Title page

Foreword

By Luis Cordova

Open source projects have a history and a lot of reasoning behind them that evolves. Symfony2 has been evolving for the past several years. Many new developers coming to use the framework feel the steep curve of mastering the framework and overcome it at the end, but are left with many doubts when it comes to developing their projects in the true Symfony way.

For symfony1 there was a book that covered the symfony1 practices. For Symfony2 there is the Book which is the major source for documentation. The Book has also Cookbook entries which help somewhat. However, even then not all the aspects and practices brewed during so many years can be summarized with those resources for developers with a hunger for understanding the why’s and best practices that only real-life field experience with Symfony can provide. There are blogs about Symfony development also, however they are sparse and blogs even like mine assume a lot of things about the readership and are more for experimental things and nuances of the framework. What has been missing up to now has been a strong technical opinion poured in a book catering to the community and pointing in the right “Symfony way”.

Matthias has had more than one year understanding Symfony from the inside out and bringing responses to the why’s to the surface for us. I could not trust anyone other than a Certified Symfony Developer to do this. Seldom times have I felt I understood Symfony so well to use it as my very own arsenal. Reading this book has been one of those times. Another time was when I heard Kris Wallsmith explaining/acting out the Symfony Security Component. This book is a must for everyone who is eager to carefully understand the framework.

Matthias will give away in this book a lot of secrets so you will find yourself often coming back to the book to see how a recommendation was issued, and what to do in your case. Matthias has also spoken about things I even did not think I would find in this book and he did it in a straightforward manner with examples and code.

I truly hope you enjoy this book.

Your friend from the symfony community,

Luis Cordova (@cordoval)

Introduction

A year with Symfony. In fact, for me it’s been more than a year, almost 6 years even. First with symfony 1 (mind the lower case and the separate 1), then with Symfony2. Symfony2 is what you may call a “mature” framework. You can do pretty advanced things with it. And when you want to do one of these very advanced things, you can also choose not to install the entire framework, but just one of its components.

Starting to work with Symfony2 meant for me: learning much about programming in general and applying many of the things I only knew from books to any code I produced from then on. Symfony2 did this for me: encourage me to do things right.

In the meantime, I’ve written a lot about Symfony2, contributed to its documentation (specifically some Cookbook articles and the documentation for the Security Component and the Config Component), I started my own blog with articles about PHP in general, Symfony2, its components and related frameworks and libraries like Silex and Twig. And I became a Symfony2 Certified Developer, during the very first exam session in Paris at the Symfony Live conference in 2012.

All this is now being followed by the book you’re reading now, called A Year With Symfony. It contains many of the best practices that me and my fellow developers at IPPZ discovered while working on large Symfony2 applications. It also contains some in-depth knowledge that you’ll need to have when you’re going one step further than just writing code inside a controller or a template.

Thank you

Before I continue, let me take a moment to thank some people who have helped me finishing this book. First of all, Luis Cordova, who has been following my steps since I first started to write about Symfony2 in 2011. He did a thorough review of the first draft. My colleagues at IPPZ have also provided me with valuable feedback, encouraging me to make some things clearer, and some things more interesting: Dennis Coorn, Matthijs van Rietschoten and Maurits Henneke. Working with them for two years, sharing concerns about maintainability, readability, reusability, and all the other -bilities, like laughability has been great fun. Thanks also to Lambert Beekhuis, organizer of the Dutch Symfony2 Usergroup meetups, for giving me some valuable feedback with regard to my English grammar.

Who should read this book

I’ve written this book with developers in mind who have worked with PHP for some time and with Symfony2 for a couple of weeks, maybe months. I assume you have used the official Symfony2 documentation to get acquainted with the basics of creating Symfony2 applications. I expect you to know the basic structure of an application (the standard directory structure, the way you can create or register a bundle), how you can create controllers and add routing configuration for them, how to create form types, and write Twig templates.

I also assume you have worked with some kind of persistence library, be it Doctrine ORM, Doctrine MongoDB ODM, Propel, etc. However, in the book I only refer to Doctrine libraries, to keep things simple. When you use some other library for persisting objects, you are likely able to find out how to apply some of the concepts described in this book to code written for another persistence library.

Conventions

Since this book is explicitly about Symfony2, from now on I talk about “Symfony” which looks a bit nicer.

Everything I say about Symfony is related to version 2. I’ve written and tested the code examples in this book with Symfony 2.3. However, they may very well also be applicable to Symfony 2.1.* and 2.2.* and maybe even to Symfony 2.0.*.

In this book I show pieces of code from Symfony itself. To make this code fit on a page and still be readable, I have sometimes modified it.

Overview of the contents

The first part of this book is called The journey from request to response. It will take you along from the first entry point of a Symfony application in the front controller to the last breath it takes before sending a response back to the client. At times I will show you how you can hook into the process and modify the flow or just change the results of the intermediate steps.

The next part is called Patterns of dependency injection. It contains a collection of patterns that are solutions to recurring problems when it comes to creating or modifying service definitions, based on a bundle’s configuration. I will show you many very practical examples that you can use to model your own bundle’s container extension, configuration class and compiler passes.

The third part is about Project structure. I suggest various ways to get your controllers much cleaner, by delegating actions to form handlers, domain managers and event listeners. We also take a look at state and how to avoid it in the service layer of your application.

A quick intermezzo follows then about Configuration conventions. This part should help you with setting up the configuration for your application. It encourages you and your team to settle on some kind of a configuration convention.

The fifth part is very important as it concerns every serious application, with user sessions and sensitive data. It is about Security. This would seem to be covered completely by all the Symfony components (after all the framework itself has been audited for security issues) and Twig, but unfortunately such a thing would not be possible. You always have to keep thinking about security yourself. This part of the book contains various suggestions on how to deal with security, where to keep an eye on, when you can rely on the framework and when you need to take care of things yourself.

The sixth part is about annotations. When Symfony2 was first released in 2011 it introduced annotations as a revolutionary way to configure an application from within the doc blocks of classes, methods and properties. The first chapter of this part explains how annotations work. After that, you will learn how to create your own annotations and how you can use annotations to influence the response that is generated for a request.

The final part covers all ways of Being a Symfony developer although in fact this part is one big encouragement to not be a Symfony developer and to write things as loosely coupled to the Symfony framework as possible. This means separating code into reusable and project-specific code, then splitting the reusable code into library and bundle code. I will discuss different other ideas that make your bundles nice, clean and friendly for other projects.

Enjoy!

Changelog

I am currently revising this book. I want it to be up-to-date with the latest releases (up to Symfony 3.0) as well as my (slightly modified) current point of view. This is a list of the most important changes that I made:

General

  • Apply gender-neutral grammar.

The journey from request to response

Patterns of dependency injection

Project structure

  • Work in progress

Configuration conventions

Security

  • Work in progress

Annotations

  • Work in progress

Being a Symfony developer

  • Work in progress

I The journey from request to response

1. The HttpKernelInterface

Symfony is famous for its HttpKernelInterface:

namespace Symfony\Component\HttpKernel;

use Symfony\Component\HttpFoundation\Request;
use Symfony\Component\HttpFoundation\Response;

interface HttpKernelInterface
{
    const MASTER_REQUEST = 1;
    const SUB_REQUEST = 2;

    /**
      * @return Response
      */
    public function handle(
        Request $request,
        $type = self::MASTER_REQUEST,
        $catch = true
    );
}

An implementation of this interface would only have to implement one method and thereby declare itself capable of converting in some way a given Request into a Response. When you take a look at any of the front controllers in the /web directory of a Symfony project, you can see that this handle() method plays a central role in processing web requests - as you might expect:

// in /web/app.php
$kernel = new AppKernel('prod', false);
$request = Request::createFromGlobals();
$response = $kernel->handle($request);
$response->send();

First, AppKernel gets instantiated. This is a class specific to your project, and you can find it in /app/AppKernel.php. It allows you to register your bundles, and to change some major settings, like the location of the cache directory or the configuration file that should be loaded. Its constructor arguments are the name of the environment and whether or not the kernel should run in debug mode.

Next a Request object is created based on the existing PHP superglobals ($_GET, $_POST, $_COOKIE, $_FILES and $_SERVER). The Request class together with other classes from the HttpFoundation component provide object-oriented ways to wrap the superglobals. These classes also cover many corner cases you may experience with different versions of PHP or on different platforms. It is wise (in a Symfony context) to always use Request to retrieve any data you would normally have taken directly from the superglobals.

Then the handle() method of the AppKernel instance gets called. Its only argument is the current Request object. The default arguments for the type of the request (“master”) and whether or not to catch and handle exceptions (yes) will be added automatically.

The result of this handle() method is guaranteed to be an instance of Response (also from the HttpFoundation component). Finally the response will be sent back to the client that made the request - for instance a browser.

1.1 Booting the kernel

Of course, the magic happens inside the handle() method of the kernel. You will find this method implemented in the Kernel class, which is the parent class of AppKernel:

// in Symfony\Component\HttpKernel\Kernel

public function handle(
    Request $request,
    $type = HttpKernelInterface::MASTER_REQUEST,
    $catch = true
) {
    if (false === $this->booted) {
        $this->boot();
    }

    return $this->getHttpKernel()->handle($request, $type, $catch);
}

First of all, it is made sure that the Kernel is booted, before the HttpKernel is asked to do the rest. The process of booting includes:

  • Initializing all the registered bundles
  • Initializing the service container

Bundles as container extensions

Bundles are known amongst Symfony developers as the place to put your own code. Each bundle should have a name that reflects what kind of things you could do with the code inside it. For instance you may have a BlogBundle, a CommunityBundle, a CommentBundle, etc. You register your bundles in AppKernel.php, by adding them to the existing list of bundles:

class AppKernel extends Kernel
{
    public function registerBundles()
    {
        $bundles = array(
            new Symfony\Bundle\FrameworkBundle\FrameworkBundle(),
            ...,
            new Matthias\BlogBundle()
        );

        return $bundles;
    }
}

This is definitely a good idea - it allows you to plug functionality into and out of your project with a single line of code. However, when looking at the Kernel and how it deals with all bundles, including yours, it becomes apparent that bundles are mainly treated as ways to extend the service container, not as libraries of code. This is why you find a DependencyInjection folder inside many bundles, accompanied by a {nameOfTheBundle}Extension class. During the process of initializing the service container, each bundle is allowed to register some services of its own to the service container, maybe add some parameters too, and possibly modify some service definitions before the container gets compiled and dumped to the cache directory:

namespace Matthias\BlogBundle\DependencyInjection;

use Symfony\Component\HttpKernel\DependencyInjection\Extension;
use Symfony\Component\Config\FileLocator;
use Symfony\Component\DependencyInjection\Loader\XmlFileLoader;

class MatthiasBlogExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $loader = new XmlFileLoader($container,
            new FileLocator(__DIR__.'/../Resources/config'));

        // add service definitions to the container
        $loader->load('services.xml');

        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        // set a parameter
        $container->setParameter(
            'matthias_blog.comments_enabled',
            $processedConfig['enable_comments']
        );
    }

    public function getAlias()
    {
        return 'matthias_blog';
    }
}

The name returned by the getAlias() method of a container extension is actually the key under which you can set configuration values (for instance in config.yml):

matthias_blog:
    enable_comments: true

You will read more about bundle configuration in Patterns of dependency injection.

Creating the service container

After all the bundles have been enabled to add their services and parameters, the container is finalized in a process that is called “compilation”. During this process it is still possible to make some last-minute changes to service definitions or parameters. It is also the right moment to validate and optimize service definitions. Afterwards, the container is in its final form, and it gets dumped into two different formats: an XML file of all resolved definitions and parameters and a PHP file ready to be used as the one and only service container in your application.

