Hand-written service containers

You say "convention over configuration;" I hear "ambient information stuck in someone's head." You say "configuration over hardcoding;" I hear "information in a different language that must be parsed, can be malformed, or not exist."

— Paul Snively (@paul_snively) March 2, 2019

Dependency injection is very important. Dependency injection containers are too. The trouble is with the tools, that let us define services in a meta-language, and rely on conventions to work well. This extra layer requires the “ambient information” Paul speaks about in his tweet, and easily lets us make mistakes that we wouldn’t make if we’d just write out the code for instantiating our services.

Please consider this article to be a thought experiment. If its conclusions are convincing to you, decide for yourself if you want to make it a coding experiment as well.

The alternative: a hand-written service container

I’ve been using hand-written service containers for workshop projects, and it turns out that it’s very nice to work with them. A hand-written service container would look like this:

final class ServiceContainer
{
    public function finalizeInvoiceController(): FinalizeInvoiceController
    {
        return new FinalizeInvoiceController(
            new InvoiceService(
                new InvoiceRepository(
                    $this->dbConnection()
                )
            )
        );
    }
    
    private function dbConnection(): Connection
    {
        static $connection;
        
        return $connection ?: $connection = new Connection(/* ... */);
    }
}

The router/dispatcher/controller listener, or any kind of middleware you have for processing an incoming web request, could retrieve a controller from the service container, and call its main method. Simplified, the code would look this:

$serviceContainer = new ServiceContainer();

if ($request->getUri() === '/finalize-invoice') {
    return $serviceContainer->finalizeInvoiceController()->__invoke($request);
}
// and so on

We see the power of dependency injection here: the service won’t have to fetch its dependencies, it will get them injected. The controller here is a so-called “entry point” for the service container, because it’s a public service that can be requested from it. All the dependencies of an entry-point service (and the dependencies of its dependencies, and so on), will be private services, which can’t be fetched directly from the container.

There are many things that I like about a hand-written dependency injection container. Every one of these advantages can show how many modern service containers have to reinvent features that you already have in the programming language itself.

No service ID naming conventions

For starters, service containers usually allow you to request services using a method like get(string $id). The hand-written container doesn’t have such a generic service getter. This means, you don’t have to think about what the ID should be of every service you want to define. You don’t have to come up with arbitrary naming conventions, and you don’t have to deal with inconsistent naming schemes in a legacy single project.

The name of a service is just the name of its factory method. Choosing a service name is therefore the same as choosing a method name. But since every method in your service container is going to create and return an object of a given type, why not use that type’s name as the name of the method? In fact, this is what most service containers have also started doing at some point: they recommend using the name of the class you want to instantiate.

Type-safe, with full support for static analysis

Several years ago I was looking for a way to check the quality of the Symfony service definitions that I wrote in Yaml. So I created a tool for validating service definitions created with the Symfony Dependency Injection Component. It would inspect the service definitions and find out if they had the right number constructor arguments, if the class name it referenced actually existed, etc. This tool helped me catch several issues that I would only have been able to find out by clicking through the entire web application.

Instead of doing complicated and incomplete analysis after writing service definitions in Yaml (or any other meta-language for that matter), if I write them in PHP, I get all the support from static analysis tools. Even if I don’t use a separate tool like PHPStan or Psalm, PhpStorm will point out any issues early on. Missing classes, missing import statements, too few or too many arguments, everything will be pointed out to me when I’m editing the ServiceContainer class in my IDE. This is a huge advantage.

Easy to refactor

Because analysis is easy, we can also expect all the help there is when refactoring our code. If we change production code, opening the ServiceContainer in your IDE will show you any issues you’ve produced. Furthermore, because there’s no special service definition format, your IDE doesn’t need a special extension to deal with it. It’s just plain PHP code. So any refactoring tool that you use (e.g. rename method, move to different namespace, etc.) will also deal with any existing usages inside the ServiceContainer class.

