2.3 Logical/Physical Coherence
When developing large-scale software, it is essential that our logical and physical designs coincide in several fairly specific ways at every level of packaging. Perhaps the most fundamental property of well-packaged software is that all logical constructs advertised within the collective interface of a physical module or aggregate — e.g., component, package, UOR (section 2.2) — are implemented directly within that module. Software that does not have this property generally cannot be described in terms of a graph where the nodes represent cohesive logical content and the directed edges represent (acyclic) dependencies on other physical modules. We refer to such undesirable software as logically and physically incoherent.
For example, Component Property 3 (section 1.6.3) states that if a logical construct having external bindage is declared in a component’s header, then that component is the only one permitted to define that construct. Recall from section 1.9 that, knowing the logical relationships among classes contained within separate components having Component Property 3, we can reliably infer physical dependencies among those components. Arbitrary .h /.cpp pairs that do not fully encapsulate the definitions of their logical constructs unnecessarily make reasoning about the design (and organizational) dependencies substantially more complicated (e.g., the misplaced definition of the output operator for the Date class in Figure 1-46, section 1.6.3). We therefore require that whatever logical constructs a component advertises as its own are defined entirely within that component, and never elsewhere.
The same benefits of logical/physical coherence that we derive from individual components apply also to library software at higher levels of aggregation. Imagine, for example, that we have two fairly large logical subsystems that we call buyside and sellside. Each subsystem is composed of several classes. For this discussion, let us assume that each of the classes is defined in its own separate component, and that the dependency graph of the unbundled components is acyclic. Figure 2-13 shows what often happens when subsystems conceived from only a logical perspective materialize. Although the logical and physical aspects of these systems coincide, the cyclic physical nature of the aggregate design does not scale, and is therefore unacceptable (section 2.2.24).
FIGURE 2-13: Cyclic physical dependencies (BAD IDEA)
Avoiding cyclic physical dependencies across aggregate boundaries is not only for the benefit of build tools, it also facilitates human cognition and reasoning. If all that were needed was to have two libraries where the envelope of component dependencies across aggregates was acyclic, then it would suffice to mechanically repartition these components as shown in Figure 2-14. But for software packaging to facilitate human cognition, in addition to being physically acyclic, the logical and physical aspects of a design must remain coherent.
FIGURE 2-14: Logical/physical incoherence (BAD IDEA)
Although the cyclic physical dependencies between the two libraries have been eliminated, the logical and physical designs have diverged. Now, neither logical subsystem is encapsulated by either physical library. As a result, our ability to infer aggregate physical dependencies from abstract logical usage — i.e., at the subsystem level — is lost. That is, if a client abstractly uses either the buyside or sellside logical subsystems, we must either know the details of that usage or otherwise assume an implied physical dependency on both libraries. Just as with cyclic physical dependencies, our ability to reason about logically and physically incoherent designs does not scale; hence, such designs are to be avoided.
Uniting the logical and physical properties of software is what makes the efficient development of large-scale systems possible. Achieving an effective modularization of logical subsystems is not always easy and might require significant adjustment to the logical design of our subsystems (see Chapter 3). As Figure 2-15 suggests, the reworked design might even yield a somewhat different logical model. Achieving designs having both logical/physical coherence and acyclic physical dependencies early in the development cycle requires forethought but is far easier than trying to tweak a design after coding is underway. Once released to clients, however, the already arduous task of re-architecting a subsystem will invariably become qualitatively more intractable, often insurmountably so.
FIGURE 2-15: Acyclic logical/physical coherence (GOOD IDEA)
Achieving logical and physical coherence along with acyclic physical dependencies across our entire code base is absolutely essential. In addition to ensuring these important properties, however, we will need a strategy that guarantees not just that the name of each architecturally significant logical and physical entity is unique throughout the enterprise, but that it can also be identified (and its definition located) just from its point of use, without having to resort to tools (e.g., an IDE). The following section addresses how we realize these additional goals in practice.