Both files can be found in the cache directory corresponding to the environment of the kernel, for instance /app/cache/dev/appDevDebugProjectContainer.xml. The XML file looks like any regular XML service definition file, only a lot bigger:

<service id="event_dispatcher" class="...\ContainerAwareEventDispatcher">
  <argument type="service" id="service_container"/>
  <call method="addListenerService">
    <argument>kernel.controller</argument>
    ...
  </call>
  ...
</service>

The PHP file contains a method for each service that can be requested. Any creation logic, like constructor arguments or method calls after instantiation can be found in this file, and it is therefore the perfect place to debug your service definitions in case anything appears to be wrong with them:

class appDevDebugProjectContainer extends Container
{
    ...

    protected function getEventDispatcherService()
    {
        $this->services['event_dispatcher'] =
            $instance = new ContainerAwareEventDispatcher($this);

        $instance->addListenerService('kernel.controller', ...);

        ...

        return $instance;
    }

    ...
}

1.2 From the Kernel to the HttpKernel

Now that the kernel is booted (i.e. all bundles are initialized, their extensions are registered, and the service container is finalized), the real handling of the request is delegated to an instance of HttpKernel:

// in Symfony\Component\HttpKernel\Kernel

public function handle(
    Request $request,
    $type = HttpKernelInterface::MASTER_REQUEST,
    $catch = true
) {
    if (false === $this->booted) {
        $this->boot();
    }

    return $this->getHttpKernel()->handle($request, $type, $catch);
}

The HttpKernel implements HttpKernelInterface and it truly knows how to convert a request to a response. The handle() method looks like this:

public function handle(
    Request $request,
    $type = HttpKernelInterface::MASTER_REQUEST,
    $catch = true
) {
    try {
        return $this->handleRaw($request, $type);
    } catch (\Exception $e) {
        if (false === $catch) {
            throw $e;
        }

        return $this->handleException($e, $request, $type);
    }
}

As you can see, most of the work is done in the private handleRaw() method, and the try/catch block is here to capture any exceptions. When the initial argument $catch was true (which is the default value for “master” requests), every exception will be handled nicely. The HttpKernel will try to find someone who can still create a decent Response object for it (see also Exception handling).

2. Events leading to a response

The handleRaw() method of the HttpKernel is a beautiful piece of code, in which it becomes clear that handling a request is not deterministic per se. There are all kinds of ways in which you can hook into the process, and completely replace or just modify any intermediate result.

2.1 Early response

The first moment you can take control of handling the request, is right at the beginning. Usually the HttpKernel will try to generate a response by executing a controller. But any event listener that listens to the KernelEvents::REQUEST (kernel.request) event is allowed to generate a completely custom response:

use Symfony\Component\HttpKernel\Event\GetResponseEvent;

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    $event = new GetResponseEvent($this, $request, $type);
    $this->dispatcher->dispatch(KernelEvents::REQUEST, $event);

    if ($event->hasResponse()) {
        return $this->filterResponse(
            $event->getResponse(),
            $request,
            $type
        );
    }

    ...
}

As you can see, the event object that gets created is an instance of GetResponseEvent and it allows listeners to set a custom Response object using its setResponse() method, for example:

use Symfony\Component\HttpFoundation\Response;
use Symfony\Component\HttpKernel\Event\GetResponseEvent;

class MaintenanceModeListener
{
    public function onKernelRequest(GetResponseEvent $event)
    {
        $response = new Response(
            'This site is temporarily unavailable',
            503
        );

        $event->setResponse($response);
    }
}

Some notable kernel.request event listeners

The framework itself has many listeners for the kernel.request event. These are mostly listeners to set some things straight, before letting the kernel call any controller. For instance one listener makes sure the application has a locale (either the default locale, or the _locale part from the URI), another one processes requests for fragments of pages.

There are however two main players in the early request stage: the RouterListener and the Firewall. The RouterListener takes the path info from the Request object and tries to match it to some known route. It stores the result of the matching process in the Request object as attributes, for instance the name of the controller that corresponds to the route that was matched:

namespace Symfony\Component\HttpKernel\EventListener;

class RouterListener implements EventSubscriberInterface
{
    public function onKernelRequest(GetResponseEvent $event)
    {
        $request = $event->getRequest();

        $parameters = $this->matcher->match($request->getPathInfo());

        ...

        $request->attributes->add($parameters);
    }
}

When for example the matcher is asked to match /demo/hello/World, and the routing configuration looks like this:

_demo_hello:
    path: /demo/hello/{name}
    defaults:
        _controller: AcmeDemoBundle:Demo:hello

The parameters returned by the match() call are a combination of the values defined under defaults: and the dynamic values matched by placeholders in the path (like {name}):

array(
  '_route' => '_demo_hello',
  '_controller' => 'AcmeDemoBundle:Demo:hello',
  'name' => 'World'
);

These end up in the Request parameter bag called “attributes”. As you would be able to guess: HttpKernel will later examine the request attributes and execute the given controller.

Another important event listener is the Firewall. As we saw above, the RouterListener does not provide the HttpKernel with a Response object, it just does some work at the beginning of a request. On the contrary, the Firewall sometimes forces a certain Response object, for instance when a user is not authenticated when they should have been, since they have requested a protected page. The Firewall (through quite a complex process) then forces a redirect to for instance a login page, or sets some headers requiring the user to enter their credentials using HTTP authentication.

2.2 Resolving the controller

We have seen above that the RouterListener sets a Request attribute called _controller, which contains some kind of reference to the controller that is to be executed. This information is not known to the HttpKernel. Instead, there is an object called ControllerResolver that is asked to return a controller for the current Request object:

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    ...

    if (false === $controller = $this->resolver->getController($request)) {
        throw new NotFoundHttpException();
    }
}

The resolver itself is an instance of ControllerResolverInterface:

namespace Symfony\Component\HttpKernel\Controller;

use Symfony\Component\HttpFoundation\Request;

interface ControllerResolverInterface
{
    public function getController(Request $request);

    public function getArguments(Request $request, $controller);
}

It will later be used to determine the right arguments for executing the controller. Its first task is to get the controller. The standard controller resolver takes the controller from the Request attribute _controller:

public function getController(Request $request)
{
    if (!$controller = $request->attributes->get('_controller')) {
        return false;
    }

    ...

    $callable = $this->createController($controller);

    ...

    return $callable;
}

Since in most cases the controller is a string in some kind of (short-hand) notation, the controller needs to be actually created before it can be returned.

2.3 Allow for replacement of the controller

Back in the HttpKernel the controller is now available and almost ready to be executed. But even though the controller resolver has done everything in its power to prepare a valid callable before executing it, there is a last chance to completely replace it with some other controller (which can be any callable again). This chance is provided by another event: KernelEvents::CONTROLLER (kernel.controller):

use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    // use the controller resolver to get the controller
    ...

    $event = new FilterControllerEvent($this, $controller, $request, $type);
    $this->dispatcher->dispatch(KernelEvents::CONTROLLER, $event);
    $controller = $event->getController();
}

By calling the setController() method of the FilterControllerEvent object, it is possible to override the controller that should be executed:

use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

class ControllerListener
{
    public function onKernelController(FilterControllerEvent $event)
    {
        $event->setController(...);
    }
}

Some notable kernel.controller listeners

The framework itself has no listeners for the kernel.controller event. There are only third-party bundles which listen to the event to anticipate the fact that the controller has been determined and that it will be executed.

The ControllerListener from the SensioFrameworkExtraBundle for instance does some very important work right before executing a controller: it collects annotations like @Template and @Cache and stores them as request attributes with the same name with an underscore as a prefix: _template and _cache. Later in the process of handling the request, these annotations (or configurations as they are called in the code of this bundle) will be used to render a template, or set some cache-related response headers.

The ParamConverterListener from the same bundle will convert extra controller arguments, for instance entities that should be resolved by taking the id value from the route:

/**
 * @Route("/post/{id}")
 */
public function showAction(Post $post)
{
    ...
}

2.4 Collect arguments for executing the controller

After any listener has been enabled to replace the controller, the controller we have now is the definitive one. The next step is to collect the arguments that should be used when executing it:

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    // use the controller resolver to get the controller
    // "kernel.controller" event
    ...

    $arguments = $this->resolver->getArguments($request, $controller);
}

The controller resolver is asked to supply the controller arguments. The standard ControllerResolver from the HttpKernel Component uses reflection and the attributes from the Request object to resolve the controller arguments. It loops over all the parameters from the controller method. The following logic is used to determine each argument:

Resolving a controller argument
Resolving a controller argument

2.5 Execute the controller

Finally, it’s time to execute the controller. The response is caught and further processed.

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    // use the controller resolver to get the controller
    // "kernel.controller" event
    // use the controller resolver to get the controller arguments
    ...

    $response = call_user_func_array($controller, $arguments);

    if (!$response instanceof Response) {
        ...
    }
}

As you may remember from the Symfony documentation, a controller should return a Response object. If it doesn’t, some other part of the application should be able to convert the return value to a Response object in some way or another.

2.6 Enter the view layer

When you choose to return a Response object directly from your controller, you can thereby effectively by-pass the templating engine:

class SomeController
{
    public function simpleAction()
    {
        return new Response(
            '<html><body><p>Pure old-fashioned HTML</p></body></html>'
        );
    }
}

However, when you return anything else (though usually an array of template variables), this return value needs to be converted to a Response object before it can be used as a decent response to be sent back to the client. The HttpKernel is explicitly not coupled to a specific templating engine like Twig. Instead it uses the event dispatcher again to allow any listener to the KernelEvents::VIEW event (kernel.view) to set a proper response based on the return value of the controller (though it may choose to ignore this value entirely):

use Symfony\Component\HttpKernel\Event\GetResponseForControllerResultEvent;

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    // use the controller resolver to get the controller
    // "kernel.controller" event
    // use the controller resolver to get the controller arguments
    // execute the controller
    ...

    $event = new GetResponseForControllerResultEvent(
        $this,
        $request,
        $type,
        $response
    );
    $this->dispatcher->dispatch(KernelEvents::VIEW, $event);

    if ($event->hasResponse()) {
        $response = $event->getResponse();
    }

    if (!$response instanceof Response) {
        // the kernel REALLY needs a response by now

        throw new \LogicException(...);
    }
}

Listeners to this event can call setResponse() on the GetResponseForControllerResultEvent object:

use Symfony\Component\HttpKernel\Event\GetResponseForControllerResultEvent;

class ViewListener
{
    public function onKernelView(GetResponseForControllerResultEvent $event)
    {
        $response = new Response(...);

        $event->setResponse($response);
    }
}

A notable kernel.view listener

The TemplateListener from the SensioFrameworkExtraBundle takes the controller result and uses it as template variables to render the template that was specified using the @Template annotation (which was stored as the request attribute _template):

public function onKernelView(GetResponseForControllerResultEvent $event)
{
    $parameters = $event->getControllerResult();

    // get the templating engine
    $templating = ...;

    $event->setResponse(
        $templating->renderResponse($template, $parameters)
    );
}

2.7 Filter the response

In the end, right before returning the Response object as the final result of handling the given Request object, any listener for the KernelEvents::RESPONSE event (kernel.response) will be notified:

private function handleRaw(Request $request, $type = self::MASTER_REQUEST)
{
    // "kernel.request" event
    // use the controller resolver to get the controller
    // "kernel.controller" event
    // use the controller resolver to get the controller arguments
    // convert the controller result to a Response object

    return $this->filterResponse($response, $request, $type);
}

private function filterResponse(Response $response, Request $request, $type)
{
    $event = new FilterResponseEvent($this, $request, $type, $response);

    $this->dispatcher->dispatch(KernelEvents::RESPONSE, $event);

    return $event->getResponse();
}

Event listeners are allowed to modify the Response object and even to replace it completely:

use Symfony\Component\HttpKernel\Event\FilterResponseEvent;

class ResponseListener
{
    public function onKernelResponse(FilterResponseEvent $event)
    {
        $response = $event->getResponse();

        $response->headers->set('X-Framework', 'Symfony2');

        // or

        $event->setResponse(new Response(...));
    }
}

Notable kernel.response listeners

The WebDebugToolbarListener from the WebProfilerBundle injects some HTML and JavaScript code at the end of the response to make sure the profiler toolbar appears (usually at the bottom of the page).

The ContextListener from the Symfony Security Component stores a serialized version of the current security token in the session. This allows for a much faster authentication process during the next request. The Security Component also has a ResponseListener which sets a cookie containing “remember-me” information. Its contents can be used to auto-login a user even when their original session was already destroyed.

3. Exception handling

It is not unlikely that during the long journey from request to response some kind of an error or exception occurs. By default, the kernel is instructed to catch any exception and even then it tries to find an appropriate response for it. As we already saw, the entire request handling gets wrapped in a try/catch block:

public function handle(
    Request $request,
    $type = HttpKernelInterface::MASTER_REQUEST,
    $catch = true
) {
    try {
        return $this->handleRaw($request, $type);
    } catch (\Exception $e) {
        if (false === $catch) {
            throw $e;
        }

        return $this->handleException($e, $request, $type);
    }
}

When $catch equals true, the handleException() method is called and is expected to create a response. This method dispatches a KernelEvents::EXCEPTION event (kernel.exception) with a GetResponseForExceptionEvent object.

use Symfony\Component\HttpKernel\Event\GetResponseForExceptionEvent;

private function handleException(\Exception $e, $request, $type)
{
    $event = new GetResponseForExceptionEvent($this, $request, $type, $e);
    $this->dispatcher->dispatch(KernelEvents::EXCEPTION, $event);

    // a listener might have replaced the exception
    $e = $event->getException();

    if (!$event->hasResponse()) {
        throw $e;
    }

    $response = $event->getResponse();

    ...
}

Listeners for the kernel.exception event are allowed to:

  • Set a proper Response object for this specific exception.
  • Replace the original Exception object.

