User organization: Subsystems may be organized along lines that mirror the organization of functionality in the business organization (partitioning occurs along departmental or user role lines). This partitioning often occurs early in the design due to an existing enterprise model that is strongly partitioned according to the structure of the organization. This pattern usually affects only the top few layers of application-specific services and can often disappear as the design evolves.

• Partitioning along user-organization lines can be a good starting point for the model.

• The structure of the user organization is not stable over a long period of time because business reorganizations occur; therefore, it is not a good long-term basis for system partitioning. The internal organization of the system should enable the system to evolve and be maintained independently of the organization of the business that it supports.

Areas of competence and skills: Subsystems may be organized to partition responsibilities for parts of the model among different groups within the development organization. Typically, this occurs in the middle and lower layers of the system, and reflects the need for specialization in skills during the development and support of an infrastructure based on complex technology. Examples of such technologies include network and distribution management, database management, communication management, and process control, among others. Partitioning along competence lines may also occur in upper layers, where special competency in the problem domain is required to understand and support key business functionality. Examples include telecommunication call management, securities trading, insurance claims processing, and air traffic control, to name a few.

System distribution: Within any of the layers of the system, the layers may be further partitioned horizontally, so to speak, to reflect the distribution of functionality.

• Partitioning to reflect distribution of functionality can help you visualize the network communication that will occur as the system runs.

• Partitioning to reflect distribution can also, however, make the system more difficult to change if the deployment model changes significantly.

Secrecy areas: Some applications, especially those requiring special security clearance to develop or support, require additional partitioning according to security access privileges. Software that controls access to secrecy areas must be developed and maintained by personnel with appropriate clearance. If the number of people with this background on the project is limited, the functionality requiring special clearance must be partitioned into subsystems that will be developed independently from other subsystems, with the interfaces to the secrecy areas the only visible aspect of these subsystems.

Variability areas: Functionality that is likely to be optional, and therefore delivered only in some variants of the system, should be organized into independent subsystems that are developed and delivered independently from the mandatory functionality of the system.