Easy to make a distinction between public entry points and private dependencies

I like how the Symfony service container allows users to make a distinction between private and public services. Public ones can be fetched using a call to get($id), private ones can only be used as dependencies for other (public or private) services. Some services indeed deserve to be public (mostly the services we earlier called “entry points”), most should remain private. Of course, the distinction between public and private services reminds us of the way we can have public and private methods too, and in fact, if you write your own service container, you will use these method scopes to accomplish the same thing.

If you hand-write the service container you can do some optimizations too, just like the Symfony container does them. For instance, if you have a private service that’s only used in one place, you can inline its instantiation. As an example, consider the private invoiceService() method:

public function finalizeInvoiceController(): FinalizeInvoiceController
{
    return new FinalizeInvoiceController(
        $this->invoiceService()
    );
}

private function invoiceService(): InvoiceService()
{
    return new InvoiceService(
        new InvoiceRepository(
            $this->dbConnection()
        )
    );
}

This method is only used by finalizeInvoiceController(), so we can safely inline it:

public function finalizeInvoiceController(): FinalizeInvoiceController
{
    return new FinalizeInvoiceController(
        new InvoiceService(
            new InvoiceRepository(
                $this->dbConnection()
            )
        )
    );
}

If, due to refactoring efforts, a private service is no longer needed, PhpStorm will tell you about it.

No need to define partial service definitions to assist auto-wiring

Auto-wiring has become quite popular, but I’m not convinced it’s the way to go. I’m sure most scenarios have been covered by now, so I’m not afraid that things won’t work out between me and auto-wiring. However, we’ll always have to do tricks to make it work. We have to give the wirer hints about which implementations to use. This means that we may be able to delete many service definitions, but we also have to keep some around, since there are some things that won’t work without them. You need to have in-depth knowledge about the dependency injection tool you use, and you need to learn the syntax for helping the auto-wirer. Worse, you may decide to adopt your production code so that the wirer can understand it.

Needless to say: if you write your service definitions in your own PHP class you’ll never need custom syntax. In fact, you don’t need to look up specific documentation at all, because you don’t have to worry about failures to resolve dependencies; you make all the decisions yourself when you write your ServiceContainer class.

No need to inject primitive-type values by their parameter name

A downside of auto-wiring containers is that they need special instructions when the injected constructor arguments aren’t objects, but primitive-type configuration values.

namespace App\Db;

final class Connection
{
    private $dsn;

    public function __construct(string $dsn)
    {
        // ...
        
        $this->dsn = $dsn;    
    }
}

Object-type constructor arguments can usually be resolved, but the service definition needs an extra hint for the $dsn argument:

    App\Db\Connection:
        arguments:
            $dsn: 'mysql:host=localhost;dbname=testdb'

This exposes an implementation aspect of the service class itself, which would normally remain hidden behind its public interface. To a client that instantiates an instance of Connection, only the parameter types should be relevant, not their names. In fact, a developer should be able to rename the parameter $dsn to something else, without breaking the application. Of course, a smart container builder will warn you about it when you rename a parameter, but this comes with some extra indirection that wouldn’t be needed at all if we’d just write the instantiation logic inside a manual ServiceContainer class, where parameter names are irrelevant (as they should be).

No magic effects on the service container

Talking about auto-wiring, I have to say I dislike the fact that the location of a file containing a class has an influence on it being defined as a service in the container. I’d want to be able to create a new class anywhere in the project, and decide for myself whether or not it gets defined as a service. Needless to say, you won’t have any magical effects like this if you write a custom service container, but personally I won’t miss them.

Easier to distinguish between services and singletons

Most service containers will automatically share instances of a service; if a service has been instantiated once, the next time you ask for it, you’ll get the exact same object instance. This is important for things like a database connection; you want to reuse the connection, instead of connecting to the database every time you need something from it. However, most services should be stateless anyway, and in that case it isn’t really necessary to retrieve the exact same instance.

A service that’s instantiated once and then shared between different clients is traditionally called a “singleton” service. Singleton services were usually implemented using the Singleton design pattern, which actually protects them from being instantiated more than once. Nowadays a service container manages service instances, and although it doesn’t use the Singleton design pattern, it still makes every service effectively a singleton service: there’s at most one instance of every service.