When none of the listeners has called setResponse() on the event object, the exception will be thrown (again), but this time it will not be handled automatically. So in case your display_errors PHP setting equals true, PHP just renders it as-is.

In case any of the listeners has set a Response object, the HttpKernel examines this object in order to set the right status code for the response:

// the developer asked for a specific status code
if ($response->headers->has('X-Status-Code')) {
    $response->setStatusCode($response->headers->get('X-Status-Code'));

    $response->headers->remove('X-Status-Code');
} elseif (
    !$response->isClientError()
    && !$response->isServerError()
    && !$response->isRedirect()
) {
    // ensure that we actually have an error response
    if ($e instanceof HttpExceptionInterface) {
        // keep the HTTP status code and headers
        $response->setStatusCode($e->getStatusCode());
        $response->headers->add($e->getHeaders());
    } else {
        $response->setStatusCode(500);
    }
}

This is quite useful: we can enforce a certain status code to be used by adding an X-Status-Code header to the Response object (remember: this only works for exceptions that are caught by the HttpKernel), or by throwing exceptions that implement HttpExceptionInterface. Otherwise the status code defaults to 500 - Internal server error. This is much better than the standard PHP behavior, which will return a response with status 200 - OK, when an error has occurred.

When an event listener has set a Response object, this response is not handled any differently than a normal response, so the last step in handling an exception is to filter the response. When another exception gets thrown while filtering the response, this exception will simply be ignored, and the unfiltered response will be sent to the client.

try {
    return $this->filterResponse($response, $request, $type);
} catch (\Exception $e) {
    return $response;
}

3.1 Notable kernel.exception listeners

The ExceptionListener from the HttpKernel Component itself tries to handle an exception by logging it (when a logger is available) and by executing a controller which may render a page with some information about the error. Usually this is the controller defined in config.yml:

twig:
    # points to Symfony\Bundle\TwigBundle\Controller\ExceptionController
    exception_controller: twig.controller.exception:showAction

Another important listener is the ExceptionListener from the Security Component. This listener checks if the original exception that was thrown is an instance of AuthenticationException or AccessDeniedException. In the first case, it starts the authentication process if this is possible. In the second case, it tries to re-authenticate the user or lets the access denied handler deal with the situation.

4. Sub-requests

Maybe you noticed the $type argument of the HttpKernel::handle() method:

public function handle(
    Request $request,
    $type = HttpKernelInterface::MASTER_REQUEST,
    $catch = true
) {
    ...
}

There are two request types defined as constants in HttpKernelInterface:

  1. HttpKernelInterface::MASTER_REQUEST, the master request
  2. HttpKernelInterface::SUB_REQUEST, a sub-request

For each request to your PHP application, the first request that is handled by the kernel is of the first type, HttpKernelInterface::MASTER_REQUEST. This is rather implicit, since it is caused by leaving the $type argument away in the front controller (app.php or app_dev.php):

$response = $kernel->handle($request);

Many event listeners listen to the kernel events as discussed above, but only act when the request is of type HttpKernelInterface::MASTER_REQUEST. For instance, the Firewall does not do anything if the request is a sub-request:

public function onKernelRequest(GetResponseEvent $event)
{
    if (HttpKernelInterface::MASTER_REQUEST !== $event->getRequestType()) {
        return;
    }

    ...
}

4.1 When are sub-requests used?

Sub-requests are used to isolate the creation of a Response object. For example, when an exception is caught by the kernel, the standard exception handler tries to execute a designated exception controller (see above). To do this, a sub-request is created:

public function onKernelException(GetResponseForExceptionEvent $event)
{
    $request = $event->getRequest();

    ...

    $request = $request->duplicate(null, null, $attributes);
    $request->setMethod('GET');

    $response = $event
        ->getKernel()
        ->handle($request, HttpKernelInterface::SUB_REQUEST, true);

    $event->setResponse($response);
}

Also, whenever you render another controller as part of a Twig template, a sub-request is created and handled:

{{ render(controller('BlogBundle:Post:list')) }}

4.2 The request stack

For each new request, starting with the master request, a Request object is added to the RequestStack (available as the request_stack service since Symfony 2.4). This object holds all Request objects in memory, until they have been processed. At all times you can call its getCurrentRequest() method, to find out which request is currently being handled.

When the HttpKernel has fully handled a master or sub-request, i.e. either a normal or an exception response has been produced for it, the kernel dispatches a KernelEvents::FINISH_REQUEST (kernel.finish_request) event (available since Symfony 2.4). Event listeners which listen to this kernel.finish_request event can use the request stack to retrieve the current request and/or its parent request and, for example, restore the locale or the routing context of the previous request.

The request stack has been introduced to overcome problems with the unpredictability of the existing request service: it changed during the lifetime of the process, but previously injected instances of the request service weren’t automatically updated (of course). Before the request stack became available, the concept of service container scopes had been introduced to tackle this problem, but it came with its own problems.

Both solutions (the request stack and container scopes) aren’t proper solutions. “The current request” is just not something that should be injected as a constructor argument. It’s not a stateless thing which can be trusted to stay the same - it’s a mutable thing, which even gets replaced multiple times during the lifetime of the process.

The request stack itself is disguised as a stateless service, but is quite stateful nonetheless. It’s mutable and unpredictable. Sometimes you may call getCurrentRequest() and receive a Request object, other times you may call the same method and receive null. This is very confusing.

When a service needs to take something from the current Request object, the best solution is to never inject the request nor the request_stack service as a constructor argument. Only pass the current Request as an argument when calling a method on the service. This means you should start passing the Request object along when calling services in your controller.

4.3 Terminating the kernel

In /web/app.php and the likes, you can see that after sending the response to the client, the kernel is explicitly terminated:

$response->send();
$kernel->terminate($request, $response);

The kernel itself terminates the HttpKernel, which dispatches a KernelEvents::TERMINATE (kernel.terminate) event. This allows event listeners to perform any task that could be postponed until after the client has received the response. By postponing certain actions, like sending emails, purging caches, etc. you can sometimes reduce the response time for the client significantly. If something goes wrong though, the client won’t see an error message. The error will only appear in your log files (if you have enabled logging).

A sample kernel terminate listener looks like this:

use Symfony\Component\HttpKernel\Event\PostResponseEvent;

class PurgeCacheEventListener
{
    ...

    public function onKernelTerminate(PostResponseEvent $event)
    {
        // purge a cache, which might take some time
    }
}

You can retrieve the HttpKernel instance as well as the master Request and Response from the event object.

Please note that it’s always a good idea to move the actual logic outside of the event listener, thereby decoupling the reaction (listening to the kernel.terminate event) from the action it triggers (in this example: purging the cache).

Notable kernel.terminate event listeners

When the kernel is terminated the ProfilerListener saves all profiles (information about Doctrine queries, dispatched events, memory usage, etc.) that have been created for the current request, thereby reducing the response time. Also, the EmailSenderListener from the SwiftMailerBundle sends any emails that have been “sent” during the request. This significantly reduces response times.

Please note that offloading work until after the response has been sent to the client only works when you use PHP-FPM to handle your PHP requests.

II Patterns of dependency injection

5. What is a bundle?

As we saw in the previous chapter: running a Symfony application means booting the kernel and handling a request or executing a command, where booting the kernel means: loading all bundles and registering their service container extensions (which can be found in the DependencyInjection folder of a bundle). The container extension usually loads a services.xml file (but this can be anything) and the bundle’s configuration, defined in a separate class, usually in the same namespace, called Configuration. These things together (bundle, container extension and configuration) can be used to wire up your bundle: you can define parameters and services so that the functionality you provide inside your bundle is available to other parts of the application. You can even go one step further and register compiler passes to further modify the service container before it gets its final form.

After creating many bundles, I concluded that much of my work as a developer which is specific for a Symfony application consists of writing code for exactly these things: bundle, extension and configuration classes and compiler passes. When you know how to write good code, you still need to know how to create good bundles, and this basically means that you need to know how to create good service definitions. There are many ways to do this, and in this chapter I will describe most of them. Knowing your options enables you to make better choices when looking for a dependency injection pattern.

6. Service patterns

A service is an object, registered at the service container under a certain id. A service definition is a bit of configuration to define such a service, so that it can be instantiated at any time by the service container.

6.1 Required dependencies

Most objects need some other objects and maybe a scalar value (like an API key) or an array of values (either scalar values or objects) to be able to do their work. These are called its dependencies. We will first discuss how you can define required dependencies for services.

Required constructor arguments

The usual way to make sure a service gets its dependencies is by injecting them as constructor arguments:

class TokenProvider implements TokenProviderInterface
{
    private $storage;

    public function __construct(TokenStorageInterface $storage)
    {
        $this->storage = $storage;
    }
}

The service definition for the TokenProvider class would look like this:

<service id="token_provider" class="TokenProvider">
  <argument type="service" id="token_storage" />
</service>

<service id="token_storage" class="...">
</service>

The first argument of the token_provider service is a reference to the token_storage service. The class used for the storage should therefore implement TokenStorageInterface, or else it is not a correct argument and you get a Fatal error.

Abstract definitions for extra arguments

Say you have another token provider, an ExpiringTokenProvider, which extends from TokenProvider but has an extra constructor argument, $lifetime:

class ExpiringTokenProvider extends TokenProvider
{
    private $lifetime;

    public function __construct(TokenStorageInterface $storage, $lifetime)
    {
        $this->lifetime = $lifetime;

        parent::__construct($storage);
    }
}

When creating a service definition for this second token provider, you could just copy the argument from the existing token_provider definition:

<service id="expiring_token_provider" class="ExpiringTokenProvider">
  <argument type="service" id="token_storage" />
  <argument>3600</argument><!-- lifetime -->
</service>

In these situations it’s better to create a parent service definition, which you can use to define everything that its child definitions should have in common:

<service id="abstract_token_provider" abstract="true">
  <argument type="service" id="token_storage" />
</service>

<service id="token_provider" class="TokenProvider"
  parent="abstract_token_provider">
</service>

<service id="expiring_token_provider" class="ExpiringTokenProvider"
  parent="abstract_token_provider">
  <argument>3600</argument><!-- lifetime -->
</service>

The abstract service has one argument and is marked as abstract. The token provider services mention abstract_token_provider as their parent. The token_provider service has no extra arguments, so it just inherits the first constructor argument from abstract_token_provider. The expiring_token_provider service also inherits the token_storage service as the first argument, but adds an extra argument for $lifetime.

Decorating services

When using abstract service definitions like this you basically introduce a hierarchy in your service definitions. The hierarchy reflects the hierarchy of classes: ExpiringTokenProvider extends TokenProvider. If you’ve been doing object-oriented programming for a while, you may know that inheritance is often not the best solution.

A hierarchy can only be formed in one way because PHP supports only single class inheritance and Symfony supports only single service definition inheritance. This greatly reduces the flexibility of your code. There’s no way to easily rearrange the hierarchy without modifying a lot of code and service definitions, thereby endangering the stability of your application.

The solution is to use composition instead of inheritance. Applied to the example of the ExpiringTokenProvider: instead of extending TokenProvider, we accept an instance of TokenProviderInterface:

class ExpiringTokenProvider implements TokenProviderInterface
{
    private $innerTokenProvider;
    private $lifetime;

    public function __construct(
        TokenProviderInterface $innerTokenProvider,
        $lifetime
    ) {
        $this->innerTokenProvider = $innerTokenProvider;

        $this->lifetime = $lifetime;
    }
}

As you can see, ExpiringTokenProvider inherits nothing from TokenProvider anymore. It only uses TokenProvider via its interface TokenProviderInterface (by doing so we nicely follow the Dependency inversion principle as well).

The corresponding service definitions would look like this:

<service id="default_token_provider" class="TokenProvider">
  <argument type="service" id="token_storage" />
</service>

<service id="expiring_token_provider" class="ExpiringTokenProvider">
  <argument type="service" id="default_token_provider" />
  <argument>3600</argument><!-- lifetime -->
</service>

<!-- we create an alias to make "token_provider" available as a service -->
<service id="token_provider" alias="expiring_token_provider" />

As of version 2.5 Symfony supports decoration for service definitions as well:

<service id="token_provider" class="TokenProvider">
  <argument type="service" id="token_storage" />
</service>

<service
  id="expiring_token_provider"
  class="ExpiringTokenProvider"
  decorates="token_provider">
  <argument type="service" id="expiring_token_provider.inner" />
  <argument>3600</argument><!-- lifetime -->
</service>

As you can see, the token_provider service will be renamed to expiring_token_provider.inner. It will be injected as the first constructor argument of expiring_token_provider. Symfony will automatically define an alias token_provider which points to expiring_token_provider.

