"Bad actors love the dependency addiction of modern developers"
Brings a new meaning to dependency injection.
Brings a new meaning to dependency injection.
My personal beef is that most of the time it acts like hidden global dependencies, and the configuration of those dependencies, along with their lifetimes, becomes harder to understand by not being traceable in the source code.
To me it‘s rather anti-functional. Normally, when you instantiate a class, the resulting object’s behavior only depends on the constructor arguments you pass it (= the behavior is purely a function of the arguments). With dependency injection, the object’s behavior may depend on some hidden configuration, and not even inspecting the class’ source code will be able to tell you the source of that bevavior, because there’s only an @Inject annotation without any further information.
Conversely, when you modify the configuration of which implementation gets injected for which interface type, you potentially modify the behavior of many places in the code (including, potentially, the behavior of dependencies your project may have), without having passed that code any arguments to that effect. A function executing that code suddenly behaves differently, without any indication of that difference at the call site, or traceable from the call site. That’s the opposite of the functional paradigm.
It sounds like you have a gripe with a particular DI framework and not the idea of Dependency Injection. Because
> Normally, when you instantiate a class, the resulting object’s behavior only depends on the constructor arguments you pass it (= the behavior is purely a function of the arguments)
With Dependency Injection this is generally still true, even more so than normal because you're making the constructor's dependencies explicit in the arguments. If you have a class CriticalErrorLogger(), you can't directly tell where it logs to, is it using a flat file or stdout or a network logger? If you instead have a class CriticalErrorLogger(logger *io.writer), then when you create it you know exactly what it's using to log because you had to instantiate it and pass it in.
Or like Kortilla said, instead of passing in a class or struct you can pass in a function, so using the same example, something like CriticalErrorLogger(fn write)
My issue with that is this: From the point of view of the code accessing the injected value (and from the point of view of that code's callers), the value appears like out of thin air. There is no way to trace back from that code where the value came from. Similarly, when defining which value will be injected, it can be difficult to trace all the places where it will be injected.
In addition, there are often lifetime issues involved, when the injected value is itself a stateful object, or may indirectly depend on mutable, cached, or lazy-initialized, possibly external state. The time when the value's internal state is initialized or modified, or whether or not it is shared between separate injection points, is something that can't be deduced from the source code containing the injection points, but is often relevant for behavior, error handling, and general reasoning about the code.
All of this makes it more difficult to reason about the injected values, and about the code whose behavior will depend on those values, from looking at the source code.
I agree with your definition except for this part, you don't need any framework to do dependency injection. It's simply the idea that instead of having an abstract base class CriticalErrorLogger, with the concrete implementations of StdOutCriticalErrorLogger, FileCriticalErrorLogger, AwsCloudwatchCriticalErrorLogger which bake their dependency into the class design; you instead have a concrete class CriticalErrorLogger(dep *dependency) and create dependency objects externally that implement identical interfaces in different ways. You do text formatting, generating a traceback, etc, and then call dep.write(myFormattedLogString), and the dependency handles whatever that means.
I agree with you that most DI frameworks are too clever and hide too much, and some forms of DI like setter injection and reflection based injection are instant spaghetti code generators. But things like Constructor Injection or Method Injection are so simple they often feel obvious and not like Dependency Injection even though they are. I love DI, but I hate DI frameworks; I've never seen a benefit except for retrofitting legacy code with DI.
And yeah it does add the issue or lifetime management. That's an easy place to F things up in your code using DI and requires careful thought in some circumstances. I can't argue against that.
But DI doesn't need frameworks or magic methods or attributes to work. And there's a lot of situations where DI reduces code duplication, makes refactoring and testing easier, and actually makes code feel less magical than using internal dependencies.
The basic principle is much simpler than most DI frameworks make it seem. Instead of initializing a dependency internally, receive the dependency in some way. It can be through overly abstracted layers or magic methods, but it can also be as simple as adding an argument to the constructor or a given method that takes a reference to the dependency and uses that.
edit: made some examples less ambiguous
The term Dependency Injection was coined by Martin Fowler with this article: https://martinfowler.com/articles/injection.html. See how it presents the examples in terms of wiring up components from a configuration, and how it concludes with stressing the importance of "the principle of separating service configuration from the use of services within an application". The article also presents constructor injection as only one of several forms of dependency injection.