What I like about using a hand-written service container is that you can make a clear distinction between services for which it’s actually important to have only one instance, and services for which it doesn’t matter. Using the same example as earlier, note that the controller service can be re-instantiated every time a client needs it, and the connection service will be stored in a static variable, so that it can be reused:

final class ServiceContainer
{
    public function finalizeInvoiceController(): FinalizeInvoiceController
    {
        return new FinalizeInvoiceController(/* ... */);
    }
    
    private function dbConnection(): Connection
    {
        static $connection;
        
        return $connection ?: $connection = new Connection(/* ... */);
    }
} 

What about performance? Well, if it starts to hurt, you can always add some more shared services. But in most cases, I don’t think it’ll be needed. In part because of the fire-and-forget nature of the PHP server, but also because most services will be instantiated and used only once or twice anyway.

Easier to override parts for testing

With a service container based on a meta-language, like Yaml service definitions, you have to build in a mechanism for modifying the service definitions for use in different environments. With Yaml you can load in multiple service definition files and override service definitions and parameters. But just like “public” and “private” services, the concept of overriding services is also built into the programming language itself, namely by overriding methods. For example, if you want to use a fake/catch-all mailer or something while in development, you can do something like this:

abstract class ServiceContainer
{
    abstract protected function mailer(): Mailer;
}

final class DevelopmentServiceContainer extends ServiceContainer
{
    protected function mailer(): Mailer
    {
        return new FakeMailer();
    }
}

final class ProductionServiceContainer extends ServiceContainer
{
    protected function mailer(): Mailer
    {
        return new SmtpMailer(/* ... */);
    }
}

Optionally testable

If you want, you can even write an integration test for your hand-written service container. Then you could prove that the DevelopmentServiceContainer actually uses a FakeMailer. Or you can verify that all the public entry-points can be instantiated without producing any runtime errors (although static analysis will catch most of the issues already).

To be honest, this is also possible when you use a meta-language for defining services; you can always run tests against the compiled service container. However, I don’t see this happening often, so I just wanted to mention the possibility here.

Composing containers? Separating containers?

A big issue with maintaining a hand-written container is that you wouldn’t want to rewrite all of the framework’s service definitions yourself. In fact, this isn’t what I’m suggesting here. If you use a framework which ships with a service container, you just don’t have to use it for your own service definitions. For instance, you can define your service container as a service in your framework’s service container, and have access to it in the usual ways the framework supports.

You can even have multiple containers, for each of the modules/contexts you distinguish in your application. This could help keeping them actually separated (however, there’s no technical way to enforce a context to protect the integrity of a context, it’ll always be a people issue too).

Note that composition of containers is something that service containers don’t usually offer, but the programming language itself is capable of. You only have to inject containers as constructor arguments of other containers.

Conclusion

In an educational setting I found that one of the biggest advantages of having your own hand-written service container for your Application and Domain layer services is that it allows you to write acceptance tests using these services. You can freely instantiate the container in your Behat FeatureContext. You can then write tests which talk to the Application layer (instead of the Infrastructure layer as they usually do). These tests will run in a very short time, but most importantly: they will be less brittle, because they don’t (and can’t) rely on all kinds of infrastructure-level peculiarities.

In a project setting, I haven’t been fortunate enough to be able to use a hand-written service container. I’ll just wait for the next opportunity to do so. If in the meantime you find yourself agreeing with the thought experiment which is this article, and have even applied the idea in practice, let me know how it worked out for you!

Finally, some suggestions for further reading:

• Mathias Verraes has a nice explanation-in-a-gist where he argues that “We don’t need no DIC libs”, written after a discussion on Twitter about this topic. The gist also includes some implementation examples.
• Marijn Huizendveld has an interesting blog post about how you can deal in a better way with environment variables (which works well with hand-written service containers too).

• Dependency Injection
• Service Container