Even though this may seem quite nice, decorating service definitions like this can be quite confusing. The code above was a bit more explicit and understandable after a quick glance.

Method calls

In some cases you don’t want to or can’t override the constructor to add extra required arguments, or some of the dependencies are not yet determined at the time the service gets created. For these situations, you may add setter methods to your class, to allow someone to inject a dependency immediately after the service was created (or in fact, at any moment afterwards):

class SomeController
{
    private $container;

    public function setContainer(ContainerInterface $container)
    {
        $this->container = $container;
    }

    public function indexAction()
    {
        $service = $this->container->get('...');
    }
}

Since (as you can see in the example above) an instance of ContainerInterface really is required to run the indexAction() method, this is a required setter call. So in your service definition you should take care that the service container gets injected:

<service id="some_controller" class="SomeController">
  <call method="setContainer">
    <argument type="service" id="service_container" />
  </call>
</service>

The advantage of using setters for dependency injection is that you are not required to have a constructor argument for the dependency anymore. Sometimes this means that you don’t need to have a constructor at all, or that you can leave an existing constructor as it is. The disadvantage is that you may forget to call the setter, so that one of the dependencies of the object is missing. In the example above, a call will then be made to the get() method of a non-object (null) which results in PHP throwing a Fatal error. In my opinion, this one disadvantage, is usually much bigger than any advantage you can think of, since it introduces a code smell called “temporal coupling”. It thereby makes your class somewhat unreliable.

To prevent these severe crashes from happening (and to help a developer who encounters the error to fix the problem) you may choose to wrap calls to dependencies that should have been injected using a setter:

class SomeController
{
    public function indexAction()
    {
        $service = $this->getContainer()->get('...');
    }

    private function getContainer()
    {
        if (!($this->container instanceof ContainerInterface)) {
            throw new \RuntimeException('Service container is missing');
        }

        return $this->container;
    }
}

This is not an optimal solution of course, since you might still accidentally use the $container property itself. For some more waring notes about injecting the service container, take a look at the chapter Constructor arguments over fetching container services

Method calls in abstract definitions

When you create container-aware services, you will have much code duplication in your service definitions. It may then be a good idea to add the call to setContainer() to an abstract service definition:

<service id="abstract_container_aware" abstract="true">
  <call method="setContainer">
    <argument type="service" id="service_container" />
  </call>
</service>

<service id="some_controller" class="SomeController"
  parent="abstract_container_aware">
</service>

6.2 Optional dependencies

Sometimes dependencies are optional. When you think about the term “optional dependencies” this feels a bit like a contradiction, because if you don’t really depend on them, they are not “dependencies”. However, there are situations where a service knows how to use another service, but does not actually need it for doing its job. For instance, a service may know how to deal with a logger to log some things for debugging purposes.

Optional constructor arguments

In case your service’s class knows how to work with a logger, it may have an optional constructor argument for it:

use Symfony\Component\EventDispatcher\EventDispatcherInterface;
use Psr\Log\LoggerInterface;

class AuthenticationListener
{
    private $eventDispatcher;
    private $logger;

    public function __construct(
        EventDispatcherInterface $eventDispatcher,
        LoggerInterface $logger = null
    ) {
        $this->eventDispatcher = $eventDispatcher;
        $this->logger = $logger;
    }

    public function onAuthenticationSuccess(AuthenticationEvent $event)
    {
        if ($this->logger instanceof LoggerInterface) {
            $this->logger->debug('Authentication was successful');
        }

        ...
    }
}

For constructor arguments that should be either an object of the given class/interface, or nothing, you can use the default value null. Then in your service definition, choose a strategy for dealing with a missing service:

<service id="authentication_listener" class="AuthenticationListener">
  <argument type="service" id="logger" on-invalid="ignore" />
</service>

The ignore strategy is currently equivalent to the null strategy, in that it will call the constructor with a null value instead of the requested service. There is also the exception strategy, which is the default strategy. This will raise an exception when the injected service could not be found.

Transforming optional dependencies into harmless required dependencies

When your service accepts “optional” dependencies your code will end up containing lots of if (... instanceof ...) checks. This is really bad for readability. The if clauses make it harder to understand what’s going on (and under what conditions). Besides, the if clause introduces new levels of indentation, which is distracting, since the expression doesn’t contain any relevant business logic. It’s just a technical thing: has the logger service been injected, or not?

There is a very simple solution for this: we just need to turn the optional dependency into a required dependency. In the example above, we remove the = null part of the $logger constructor parameter. Then we can also remove all the if clauses checking for instanceof LoggerInterface, since we can be sure of that already.

use Symfony\Component\EventDispatcher\EventDispatcherInterface;
use Psr\Log\LoggerInterface;

class AuthenticationListener
{
    ...
    private $logger;

    public function __construct(..., LoggerInterface $logger)
    {
        ...
        $this->logger = $logger;
    }

    public function onAuthenticationSuccess(AuthenticationEvent $event)
    {
        $this->logger->debug('Authentication was successful');

        ...
    }
}

Now what if you don’t have a “logger” service in your project (or it isn’t available when running the kernel in the prod environment)? In that case you only need to provide a harmless implementation of the LoggerInterface. In this case “harmless” means that it shouldn’t really append any lines to the log file, or keep those lines in memory; it should just do nothing, like this:

class HarmlessLogger implements LoggerInterface
{
    public function debug($message, array $context = array())
    {
        // do nothing!
    }
    ...
}

This type of object is called a Null object, since it is the object equivalent of a null value.

By the way, you don’t implement your own “harmless” logger, since the psr/log comes bundles with a Psr\Log\NullLogger which does the same thing: nothing. You can simply add a service definition for the NullLogger like this:

<service id="logger" class="Psr\Log\NullLogger" />

6.3 A collection of services

In most situations you inject single services as constructor arguments. But sometimes you need to inject a collection of services, that are treated in the same way, for instance when you want to provide several alternative ways (strategies) to achieve something:

class ObjectRenderer
{
    private $renderers;

    public function __construct(array $renderers)
    {
        $this->renderers = $renderers;
    }

    public function render($object)
    {
        foreach ($this->renderers as $renderer) {
            if ($renderer->supports($object) {
                return $renderer->render($object);
            }
        }
    }
}

In a service definition this may look like:

<service id="object_renderer" class="ObjectRenderer">
  <argument type="collection">
    <argument type="service" id="domain_object_renderer" />
    <argument type="service" id="user_renderer" />
  </argument>
</service>

This collection type argument will be converted to an array, containing the services referenced by their ids:

array(
    0 => ...
    1 => ...
)

Optionally you could also give each argument inside the collection a key attribute.

<service id="object_renderer" class="ObjectRenderer">
  <argument type="collection">
    <argument
      key="domain_object" type="service"
      id="domain_object_renderer" />
    <argument
      key="user" type="service"
      id="user_renderer" />
  </argument>
</service>

The value of the key attribute will be used as the key for each value of the collection:

array(
    'domain_object' => ...
    'user' => ...
)

Multiple method calls

When you read the code of the ObjectRenderer class again in “strict” mode, it appears that you cannot trust the $renderers array to contain only valid renderers (which, let’s say, implement RendererInterface). Therefore, you may decide to dedicate a special method for adding a renderer:

class ObjectRenderer
{
    private $renderers;

    public function __construct()
    {
        $this->renderers = array();
    }

    public function addRenderer($name, RendererInterface $renderer)
    {
        $this->renderers[$name] = $renderer;
    }
}

Of course, when the name is irrelevant, leave out the $name parameter. What matters is: whenever anybody calls the addRenderer method and provides an object which is not an implementation of RendererInterface, they will not succeed, because of the type-hint mismatch.

The service definition should now be changed to call the addRenderer() method for each of the available renderers:

<service id="object_renderer" class="ObjectRenderer">
  <call method="addRenderer">
    <argument>domain_object</argument>
    <argument type="service" id="domain_object_renderer" />
  </call>
  <call method="addRenderer">
    <argument>user</argument>
    <argument type="service" id="user_renderer" />
  </argument>
</service>

The best of both worlds

It may be better to combine the two options above, allowing developers to provide both an initial collection of renderers through a constructor argument and/or to add renderers one by one using the addRenderer() method:

class ObjectRenderer
{
    private $renderers;

    public function __construct(array $renderers)
    {
        foreach ($renderers as $name => $renderer) {
            $this->addRenderer($name, $renderer);
        }
    }

    public function addRenderer($name, RendererInterface $renderer)
    {
        $this->renderers[$name] = $renderer;
    }
}
The bester of both worlds

The object_renderer service now offers two ways in which a user can add renderers. This may seem to be great from a usability perspective. However, it’s bad for the ObjectRenderer service itself: it has become mutable. A developer may call addRenderer() whenever they like. This wil change the behavior of the ObjectRenderer wherever it’s currently in use. This is a dangerous situation and will result in unpredictable, hard-to-debug behavior. To read more about stateless services, read the article with that name by Igor Wiedler.

What we were really trying to accomplish when we introduced the addRenderer() method, was to force each renderer to be of type RendererInterface. There is a better way to do this:

class ObjectRenderer
{
    private $renderers;

    public function __construct(array $renderers)
    {
        foreach ($renderers as $name => $renderer) {
            if (!$renderer instanceof RendererInterface) {
                throw new \InvalidArgumentException(
                    'Expected only instances of RendererInterface`
                );
            }

            $this->renderers[$name] = $renderer;
        }
    }
}

In general I would recommend you to use a standard assertion library for these kinds of “type” checks. For example, I like to use beberlei/assert. Using this library, you can rewrite the previous code sample like this:

class ObjectRenderer
{
    private $renderers;

    public function __construct(array $renderers)
    {
        \Assert\that($renderers)
            ->all()
            ->isInstanceOf(RendererInterface::class);

        $this->renderers = $renderers;
    }
}

Service tags

We already have a nice setup for manually adding renderers, but what if other parts of your application (like other bundles) should be able to register specific renderers? The best way to do this is by using service tags:

<!-- in some other bundle -->
<service id="date_time_renderer" class="DateTimeRenderer">
  <tag name="specific_renderer" alias="date_time" />
</service>

Each tag has a name, which you can pick yourself. Each tag can also have any number of extra attributes (like alias above). These attributes allow you to further specify any requested behavior.

To collect services with these tags, you have to create a compiler pass like this one:

namespace Matthias\RendererBundle\DependencyInjection\Compiler;

use Symfony\Component\DependencyInjection\Compiler\CompilerPassInterface;
use Symfony\Component\DependencyInjection\ContainerBuilder;
use Symfony\Component\DependencyInjection\Reference;

class RenderersPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        // collect all tagged services in the entire project
        $taggedServiceIds
            = $container ->findTaggedServiceIds('specific_renderer');

        $objectRendererDefinition
            = $container->getDefinition('object_renderer');

        foreach ($taggedServiceIds as $serviceId => $tags) {

            // services can have many tag elements with the same tag name
            foreach ($tags as $tagAttributes) {

                // call addRenderer() to register this specific renderer
                $objectRendererDefinition
                    ->addMethodCall('addRenderer', array(
                        $tagAttributes['alias'],
                        new Reference($serviceId),
                    ));
            }
        }
    }
}

Register this compiler pass in your bundle class:

use Matthias\RendererBundle\DependencyInjection\Compiler\RenderersPass;
use Symfony\Component\DependencyInjection\ContainerBuilder;

class RendererBundle extends Bundle
{
    public function build(ContainerBuilder $container)
    {
        $container->addCompilerPass(new RenderersPass());
    }
}

Inside the process() method of the compiler pass, first all tags with the name specific_renderer are collected. This will result in an array of which the keys are service ids and the values are arrays of attribute arrays. This is because each service definition can have multiple tags with the same name (but maybe with different attributes).

Then, the service definition for the ObjectRenderer class is retrieved and while iterating over the tags, a Reference is created which refers to each renderer service that is tagged as “specific_renderer” and together with the provided value for alias, these are used as arguments to a call to the method addRenderer().

All of this means that when the object_renderer service is requested, first of all an instance of ObjectRenderer gets created. But afterwards, some calls are made to its addRenderer() method to add the specific renderers tagged as specific_renderer.

