Dependency injections frameworks are pure ideology24 Oct 2020
Mingus believes now that it got too far away from jazz – spontaneity – since almost all of the music was written. He remembers one rehearsal at which Teddy had left several bars open for blowing and everyone jumped on him with ‘Man, are you lazy? Write it out!’
Dependency injection is a great idea. Supplying dependencies to a class, rather than having the class create them themselves, is a practical illustration of the benefits of the single-responsibility principle. Creating dependencies is a different responsibility from using them, and the class that uses dependencies probably doesn’t (or shouldn’t) have the information required to create the dependencies; so you can isolate this responsibility by in a separate class (or multiple classes, for different environments or for testing) by injecting the dependencies. But dependency injection doesn’t require a framework, and indeed dependency injection frameworks are useless if they’re not actively harmful; their continued prevelance is the zombie-like persistence of a dead ideology.
Dependency injection frameworks developed as a lightweight alternative to Enterprise Java Beans, which were the microservices of the late 90s - a technology for creating applications by wiring up independent components. The EJB specification included lots of parts (distributed objects, persistence), and mandated, or at least strongly encouraged, applications to be built using all these specific EJB technologies. Dependency injection frameworks were more lightweight in the sense that they focussed just on this wiring up of independent components. Indeed, because they were more lightweight they made this kind of componentisation easier, and so more widespread.
The classic dependency injection framework was implemented by early versions of Spring, where each component is defined as an independent Java class, with the injections of one component into another coordinated by XML files. The idea is that the connections between components are configuration, and so can be externalised; new systems can be constructed without any code changes, just by re-assembling the components. This turned out to be a bad idea. Connecting components to one another is programming, and should be treated as such: it needs to be tested, and benefits from compile-time type-checking and IDE support. So coordinating dependency injection via externalised XML is now discouraged; modern dependency injection frameworks express dependency requirements with in-code metadata annotations.
The problem with these modern dependency injection frameworks is that they’re pointless. We already have a way to declare and provide dependencies in code. The Spring developers already implicitly recognise this, as the current recommended practice is to declare required dependencies simply by having them as constructor parameters (rather than by annotating fields); I hope at some point they’ll notice that it’s also possible to provide dependencies by passing them to constructors. Wiring dependencies by writing constructor calls is better than delegating the job to a dependency injection framework because it’s more explicit. What gets injected where is visible to developers and, perhaps even more importantly, to tooling, so the presence of required dependencies can be checked at compile time, and the specific implementation of particular dependencies can be navigated through standard IDE functionality.
So, when I see people using dependency injection frameworks, I’m reminded of the Charles Mingus anecdote quoted above – I think, ‘Are you lazy? Just write the constructor calls.’ But, actually, the continued existence of dependency injection frameworks isn’t due to individual developer laziness; it’s the lingering remnant of the componentized dream which reached its full realization in the original Spring’s externalised configuration. This is one form of a longer-term industry dream, that of programs without programmers, which inspired fourth-generation languages in the 70s, and which we see in today’s low-code platform. And perhaps one day this dream will come true, but it didn’t come true through dependency injection frameworks; and we don’t need to keep using these frameworks through loyalty to this half-remembered dream.
(Some people will say, if you don’t use a dependency injection framwork, how will you manage the generation of proxies that implement aspect-oriented programming? The answer here is that AOP is also bad, and you shouldn’t use it.)