Dependency Inversion demands that abstractions belong to callers. This isn’t an immediately intuitive choice. I’ll break down the choices for where our abstractions can live and how each choice impacts kinds of dependency.
Dependence often get’s talked about as if there were only one kind. A component depends on another or it doesn’t. But that’s not true. There are many kinds of dependence (or coupling).
The main kinds of dependency I’ll be considering here are
- Design-time dependency: Dependent code won’t compile if it doesn’t have a reference to the dependency code
- Runtime-time dependency: Dependent will execute the dependency code at runtime. May or may not involve design-time dependence
- Semantic coupling: The dependent code uses knowledge of concepts the dependency defines. Could be modeled concepts, state, side-effects, implicit (or explicit) assumptions, or other knowledge.
I’ll also often refer to packages for groups of code that get would get distributed together. In .NET-land this could mean nuget packages or projects (which decide what gets compiles into one DLL). The same general ideas apply to code groupings across languages.
Sample of Runtime without Design-time Dependence
It’s not immediately intuitive that code can depend on other code at runtime, but not at design time. However, that’s one of the main motivations for inheritance, and more broadly for dynamic polymorphism.
Consider this quick example. Which implementation of
Baz depend on?
It’s a trick question. Baz doesn’t require any concrete
IFoo implementations. It has no design-time dependence on specific IFoo implementations. Baz could be compiled with no IFoo implementations at all. However, a specific implementation of IFoo will need to passed into Baz in order for it to run. Baz has runtime dependence on IFoo, but not a design-time or compile-time dependence. The concrete IFoo that Baz depends on can be decided independently for each instantiation of Baz.
Where Abstractions Live
The most straightforward dependency approach is dependency retention. That is, no abstraction. A concrete component directly references another concrete component.
For example, this class B directly depends on class A.
In diagram form
The caller is coupled to it’s dependency at design-time and runtime. It also probably has a degree of semantic coupling since it is fully aware of the dependency’s design choices, like data contracts. The caller must directly map it’s information into the dependency’s contract. The dependency thus shapes decisions in the caller.
Changing what dependency the caller uses requires code changes to the caller.
Desire to swap out dependency implementations generally leads to abstract types (e.g. interfaces). Testing, for example, is much easier if class dependencies can be swapped for faster and more observable test implementations. When I say abstract types, I mean types that determine the contract a component accepts but not the implementation. These types could require specific methods or set the shape for input data and output data.
All types in this code example could be called “abstractions” and part of the contract.
I quoted abstractions because, as Mark Seemann said, abstractions are about semantics. A poorly defined interface can define a shape without truly hiding internal expectations of the dependency. We’ll see more of this in the next section.
Abstraction Lives with Implementation
Suppose we group these “abstractions” with their implementations. Seems intuitive.
Now we can technically swap implementations, but the reality tends to not be so nice.
For one, we haven’t separated design-time and runtime dependency. Since the abstraction and implementation live together in one package, they are also always depended on together.
Alternative implementations have to live in the same package or reference the package that also includes the original implementation. The original implementation is often explicitly not desired when you’re creating an alternative implementation.
Implementation-oriented abstractions tends to violate Interface Segregation. The abstractions also don’t know what they’re being used for, so they have to try to be generic and satisfy many callers. The caller ends up depending on concepts they don’t need. Further, the abstraction can’t focus on a particular use-case so it accumulates too much responsibility and alternatives become too complex to practically create. The also abstraction usually becomes a “header interface” that just reflects the concrete implementation, resulting in very little conceptual decoupling.
Shared Abstraction Reference
The coupling of design-time and runtime dependence can be broken by moving the abstractions into their own package.
A separate abstraction package does separate design-time and runtime dependency. However, it doesn’t resolve semantic tensions. The abstractions still need to orient their semantics to something. Here’s the main options
- implementation-owned semantics -> Forces the caller to still shape itself around the implementation’s semantics. Suffers the disadvantages already discussed for Implementation-oriented abstractions. Also, offten results in the caller depending on sneaky assumptions of the implementation and undermining the decoupling
- caller and implementation semantics -> Suffers all the disadvantages of Implementation-oriented abstractions, but also leaks semantic coupling between consumers of the dependency. This causes widespread coupling in a system that is hard to track and thus leads to defects.
- the abstractions own themselves -> Abstractions on their own terms can be ok when they model system-wide common concepts. For example, you probably don’t want every service defining a custom logging interface. I’d recommend avoiding these until you have a firm handle on Dependency Inversion.
- caller-owned semantics -> There’s no need for a separate abstractions package. They can live with their caller. It’s not immediately intuitive how to accomplish this, but we’ll get there.
A major danger of a shared abstraction assembly is just having an “abstractions” package for most abstractions in the system. This is (usually) about the same as making all your data global. It makes reference easy, but then everything knows about every abstraction and any change to those abstractions is probably unsafe and widespread. Changing any abstraction requires considering the entire system. Yuck.
Caller-owned abstractions resolve semantic tensions while maintaining design-time and runtime dependency separation.
Abstractions oriented to the caller know exactly how they’re going to be used, so they can define only what the caller needs and in the exact terms the caller needs. There is no tension trying to balance design forces of multiple consumers. It is much easier to keep abstractions focused, small, and thus swap different implementations readily.
The confusing part is now your dependencies rely on the services that uses them. How can that work?
It works because the kinds of dependence are different. The caller has only a runtime dependence, while the implementation has a design-time dependence.
This does mean that the dependencies are semantically dependent on the caller. They must work around mapping the caller’s contracts into the dependency’s concepts.
We can keep both callers and dependencies independent from each other with adapters. Adapters exist to map between caller contracts and dependency contracts. The adapter depends on both the caller and the code fulfilling the runtime/computational need so that neither has to depend on the other at design-time. Adapters free both the caller and the dependency to define contracts on their own terms and pretend other components don’t exist. Adapters cut design-time and semantic dependence while maintaining deferred runtime dependence.
Adapters are light-weight. The design freedom they gain usually outweighs the cost of an extra layer. The adapters also create a composable suite of caller-dependency relationships. The system can achieve substantial behavior changes without changing any services (i.e. the callers or the implementations).
Here’s a quick code example.
IMessageNotifier is the abstraction or “port” in this example. It could have adapters to all kinds of implementations.
For example, different email providers
Or even aggregates of other notifiers
Adapters are so powerful, they form the backbone of many of the most established architecture patterns. For example, Ports and Adapters or Clean Architecture.
There are many kinds of dependence and it’s valuable to consider which we keep and which we avoid. This post focused on design-time, runtime, and semantic coupling.
While not immediately intuitive, callers must own abstractions. Any other placement for abstractions fails to break semantic coupling, leads to unstable large scopes for abstractions, and most placements don’t even break design-time coupling.
Caller-owned abstractions with adapters break design-time and semantic coupling leading to far more stable and composable services.