Single method call

There are many possible approaches when collecting services in a compiler pass. For instance, you can collect service references in an array and set them all at once, by adding a method call to setRenderers():

class RenderersPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        $taggedServiceIds = ...;

        $objectRendererDefinition = ...;

        $renderers = array();

        foreach ($taggedServiceIds as $serviceId => $tags) {
            foreach ($tags as $tagAttributes) {
                $name = $tagAttributes['alias'];
                $renderer = new Reference($serviceId);
                $renderers[$name] = $renderer;
            }
        }

        $objectRendererDefinition
            ->addMethodCall('setRenderers', array($renderers));
    }
}

Replacing a single argument

When it’s possible - like in one of the examples above - to inject a collection of renderers as a constructor argument, there is another way you can do this: set a constructor argument directly:

class RenderersPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        $taggedServiceIds = ...;

        $objectRendererDefinition = ...;

        $renderers = array();

        // collect service references
        ...

        $objectRendererDefinition->replaceArgument(0, $renderers);
    }
}

*Replacing an argument can only be done when you have defined it in the first place (for instance as an empty argument):

<service id="object_renderer" class="ObjectRenderer">
  <argument /><!-- specific renderers -->
</service>

Service ids instead of references

Whenever you request the object_renderer service, all the specific renderers are instantiated too. Depending on the cost of instantiating these renderers it could be a good idea to add support for lazy-loading. This can be accomplished by making the ObjectRenderer container-aware and by injecting service ids, not the services themselves:

class LazyLoadingObjectRenderer
{
    private $container;
    private $renderers;

    public function __construct(ContainerInterface $container)
    {
        $this->container = $container;
    }

    public function addRenderer($name, $renderer)
    {
        $this->renderers[$name] = $renderer;
    }

    public function render($object)
    {
        foreach ($this->renderers as $name => $renderer) {
            if (is_string($renderer)) {
                // $renderer is assumed to be a service id
                $renderer = $this->container->get($renderer);
            }

            // check if the renderer is an instance of RendererInterface
            ...
        }
    }
}

The compiler pass should be modified to not pass references to services, but just service ids:

class RenderersPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        $taggedServiceIds = ...;

        $objectRendererDefinition = ...;

        foreach ($taggedServiceIds as $serviceId => $tags) {
            foreach ($tags as $tagAttributes) {
                $objectRendererDefinition
                    ->addMethodCall('addRenderer', array(
                        $tagAttributes['alias'],
                        // don't create a Reference, just pass the string
                        $serviceId,
                    ));
            }
        }
    }
}

Also don’t forget to provide the service container as a constructor argument:

<service id="object_renderer" class="LazyLoadingObjectRenderer">
  <argument type="service" id="service_container" />
</service>

Of course, any of the strategies mentioned above can be used with this lazy-loading class (single method call, multiple method calls or argument replacement).

Before you consider changing your class to make use of the service container directly, please read the chapter Reduce coupling to the framework, specifically The performance issue.

6.4 Delegated creation

Instead of fully defining services upfront using a service definition with a class, arguments and method calls, you can also leave the details to be figured out at runtime, by delegating the creation of services to a factory method. Factory methods can be either static methods or object methods. In the first case, you can provide the class name and the method name as attributes of the service definition:

<service id="some_service" class="ClassOfResultingObject"
  factory-class="Some\Factory" factory-method="create">
  <argument>...</argument>
</service>

When some_service is being requested for the first time, the service will be retrieved by calling Some\Factory::create() statically with any arguments provided. The result will be stored in memory, so the factory method is called only once.

Most factory methods nowadays are not static anymore, which means that the factory method should be called on an instance of the factory itself. This instance should be defined as a service:

<!-- a service created by some_factory_service -->
<service id="some_service" class="ClassOfResultingObject"
  factory-service="some_factory_service" factory-method="create">
  <argument>...</argument>
</service>

<!-- the factory -->
<service id="some_factory_service" class="Some\Factory">
</service>

Not so useful

Though the options for delegating the creation of services to other services seem really great, I have not used them very often. They are almost exclusively useful when you are creating service definitions for older PHP classes since in the (not so distant) past, creation logic was often hidden inside static factory classes (remember Doctrine_Core::getTable()?).

My objection to factory classes with static factory methods is that static code is global code and that executing that code may have side effects which can not be isolated (for instance in a test scenario). Besides, any dependency of such a static factory method has to be by definition static itself, which is also really bad for isolation and prevents you from replacing (part of) the creation logic by your own code.

Factory objects (or factory services) are slightly better. However, the need for them very likely points to some kind of design problem. A service should not need a factory, since it will be created only once in a predetermined (and deterministic) way and from then on be perfectly reusable by any other object. The only things that are dynamic about a service, should be the arguments of the methods that are part of its public interface (see also State and context).

Service factories: sometimes useful

One particularly nice example of using a factory service and method for retrieving a service is the case of a Doctrine repository. When you need one, you would normally inject an entity manager as a constructor argument and later retrieve a specific repository:

use Doctrine\ORM\EntityManager;

class SomeClass
{
    public function __construct(EntityManager $entityManager)
    {
        $this->entityManager = $entityManager;
    }

    public function doSomething()
    {
        $repository = $this->entityManager->getRepository('User');

        ...
    }
}

But using a factory service and method you could directly inject the correct repository itself:

class SomeClass
{
    public function __construct(UserRepository $userRepository)
    {
        $this->userRepository = $userRepository;
    }
}

This is the corresponding service definition:

<service id="some_service" class="SomeClass">
  <argument type="user_repository" />
</service>

<service id="user_repository" class="UserRepository"
  factory-service="entity_manager" factory-method="getRepository">
  <argument>User</argument>
</service>

By looking at the constructor arguments of SomeClass it is immediately clear that it needs a User repository, which is much more specific and communicative than the earlier example in which SomeClass needed an EntityManager. Besides making the class itself much cleaner, it will also make it much easier to create a stand-in object for the repository when you are writing a unit test for this class. Instead of creating a mock for both the entity manager and the repository, you only have to create one for the repository itself.

6.5 Manually creating services

Usually you create services by loading service definitions from a file:

use Symfony\Component\HttpKernel\DependencyInjection\Extension;
use Symfony\Component\Config\FileLocator;
use Symfony\Component\DependencyInjection\Loader\XmlFileLoader;

class SomeBundleExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $locator = new FileLocator(__DIR__.'/../Resources/config');
        $loader = new XmlFileLoader($container, $locator);
        $loader->load('services.xml');
    }
}

But some services can not be defined in a configuration file. They should be defined dynamically, because their name, class, arguments, tags, etc. are not fixed.

Definition

Manually creating a service definition means creating an instance of Definition, and optionally providing a class name. The definition will get its identifier when it is set on the ContainerBuilder instance:

use Symfony\Component\DependencyInjection\Definition;

$class = ...; // set a class name for the definition

$definition = new Definition($class);

$container->setDefinition('the_service_id', $definition);

The equivalent of this in XML would be:

<service id="the_service_id" class="...">
</service>

You can make the definition non-public if it only exists as a dependency of other services:

$definition->setPublic(false);

Arguments

When the service requires some constructor arguments, you may set them all at once:

use Symfony\Component\DependencyInjection\Reference;

$definition->setArguments(array(
    new Reference('logger') // reference to another service
    true // boolean argument,
    array(
        'table_name' => 'users'
    ) // array argument
    ...
);

Arguments should be either references to other services, array values or scalar values (or a mix of these). This is because all service definitions will eventually be stored as a simple PHP file. A reference to another service can be created by using a Reference object with the id of the service that should be injected.

You can also add the arguments one by one, in the right order:

$definition->addArgument(new Reference('logger'));
$definition->addArgument(true);
...

Finally, when you are modifying an existing service definition with a list of arguments, you could replace them by providing their numeric index:

$definition->setArguments(array(null, null));

...

$definition->replaceArgument(0, new Reference('logger'));
$definition->replaceArgument(1, true);

The equivalent in XML would be:

<service id="..." class="...">
  <argument type="service" id="logger" />
  <argument>true</argument>
</service>

Tags

There is another thing you may want to do when working with Definition objects: adding tags to them. A tag consists of the name of the tag and an array of attributes. A definition can have multiple tags with the same tag name:

$definition->addTag('kernel.event_listener', array(
    'event' => 'kernel.request'
);
$definition->addTag('kernel.event_listener', array(
    'event' => 'kernel.response'
);

In XML you would write this like:

<service id="..." class="...">
  <tag name="kernel.event_listener" event="kernel.request">
  <tag name="kernel.event_listener" event="kernel.response">
</service>

Aliases

Before talking about what you can do with all this knowledge, there is one last thing you’ll need: creating aliases for services:

$container->setAlias('some_alias', 'some_service_id');

Now whenever you request the service some_alias, you will in fact get the service some_service_id.

6.6 The Configuration class

Before we continue, I need to explain a few things about the Configuration class. You may have noticed it earlier, and maybe you have even created one yourself.

Most of the times you will use a Configuration class to define all the possible configuration options for your bundle (though the Config Component is completely decoupled from the HttpKernel Component, so in theory you could use any other means for defining a bundle configuration). The name of the class or its namespace is actually irrelevant, as long as it implements ConfigurationInterface:

use Symfony\Component\Config\Definition\ConfigurationInterface;
use Symfony\Component\Config\Definition\Builder\TreeBuilder;

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('name_of_bundle');

        $rootNode
            ->children()
                // define configuration nodes
                ...
            ->end()
        ;

        return $treeBuilder;
    }
}

There is one public method: getConfigTreeBuilder(). This method should return a TreeBuilder instance which is a builder you use to describe all configuration options, including their validation rules. Creating a config tree starts with defining a root node:

$rootNode = $treeBuilder->root('name_of_bundle');

The name of the root node should be the name of the bundle, without “bundle”, but lower-cased and with underscores. So the node name for MatthiasAccountBundle will be matthias_account.

The root node is an array node. It can have any child node you like:

$rootNode
    ->children()
        ->booleanNode('auto_connect')
            ->defaultTrue()
        ->end()
        ->scalarNode('default_connection')
            ->defaultValue('default')
        ->end()
    ->end()
;

Usually, you will use an instance of the Configuration class inside a bundle’s extension class, to process a given set of configuration arrays. These configuration arrays have been collected by the kernel, by loading all the relevant configuration files (like config_dev.yml, config.yml, parameters.yml, etc.).

class MatthiasAccountExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );
    }
}

The processConfiguration() method of the Extension class instantiates a Processor and finalizes the config tree retrieved from the Configuration object. It then asks the processor to process (validate and merge) the raw configuration arrays:

final protected function processConfiguration(
    ConfigurationInterface $configuration,
    array $configs
) {
    $processor = new Processor();

    return $processor->processConfiguration($configuration, $configs);
}

When there were no validation errors, you can then use the configuration values in any way you like. You can define or modify container parameters or service definitions based on the configuration values. In the following chapters we will discuss many different ways to do this.

6.7 Dynamically add tags

Say you want to create a generic event listener, which listens to a configurable list of events, like kernel.request, kernel.response, etc. This is what your Configuration class might look like:

use Symfony\Component\Config\Definition\ConfigurationInterface;

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('generic_listener');

        $rootNode
            ->children()
                ->arrayNode('events')
                    ->prototype('scalar')
                    ->end()
                ->end()
            ->end()
        ;

        return $treeBuilder;
    }
}

It allows a list of event names to be configured like this:

generic_listener:
    events: [kernel.request, kernel.response, ...]