That is how everyone understood dependency injection when it became popular 10-20 years ago: A way to customize behavior at the top application/deployment level by configuration, without having to pass arguments around throughout half the code base to the final object that uses them.
Apparently there has been a divergence of how the term is being understood.
[0] https://en.wikipedia.org/wiki/Strategy_pattern
[1] The fact that Car is abstract in the example is immaterial to the pattern, and a bit unfortunate in the Wikipedia article, from a didactic point of view.
DI in Java is almost completely disconnected from what the Strategy pattern is, so it doesn't make sense to use one to refer to the other there.
It's equivalent to partial application.
An uninstantiated class that follows the dependency injection pattern is equivalent to a family of functions with N+Mk arguments, where Mk is the number of parameters in method k.
Upon instantiation by passing constructor arguments, you've created a family of functions each with a distinct sets of Mk parameters, and N arguments in common.
That's the best way to think of it fundamentally. But the main implication of that which is at some point something has to know how to resolve those dependencies - i.e. they can't just be constructed and then injected from magic land. So global cradles/resolvers/containers/injectors/providers (depending on your language and framework) are also typically part and parcel of DI, and that can have some big implications on the structure of your code that some people don't like. Also you can inject functions and methods not just constructors.
This is all well and good, but you also need a bunch of code that handles resolving those dependencies, which oftentimes ends up being complex and hard to debug and will also cause runtime errors instead of compile time errors, which I find to be more or less unacceptable.
Edit: to elaborate on this, I’ve seen DI frameworks not be used in “enterprise” projects a grand total of zero times. I’ve done DI directly in personal projects and it was fine, but in most cases you don’t get to make that choice.
Just last week, when working on a Java project that’s been around for a decade or so, there were issues after migrating it from Spring to Spring Boot - when compiled through the IDE and with the configuration to allow lazy dependency resolution it would work (too many circular dependencies to change the code instead), but when built within a container by Maven that same exact code and configuration would no longer work and injection would fail.
I’m hoping it’s not one of those weird JDK platform bugs but rather an issue with how the codebase is compiled during the container image build, but the issue is mind boggling. More fun, if you take the .jar that’s built in the IDE and put it in the container, then everything works, otherwise it doesn’t. No compilation warnings, most of the startup is fine, but if you build it in the container, you get a DI runtime error about no lazy resolution being enabled even if you hardcode the setting to be on in Java code: https://docs.spring.io/spring-boot/api/kotlin/spring-boot-pr...
I’ve also seen similar issues before containers, where locally it would run on Jetty and use Tomcat on server environments, leading to everything compiling and working locally but throwing injection errors on the server.
What’s more, it’s not like you can (easily) put a breakpoint on whatever is trying to inject the dependencies - after years of Java and Spring I grow more and more convinced that anything that doesn’t generate code that you can inspect directly (e.g. how you can look at a generated MapStruct mapper implementation) is somewhat user hostile and will complicate things. At least modern Spring Boot is good in that more of the configuration is just code, because otherwise good luck debugging why some XML configuration is acting weird.
In other words, DI can make things more messy due to a bunch of technical factors around how it’s implemented (also good luck reading those stack traces), albeit even in the case of Java something like Dagger feels more sane https://dagger.dev/ despite never really catching on.
Of course, one could say that circular dependencies or configuration issues are project specific, but given enough time and projects you will almost inevitably get those sorts of headaches. So while the theory of DI is nice, you can’t just have the theory without practice.
Hidden dependencies are: untyped context variable; global "service registry", etc. Those are hidden, the only way to find out which dependencies given module has is to carefully read its code and code of all called functions.
I'm talking more specifically about Aspect Oriented Programming though and DI containers in OOP, which seemed pretty clever in theory, but have a lot of issues in reality.
I take no issues with currying in functional programming.
There are other good uses of it but it absolutely can get out of control, especially if implemented by someone whose just discovered it and wants to use it for everything.