The standard way to register an event listener would be to add tags to the event listener’s service definition in services.xml:

<service id="generic_event_listener" class="...">
  <tag name="kernel.event_listener" event="..." method="onEvent" />
  <tag name="kernel.event_listener" event="..." method="onEvent" />
</service>

But in this situation, we don’t know in advance to which events the listener should listen, so these events can not be defined in a configuration file. Luckily - as we saw earlier - we can add tags to service definitions on the fly. This can be done inside the container extension:

class GenericListenerExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        // load services.xml
        $loader = ...;
        $loader->load('services.xml');

        $eventListener = $container
            ->getDefinition('generic_event_listener');

        foreach ($processedConfig['events'] as $eventName) {
            // add a kernel.event_listener tag for each event
            $eventListener
                ->addTag('kernel.event_listener', array(
                    'event' => $eventName,
                    'method' => 'onEvent'
                ));
        }
    }
}

There is one extra step you might take to prevent a dangling listener service when there are no events to which it should listen:

if (empty($processedConfig['events'])) {
    $container->removeDefinition('generic_event_listener');
}

6.8 Strategy pattern for loading exclusive services

Often bundles provide multiple ways of doing a single thing. For instance, a bundle that provides some kind of mailbox functionality may have different storage implementations, like one storage manager for Doctrine ORM and one for MongoDB. To make the choice for a specific storage manager configurable, create a Configuration class like this:

use Symfony\Component\Config\Definition\ConfigurationInterface;

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('browser');

        $rootNode
            ->children()
                ->scalarNode('storage_manager')
                    ->validate()
                        ->ifNotInArray(array('doctrine_orm', 'mongo_db')
                        ->thenInvalid('Invalid storage manager')
                    ->end()
                ->end()
            ->end()
        ;

        return $treeBuilder;
    }
}

Then given there are two service definition files for each of the storage managers, like doctrine_orm.xml:

<services>
  <service id="mailbox.doctrine_orm.storage_manager" class="...">
  </service>
</services>

And mongo_db.xml:

<services>
  <service id="mailbox.mongo_db.storage_manager" class="...">
  </service>
</services>

You could then load either of these files by doing something like this in your container extension:

class MailboxExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        // create an XmlLoader
        $loader = ...;

        // load only the services for the given storage manager
        $storageManager = $processedConfig['storage_manager'];
        $loader->load($storageManager.'.xml');

        // make the specific storage manager available as the general one
        $container->setAlias(
            'mailbox.storage_manager',
            'mailbox.'.$storageManager.'.storage_manager'
        );
    }
}

A convenient alias is created in the end to allow other parts of the application to just request the mailbox.storage_manager service, instead of worrying about the storage-specific service that should be used. However, the way this is done is too rigid: the id of each specific storage manager service should conform to the pattern mailbox.{storageManagerName}.storage_manager. It would be better to define the alias inside the service definition files themselves:

<services>
  <service id="mailbox.doctrine_orm.storage_manager" class="...">
  </service>

  <service id="storage_manager"
    alias="mailbox.doctrine_orm.storage_manager">
  </service>
</services>

Using the strategy pattern for loading service definitions has many advantages:

  • Only the services that are useful in the current application will be loaded. When you don’t have a MongoDB server up and running, there will be no services that accidentally refer to it.
  • The setup is open for extension, since you can add the name of another storage manager to the list in the Configuration class and then add a service definition file with the necessary services and an alias.

6.9 Loading and configuring additional services

Say you have a bundle dedicated to input filtering. Probably you offer several different services, like services for filtering form data, and services for filtering data stored using Doctrine ORM. It should be possible to enable or disable any of these services or collections of services at any time because they may not all be applicable to your specific situation. There is a handy shortcut for configuration definitions to accomplish a thing like this:

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('input_filter');

        $rootNode
            ->children()
                ->arrayNode('form_integration')
                    // will be enabled by default
                    ->canBeDisabled()
                ->end()
                ->arrayNode('doctrine_orm_integration')
                    // will be disabled by default
                    ->canBeEnabled()
                ->end()
            ->end()
        ;

        return $treeBuilder;
    }
}

With a configuration tree like this, you can enable or disable specific parts of the bundle in config.yml:

input_filter:
    form_integration:
        enabled: false
    doctrine_orm_integration:
        enabled: true

Inside your container extension you can then load the appropriate services:

class InputFilterExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        if ($processedConfig['doctrine_orm_integration']['enabled']) {
            $this->loadDoctrineORMIntegration(
                $container,
                $processedConfig['doctrine_orm_integration']
            );
        }

        if ($processedConfig['form_integration']['enabled']) {
            $this->loadFormIntegration(
                $container,
                $processedConfig['form_integration']
            );
        }

        ...
    }

    private function loadDoctrineORMIntegration(
        ContainerBuilder $container,
        array $configuration
    ) {
        // load services, etc.
        ...
    }

    private function loadFormIntegration(
        ContainerBuilder $container,
        array $configuration
    ) {
        ...
    }
}

Each of the stand-alone parts of the bundle can be loaded separately like this.

A cleaner configuration class

One or two of these stand-alone bundle parts can be easily handled, but soon the Configuration class will contain many lines of code in just one method. You can clean this up a bit by using the append() method in combination with some private methods:

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();

        $rootNode = $treeBuilder->root('input_filter');

        $rootNode
            ->append($this->createFormIntegrationNode())
            ->append($this->createDoctrineORMIntegrationNode())
        ;

        return $treeBuilder;
    }

    private function createDoctrineORMIntegrationNode()
    {
        $builder = new TreeBuilder();

        $node = $builder->root('doctrine_orm_integration');

        $node
            ->canBeEnabled()
            ->children()
                // maybe add some more configuration
                ...
            ->end();

        return $node;
    }

    private function createFormIntegrationNode()
    {
        ...
    }
}

6.10 Configure which services to use

Instead of using the Strategy pattern for loading services you may also allow developers to manually configure a service they want to use. For example, if your bundle needs some kind of encrypter service and the bundle does not provide one itself, you would want to ask the developer to provide the encrypter by its service id:

matthias_security:
    encrypter_service: my_encrypter_service_id

Your Configuration class should then look like this:

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('matthias_security');

        $rootNode
            ->children()
                ->scalarNode('encrypter_service')
                    ->isRequired()
                ->end()
            ->end()
        ;

        return $treeBuilder;
    }
}

Inside the bundle’s extension class, you could then create an alias for the configured service.

class MatthiasSecurityExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        $container->setAlias(
            'matthias_security.encrypter',
            $processedConfig['encrypter_service']
        );
    }
}

So even though the service id of the encrypter may originally be anything, now you have a stable reference to it - an alias - which you can use inside any of your bundle’s service definitions:

<service id="matthias_security.encrypted_data_manager" class="...">
  <argument type="service" id="matthias_security.encrypter" />
</service>

Of course the assumption here is that the manually configured encrypter service is a valid encrypter object. You can not be sure at configuration time that this is true, so you will have to verify this at runtime. The usual way to do this would be to add the appropriate type-hint to classes of services that use the encrypter service:

class EncryptedDataManager
{
    public function __construct(EncrypterInterface $encrypter)
    {
        // $encrypter is a valid encrypter
    }
}

6.11 Completely dynamic service definitions

There are situations when you know almost nothing about the services you need, until you have processed the configuration. Say you want the users of your bundle to define a set of resources as services. These resources can be of type directory or file. You want to create these services on the fly since they are different for each application and you need to collect them using a custom service tag, called resource. Your Configuration class may look like this:

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('resource_management');

        $rootNode
            ->children()
                ->arrayNode('resources')
                    ->prototype('array')
                        ->children()
                            ->enumNode('type')
                                ->values('directory', 'file')
                            ->end()
                            ->scalarNode('path')
                            ->end()
                        ->end()
                    ->end()
                ->end()
            ->end()
        ;

        return $treeBuilder;
    }
}

An example resource configuration:

resource_management:
    resources:
        global_templates:
            type: directory
            path: Resources/views
        app_kernel:
            type: file
            path: AppKernel.php

When the resources are defined like this, you can create service definitions for them in the container extension:

class ResourceManagementExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );

        $resources = $processedConfig['resources'];

        foreach ($resources as $name => $resource) {
            $this->addResourceDefinition($container, $name, $resource);
        }
    }

    private function addResourceDefinition(
        ContainerBuilder $container,
        $name,
        array $resource
    ) {
        // determine the class
        $class = $this->getResourceClass($resource['type']);

        $definition = new Definition($class);

        // add a specific tag
        $definition->addTag('resource');

        $serviceId = 'resource.'.$name;

        $container->setDefinition($serviceId, $definition);
    }

    private function getResourceClass($type)
    {
        if ($type === 'directory') {
            return 'Resource\Directory';
        } elseif ($type === 'file') {
            return 'Resource\File';
        }

        throw new \InvalidArgumentException('Type not supported');
    }
}

When these manually created service definitions need arguments, method calls, etc. use the techniques described above to add them dynamically too.

7. Parameter patterns

The service container has service definitions and parameters. Parameters are simple values that can consist of constant scalar values and arrays, in any combination. So 'matthias' would be a valid parameter, and so would 23 be. But also array(23 => 'matthias'), array(23 => array('matthias')), etc.

You can define parameters using the key of your choice. The naming convention however would be: name_of_your_bundle_without_bundle.parameter_name to prevent collisions with parameters defined in other bundles. These parameters can be defined in various places.

7.1 Parameters.yml

Some of the more essential parameters of your application (which probably have no default value) can be found in /app/config/parameters.yml. Parameters are loaded together with service definitions and configuration values for container extensions. That is why the standard config.yml starts with these lines:

imports:
    - { resource: parameters.yml }
    - { resource: security.yml }

framework:
    secret:          %secret%
    ...

First parameters.yml and security.yml will be imported. The file parameters.yml starts with:

parameters:
    database_driver: pdo_mysql
    ...

And security.yml starts with:

security:
    encoders:
        Symfony\Component\Security\Core\User\User: plaintext
    ...

These files will be imported as they are. So config.yml could just as well look like:

parameters:
    ...
security:
    ...
framework:
    ...

Since all configuration arrays will be merged in the end, it is possible in config.yml to override any parameter defined in parameters.yml:

parameters:
    ... # loaded from parameters.yml
    database_driver: pdo_sqlite

Even service definitions can be created within config.yml (or any service container configuration file for that matter):

parameters:
    ...
services:
    some_service_id:
        class: SomeClass

7.2 Parameter resolving

Values defined in config.yml, or parameters.yml but also in service definitions and definition arguments can contain placeholders for values that should be defined elsewhere as parameters. When the service container gets compiled, values containing placeholders will be resolved. For instance in the example above we defined the database_driver parameter in parameters.yml. In config.yml we can refer to this parameter using the %database_driver% placeholder:

doctrine:
    dbal:
        driver: %database_driver%

When creating service definitions, the Symfony bundles usually take the same approach when it comes to the class names used:

<parameters>
    <parameter key="form.factory.class">
        Symfony\Component\Form\FormFactory
    </parameter>
</parameters>

<service id="form.factory" class="%form.factory.class%">
</service>

Parameters for class names

Using parameters for class names would allow other parts of the application to replace the parameter, instead of directly modifying the service definition. It is not very likely however that this will happen, and when every service definition has its class in a parameter, this parameter also ends up in the final service container, so that even at runtime you could call $container->getParameter('form.factory.class') and retrieve the class name of the form factory. This seems very redundant to me and I would not recommend it when creating your own service definitions.

Whenever you would want to change the class of a definition, you could do this inside a compiler pass:

use Symfony\Component\DependencyInjection\ContainerBuilder;
use Symfony\Component\DependencyInjection\Compiler\CompilerPassInterface;

class ReplaceClassCompilerPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        $myCustomFormFactoryClass = ...;

        $container
            ->getDefinition('form.factory')
            ->setClass($myCustomFormFactoryClass);
    }
}

Manually resolving parameters

When you are using parameters in your bundle extension (or in a compiler pass), the values of these parameters are not fully resolved yet. For example, your bundle may define a parameter my_cache_dir, referring to the %kernel.cache_dir% which contains the location of the cache directory used by the kernel:

parameters:
    my_cache_dir: %kernel.cache_dir%/my_cache

In the load() method of your container extension you would like to create this directory if it does not already exist:

class MyExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        $myCacheDir = $container->getParameter('my_cache_dir');

        ...
    }
}

When the load() method gets called, the my_cache_dir parameter is still literally “%kernel.cache_dir%/my_cache”. Luckily you can use the ParameterBag::resolveValue() method to replace all placeholders with their current values:

$myCacheDir = $container->getParameter('my_cache_dir');

$myCacheDir = $container->getParameterBag()->resolveValue($myCacheDir);

// now you can create the cache directory
mkdir($myCacheDir);

7.3 Define parameters in a container extension

Many times you will find yourself in the following situation:

  • You want the developer to provide some value for your bundle’s configuration in config.yml.
  • You then need to use this specific value as an argument for one of the services of your bundle.

Say you have a BrowserBundle and you want the developer to provide a timeout value for the browser service:

browser:
    timeout: 30

Your bundle’s Configuration class should look like:

class Configuration implements ConfigurationInterface
{
    public function getConfigTreeBuilder()
    {
        $treeBuilder = new TreeBuilder();
        $rootNode = $treeBuilder->root('browser');

        $rootNode
            ->children()
                ->scalarNode('timeout')
                ->end()
            ->end()
        ;

        return $treeBuilder;
}

Then inside your container extension you need to process the configuration values from config.yml and the likes:

class BrowserExtension extends Extension
{
    public function load(array $configs, ContainerBuilder $container)
    {
        // load the service definitions
        $fileLocator = new FileLocator(__DIR__.'/../Resources/config');
        $loader = new XmlFileLoader($container, $fileLocator);
        $loader->load('services.xml');

        // process the configuration
        $processedConfig = $this->processConfiguration(
            new Configuration(),
            $configs
        );
    }
}

Your browser service is defined in services.xml:

<service id="browser" class="...">
  <argument>%browser.timeout%</argument>
</service>

The value for timeout that the user has provided in config.yml (“30”) will be available in the container extension’s load() method as $processedConfig['timeout'] so the only thing you need to do in that method is copy this value as a parameter to the service container:

$container->setParameter('browser.timeout', $processedConfig['timeout']);

7.4 Override parameters with a compiler pass

Sometimes you want to analyze and override a parameter defined in another bundle, after this bundle has had the chance to define it. For example you may want to modify the hierarchy of user roles defined in security.yml, which will be available as the container parameter security.role_hierarchy.roles. This is the standard hierarchy:

array (
  'ROLE_ADMIN' =>
    array (
      0 => 'ROLE_USER',
    ),
  'ROLE_SUPER_ADMIN' =>
    array (
      0 => 'ROLE_USER',
      1 => 'ROLE_ADMIN',
      2 => 'ROLE_ALLOWED_TO_SWITCH',
    ),
)

Say you have another mechanism for determining a role hierarchy (maybe you retrieve them from some other configuration file), you can then modify or replace the role hierarchy entirely by creating a dedicated compiler pass:

class EnhanceRoleHierarchyPass implements CompilerPassInterface
{
    public function process(ContainerBuilder $container)
    {
        $parameterName = 'security.role_hierarchy.roles';

        $roleHierarchy = $container->getParameter($parameterName);

        // modify the role hierarchy
        ...

        $container->setParameter($parameterName, $roleHierarchy);
    }
}

Don’t forget to register this compiler pass in your bundle class:

class YourBundle extends Bundle
{
    public function build(ContainerBuilder $container)
    {
        $container->addCompilerPass(new EnhanceRoleHierarchyPass());
    }
}

III Modelling your domain with Doctrine

8. Introduction

When you start working with Symfony, you probably learn how everything works by reading “The Book”. If you follow the advice given there, you’ll quickly see some useful results. You should be able to render a page with a form, bind it to a Doctrine entity and persist it in your database. Parts of your code can be auto-generated, and other parts will simply work because the Symfony Form component and Doctrine ORM are designed to work well together. This is why frameworks like Symfony are said to support “Rapid Application Development”, or RAD. Even though RAD used to mean something quite different (it’s more about project methodology), it is how many people call this style of software development, so I’ll keep using the word “RAD” in this chapter.

It turns out that RAD is very useful, but mainly for prototypes only. Prototypes should be considered quick sketches, made with the least amount of work. They are not meant to survive longer than the first stage of the project or even end up on a production server. In my experience prototypes do end up on production servers more often than not, so be very careful about going this way.

Even though it’s great that Symfony (and Doctrine) integrate so seamlessly and allow you to quickly set up a working project, there are many disadvantages to it. It’s deceptively easy to generate entities, bundles, controllers, etc. But the reality of software development always catches up with us. The world gives us complicated business problems as well as continuous change. A lot of the code that we Symfony developers produce (or use) daily won’t help us and in fact, will work against us.

We’ll discuss several of these problems and provide some solutions too. This chapter will focus primarily on creating a proper domain model in the core of your Symfony application. As the standard ORM for Symfony we will of course use Doctrine, but we won’t focus too much on it, since we are basically trying to create the domain model we want, not the domain model that Doctrine (or Symfony) needs. There we go:

8.1 Problem 1: An anemic domain model

If we simply use app/console doctrine:generate:entity to generate our entities, we end up with an anemic domain model. Our entities will just be simple data holders, with only getters and setters. They don’t show any behavior by themselves, all the knowledge about the business is elsewhere.

This will become especially problematic when different parts of the application start working with the same entities. If they don’t have any useful behavior themselves, we will end up taking a lot of information from them (by calling those getters) and figuring it out ourselves. Besides being a violation of the Tell, don’t ask principle, this means that business logic will be scattered all across the code base and that changing aspects of those logic becomes harder over time.

Consider the following code sample:

class BankAccount
{
    private $currency;
    private $amount;

    public function setCurrency($currency)
    {
        $this->currency = $currency;
    }

    public function setAmount($amount)
    {
        $this->amount = $amount;
    }

    public function getCurrency()
    {
        return $this->currency;
    }

    public function getAmount()
    {
        return $this->amount;
    }
}

$bankAccount = new BankAccount();
// do we have a valid bank account now?

$bankAccount->setCurrency('USD');
// and now?

$bankAccount->setAmount(1000);
// now we do!

$bankAccount->setCurrency('EUR');
// should we be able to do this?

// at least we're richer now!

Of course, this is a very naive piece of code. When “setting” a new currency we could of course apply conversion rates to the amount:

$newCurrency = 'EUR';

$conversionRate = conversionRate(
    $bankAccount->getCurrency(),
    $newCurrency
);

$newAmount = $bankAccount->getAmount() * $conversionRate;
$bankAccount->setAmount($newAmount);
$bankAccount->setCurrency($newCurrency);

This is a nice example of business logic that exists outside of the BankAccount class. Tell, don’t ask is being violated because we first get some information from BankAccount, then do some calculations and put new information back in.

The example also demonstrates some even more dangerous aspect of entities with getters and setters: they allow you to change everything inside it in any possible way. We can change the amount without changing the currency, or change the currency without changing the amount. We can even change both (although not at the exact same time), and end up with quite a mess.

Basically, BankAccount’s internals are completely exposed to the world, so we might just as well have defined the class like this:

class BankAccount
{
    public $currency;
    public $amount;
}

Which points us to some other problem with anemic domain model classes: BankAccount doesn’t validate the given currency or amount in any way. As far as BankAccount is concerned, we could also provide some invalid data, like so:

$bankAccount = new BankAccount();
$bankAccount->currency = 'AMSTERDAM';
$bankAccount->amount = new \DateTime();

8.2 Problem 2: Everything is an entity

If we follow the RAD development cycle and let our entities be “generated” for every “thing” that we find in our domain, we are likely to end up with entities which don’t deserve to be called an “Entity”, because they are in fact a Value object. Value objects are used to describe aspects of an entity. For these objects an identity (like a “primary ID”) is irrelevant; they are compared based on the equality of their essential properties. Also, value objects don’t undergo changes, like entities do: they are simply being replaced with new value objects.

With value objects we can prevent an entity to be described in an invalid or inconsistent way. In the previous BankAccount code sample you could still replace currency and amount separately from one another, but also provide invalid data for it. If we want to model currency conversion in a more correct way, we should make sure that currency and amount can only change together. Since we’re talking about how much money a bank account contains, we could pick the name Money to represent “currency and amount”:

class Money
{
    private $currency;
    private $amount;

    public function getCurrency()
    {
        return $this->currency;
    }

    public function getAmount()
    {
        return $this->amount;
    }
}

Validation upon construction

With anemic domain models we’re used to working with objects that are in an invalid state: they can be constructed with no data at all, and they can be changed in any way (including unforeseen ways). Introducing value objects is a first step to move away from that situation. A value object can only be created using valid and consistent data:

class Money
{
    private $currency;
    private $amount;

    public function __construct($currency, $amount)
    {
        if (!is_string($currency)) {
            throw new \InvalidArgumentException(
                'Currency should be a string'
            );
        }
        if (!in_array($currency, ['USD', 'EUR', ...])) {
            throw new \InvalidArgumentException(
                'Unknown currency'
            );
        }
        $this->currency = $currency;

        if (!is_int($amount)) {
            throw new \InvalidArgumentException(
                'Amount should be an integer'
            );
        }

        if ($amount < 0) {
            throw new \InvalidArgumentException(
                'Amount should be positive'
            );
        }

        $this->amount = $amount;
    }
}

This looks like a lot of code, but it’s the only way we can guarantee a correct instantiation of a Money object. Now whenever we encounter a Money object in the wild, we’ll know that it’s created in a valid way (unless we cheat using for example reflection of course).

Improving the entity

Once we have value objects like Money which guard their own consistency, we can improve our entities as well. As we saw before, a BankAccount can be constructed without any mandatory data. Of course, a BankAccount without a currency and an amount should not be able to exist, since we couldn’t reason about it nor work with it in predictable ways. So let’s force it to be instantiated with a Money instance provided to it:

class BankAccount
{
    private $money;

    public function __construct(Money $money)
    {
        $this->money = $money;
    }
}

Almost all of the problems about anemic domain models we discussed above have vanished. Now the only thing that’s left to do is to move the “conversion” behavior into BankAccount. A first step might be this:

class BankAccount
{
    private $money;

    public function __construct(Money $money)
    {
        $this->money = $money;
    }

    public function convertToCurrency($currency)
    {
        $conversionRate = conversionRate(
            $this->money->getCurrency(),
            $newCurrency
        );

        $newAmount = $this->money->getAmount() * $conversionRate;

        /*
         * We can't (and shouldn't be able to!) modify the
         * existing Money instance, so we replace it:
         */
        $this->money = new Money(
            $newCurrency,
            $newAmount
        );
    }
}

// usage example:
$bankAccount = new BankAccount(new Money('USD', 1000));
$bankAccount->convert('EUR');

Converting to a different currency is now an atomic operation. Objects will never be in an invalid state. One remark though: since the conversionRate() function is probably an interchangeable service, we shouldn’t call it from within the entity, instead we should provide it when calling the method:

class BankAccount
{
    ...

    public function convertToCurrency(
        $currency,
        CurrencyConverter $currencyConverter
    ) {
        $conversionRate = $currencyConverter
            ->conversionRate(
                $this->money->getCurrency(),
                $newCurrency
            );

        ...
    }
}

Providing a collaborating service as a method argument like this is called a double dispatch. Since CurrencyConverter is an interface, we’ve successfully applied the Dependency inversion principle here: BankAccount is not tied to any particular currency converter.

There are many other ways in which introducing value objects in your domain model can be helpful. See also the suggested reading material at the end of this chapter.

Usually when we define entities and their relations, we’ll end up relating every entity in some way to another one, giving us one very large interconnected model. For example User entity is often forced to represent multiple domain-related things, like the Author of a Book, the Sender of a Message, etc.

You don’t want a User which can return all the Books of which it was the author, all the Messages of which it was the sender, etc. First of all, this blurs the lines of what should be considered your transactional boundary. If you allow Doctrine to travel all the way across one big object graph on flush, all changes to any number of entities might be persisted in one transaction. In general it’s good to isolate areas of change. You can ask yourself: when a book changes, does its author (usually) change too? If not, you should not allow them to be persisted within the same transaction (there’s much more to say about this - you should read Effective Aggregate Design by Vaughn Vernon).

Linking objects using just their ids

One very important yet simple trick to break up a one-size-fits-all domain model is to link domain objects to each other using only their identifiers instead of making the entire object graph traversable. Then, instead of traversing over the object graph, you can just look up all BlogPosts and Comments that were produced by a given user.

Initially, your domain model might look like this:

namespace Domain;

class User
{
    private $id;
    private $blogPosts;
    private $comments;
}

class BlogPost
{
    private $user;
    ...
}

class Comment
{
    private $user;
    ...
}

Instead of allowing User, BlogPosts and Comments to be modified in the same transaction, we can separate the two by just keeping the ID of the user in BlogPost and Comment, and remove any reference to actual BlogPost or Comment instances in User:

namespace Domain;

class User
{
    private $id;
}

class BlogPost
{
    private $userId;
    ...
}

class Comment
{
    private $userId;
    ...
}

It isn’t even that big a change: User is now somewhat of an “upstream” entity for BlogPost and Message. However, they no longer have a class-level dependency on User. They just need to keep a user’s identifier, to allow clients to look up which user was the creator of a blog post or a message. Please note that userId doesn’t need to be a foreign key in the database anymore. There won’t be any cascading behavior or foreign key constraints in your MySQL database as well (we don’t want to make our database responsible for domain-level decisions).

Splitting up entity roles

Besides blurry transaction boundaries, another problem with a completely traversable domain model is that some entities will tend to play too many roles at the same time. The modelled concepts will become too vague and the classes themselves too big. Each of those roles requires different state (properties) and behavior (methods), giving the class too many responsibilities.

When we allow one domain class to play multiple roles, we ignore the fact that some concepts within our domain model simply don’t represent the same thing in every context. We’d be better off explicitly marking (bounded) contexts in our application.

We start off with a User, associated with Books and Messages:

namespace Domain;

class User
{
    private $id;

    /**
     * @var string
     */
    private $name; // only used as the name of a book author

    /**
     * @var string
     */
    private $email; // only used as the address of a message sender

    /**
     * @var Book[]
     */
    private $books;

    /**
     * @var Messages[]
     */
    private $messages;
}

class Book
{
    /**
     * @var User
     */
    private $user;
    ...
}

class Message
{
    /**
     * @var User
     */
    private $user;
    ...
}

These classes are not so well designed:

  • User contains data (and possibly logic) related to the role of “author of a book” as well as the role of “sender of a message”.
  • Book and Message are both coupled to the entire User class. This class is completely irrelevant to Book and Message themselves.
  • Book and Message don’t have a notion of “author” or “sender”. They only mention their relation to a “user”, which is generic and not specific for the problem domain.

We can simply split User (or possibly even better: get rid of it), by moving “author”- and “sender”-related logic to their own classes:

class Author
{
    /**
     * @var string
     */
    private $name;
}

class Book
{
    /**
     * @var Author
     */
    private $author;

    ...
}

class Sender
{
    /**
     * @var string
     */
    private $email;
}

class Message
{
    /**
     * @var Sender
     */
    private $sender;
}

Finally the domain concepts “author” and “sender” are properly represented in the domain model. Also, there’s no User anymore: all author- or sender-specific data and/or behavior has been moved to Author and Sender respectively.

Now you probably need a way to find out whether the author of a book is the same “person” as the sender of a message. This is where it makes sense to keep a unique identifier for that “person” inside Sender and Message:

class Sender
{
    /**
     * @var string
     */
    private $id;
    ...
}

It’s important to note that this identifier won’t be (auto-)assigned by the database. It’s not used as a primary key. It’s just an identifier (we’ll discuss identifiers in depth in the next session).

Once you have an identifier for Sender and Author, you might as well just apply one of the previous strategies for associating objects: just use that identifier. The end result will therefore look something like this:

class Author
{
    /**
     * @var string
     */
    private $id;

    /**
     * @var string
     */
    private $name;
}

class Book
{
    /**
     * @var string
     */
    private $authorId;

    ...
}

class Sender
{
    /**
     * @var string
     */
    private $id;

    /**
     * @var string
     */
    private $email;
}

class Message
{
    /**
     * @var string
     */
    private $senderId;
}

Of course it’s likely that the actual person that is using your web application is represented as an entity somewhere in your domain model. Maybe it’s a class like SecurityIdentity below:

class SecurityIdentity
{
    private $id;
    private $username;
    private $passwordHash;

    public function __construct($id, $username, $passwordHash)
    {
        // validate and assign the given values
        ...
    }

    public function id()
    {
        return $this->id;
    }

    public function username()
    {
        return $this->username;
    }

    public function passwordHash()
    {
        return $this->passwordHash();
    }
}

If you make sure that the ID of a SecurityIdentity is the same for corresponding Author and Sender objects, then you can get an actual Author based on the authenticated user by simply taking it from its repository:

interface AuthorRepository
{
    /**
     * @param string $id
     * @throws AuthorNotFound
     * @return Author
     */
    public function authorWithId($id);
}

Or, if Author is a value object (i.e. it doesn’t undergo changes, but simply describes part of an entity), just create it on the fly, using a factory:

class AuthorFactory
{
    public function fromSecurityIdentity(SecurityIdentity $securityIdentity)
    {
        return new Author($securityIdentity->id());
    }
}

// example usage:

// $securityIdentity represents the currently authenticated user
$securityIdentity = ...;
$authorFactory = new AuthorFactory();

$author = $authorFactory->fromSecurityIdentity($securityIdentity);

// use $author in the way your domain model demands it, e.g.:
$book = Book::writtenBy($author);

8.4 Problem: entities are coupled to the framework

By now you will probably get the idea: so far I’ve not even mentioned Doctrine mapping annotations at all; we (almost) didn’t discuss the actual persistence of our domain objects using Doctrine. The approach I’m pushing here is called “domain-first”. I think you’re better off working on a proper domain model, and only afterwards think about storing domain objects in a database.

You should pay a lot of attention to a proper domain model, which is not forced in any direction by some technical/infrastructural choice you or someone else has made already. When you’ve spent a good amount of time defining your classes in such a way that no invalid objects can be created, it would be a shame to give up on their quality by allowing Doctrine or, for example the Symfony Security Component to force itself onto the public interfaces of your nicely defined domain model classes.

The omnipresent User class I previously mentioned is a good example of how your domain model deserves something better than to be taken over by requirements made by the Symfony Security Component. A little note before we continue: I think it’s a great component, so you should definitely embrace it. Just keep it outside of your precious domain code.

Address framework concerns in their own classes

The default way of using the Symfony Security Component with a Doctrine User entity is to use the entity user provider and load users from the database using Doctrine. The model-first approach doesn’t speak of databases, so we don’t want to integrate on that level. Given the SecurityIdentity class above, we just need to make sure that Symfony knows how to deal with it. Of course Symfony offers some very easy ways to solve this. The only thing we need to do is to introduce our own security user class, which implements UserInterface and then provide a custom user provider class, which implements UserProviderInterface.

First we define our own SymfonyUser class which can be used to represent the currently authenticated user. To satisfy Symfony, we have to implement UserInterface and EquatableInterface:

namespace Infrastructure\Web\Security;

use Symfony\Component\Security\Core\User\UserInterface;

class SymfonyUser implements UserInterface, EquatableInterface
{
    private $id;
    private $username;
    private $password;

    public function __construct($id, $username, $password)
    {
        $this->id = $id;
        $this->username = $username;
        $this->password = $password;
    }

    public function id()
    {
        return $this->id;
    }

    public function getUsername()
    {
        return $this->username;
    }

    public function getPassword()
    {
        return $this->password;
    }

    public function getSalt()
    {
        // return the password salt if applicable

        return null;
    }

    public function getRoles()
    {
        // determine the applicable roles in any way you like, e.g.

        return ['ROLE_USER'];
    }

    public function eraseCredentials()
    {
        /*
         * This method only needs to be implemented if
         * your password is stored in plain text, and
         * will be called before serializing this object
         * to be persisted in the user's session.
         */
    }
}

This SymfonyUser class is a simple data holder which can be used by Symfony to retrieve some information about the currently authenticated user. SymfonyUser only exposes the information that Symfony would need for that purpose. It doesn’t carry other information with it, and the public interface of the SecurityIdentity or any of the other role-specific user classes isn’t polluted by methods like eraseCredentials() or `getRoles().

Now we only need to define a user provider which produces instances of the new SymfonyUser class. This user provider should take existing SecurityIdentity entities (see on of the previous example for what it looks like) from their repository and turn them into Symfony-targeted SymfonyUser objects. Let’s first define an interface for SecurityIdentityRepository:

namespace Identity\Domain\Model;

interface SecurityIdentityRepository
{
    /**
     * @param string $username
     * @return SecurityIdentity
     * @throws UnknownSecurityIdentity
     */
    public function securityIdentityWithUsername($username);
}

class UnknownSecurityIdentity extends \RuntimeException
{
    ...
}

We can provide any implementation for this repository, but it doesn’t really matter where these SecurityIdentity objects come from (more about this later). Now that we’ve defined the public interface of SecurityIdentityRepository, we can start using it in the UserProvider class. This class implements Symfony’s UserProviderInterface, thereby allowing you to use it as a user provider which can replace the traditional entity user provider:

namespace Infrastructure\Web\Security;

// TODO check namespace for model stuff
use Identity\Domain\Model\SecurityIdentityRepository;
use Identity\Domain\Model\UnknownSecurityIdentity;
use Symfony\Component\Security\Core\Exception\UnsupportedUserException;
use Symfony\Component\Security\Core\Exception\UsernameNotFoundException;
use Symfony\Component\Security\Core\User\UserInterface;
use Symfony\Component\Security\Core\User\UserProviderInterface;

class SymfonyUserProvider implements UserProviderInterface
{
    private $securityIdentityRepository;

    public function __construct(
        SecurityIdentityRepository $securityIdentityRepository
    ) {
        $this->securityIdentityRepository = $securityIdentityRepository;
    }

    public function loadUserByUsername($username)
    {
        try {
            $securityIdentity = $this->securityIdentityRepository
            ->securityIdentityWithUsername($username);

            /*
             * Simply construct a new SymfonyUser object that holds
             * the relevant data:
             */
            return new SymfonyUser(
                $securityIdentity->id(),
                $securityIdentity->username(),
                $securityIdentity->passwordHash()
            );
        } catch(UnknownSecurityIdentity $previous) {
            /*
             * Convert the domain-specific exception to something
             * Symfony understands:
             */
            throw new UsernameNotFoundException(
                sprintf(
                    'Unknown username "%s"',
                    $username
                ),
                null,
                $previous
            );
        }
    }

    public function refreshUser(UserInterface $user)
    {
        if (!$user instanceof SymfonyUser) {
            throw new UnsupportedUserException();
        }

        return $this->loadUserByUsername($user->getUsername());
    }

    public function supportsClass($class)
    {
        return $class === SymfonyUser::class;
    }
}

Please take a look at the Symfony documentation on implementing a custom user provider on how to hook this up in the service container.

This piece of code is a nice example of the level of decoupling that can be achieved by implementing the Adapter pattern. In this case, we already had something that resembled a “user” entity, but we needed its interface to match the one expected by Symfony. Instead of polluting the public interface of SecurityIdentity, we introduced an extra class (SymfonyUser), which itself knows nothing about SecurityIdentity entities, but is just a simple data holder.

In the same way, SymfonyUserProvider basically wraps the existing SecurityIdentityRepository (a design strategy called composition), thereby allowing Symfony to use it to retrieve user objects it can work with.

8.5 Conclusion

In this chapter we’ve looked at some of the most important ways in which you can shape up your domain model. We still need to cover some other big improvements, but we’ll look at them in the context of the overal application structure (in the next chapter that is).

Usually letting Doctrine generate your entities leads to an anemic domain model, i.e. entities with just getters and setters, (almost) no domain logic and (almost) no proper validation. We can move away from such a poor model by moving related behavior out of services and into the entities themselves. To solve the problem of validation and prevent entities to be filled with invalid data in the first place, we pointed out how you can introduce value objects in your domain model. Value objects describe your entity, don’t need an identity, are immutable and guard their own consistency by use of assertions.

IV Application architecture

8.6 Problem: entities used on both the write and read side

8.7 Problem: Coupling to the delivery mechanism

  • All the central use cases coupled to the web as a delivery mechanism (e.g. you can only use an HTML page with a form to perform a certain task).

8.8 Problem: Slow tests

  • A slow test suite, since we can’t test the use cases of our application without setting up and populating an entire database.

9. What is a proper architecture?

Following the Symfony documentation - CRUD Characteristics of CRUD-y code: create, update, delete controllers, forms which map data from and to an entire entity including its related objects, an entity which is fully mapped to one table (and related tables). Controllers which just copy data to and from objects, forms and entity managers. RAD. A “Symfony” project. Piece of example code

MVC -> view updates based on changes in the model (observes changes in the model), controller loads and stores data from the model, view calls controllers. Entirely the wrong thing for web applications where flow is linear: request to response. MVC is useful in client-facing applications with a longer life-cycle. Layers and boundaries: layers each have their own area of expertise, and their own level of abstraction. Domain layer, application, infrastructure layer. Layers allow you to allocate things. Layers allow you to define rules for dependencies. The dependency rule: depend only on things from the same layer or deeper layers.

Looking at a project: It shouldn’t be the framework that determines the layout of your code https://www.youtube.com/watch?v=WpkDN78P884 Screaming architecture https://blog.8thlight.com/uncle-bob/2011/09/30/Screaming-Architecture.html Intention revealing interfaces

Frameworks solve infrastructural issues very well We don’t want coupling to the delivery mechanism. Use cases should be defined before HTML forms. Should be able to be executed before there even is a web page for the particular use case. This even removes a dependency between people. Any way in which an application has connections with the real world should be interchangeable. Flexibility, allows the best solution for the given problem (not one-size-fits-all MySQL database for example)

Hexagonal architecture Application as object: encapsulation, methods as input ports Define the ports Write adapters for them Adapters are interchangeable The database now is an implementation detail Switching it out will allow you to write faster tests, which don’t need an actual database, populated with some fixture data. Separation between core and infrastructure allows you to put a focus on quality where it’s most needed. In the infrastructure layer you can fully embrace the frameworks and libraries that you use, including their conventions.

9.1 What does it mean for my code?

In general: Directory structure, project layout Domain, Application, Infrastructure Dependency inversion (interfaces, abstract concepts) Statelessness: service vs entity

For Symfony: Bundles no longer the central unit. Bundles are framework specific. Bundles mix many concerns. According to the docs: put everything in AppBundle. No fat controllers, but before pushing to the next (application service layer): remove low-level details from the current layer. So form handlers have no use anymore Routes and console commands: determine the input ports.

Use forms to populate a command object Use validator to validate command Show errors to user in a form Use SimpleBus for handling commands Then: let domain layer handle any inconsistencies that might occur composer require beberlei/assert

Console commands create command objects too and handles them just as if they were provided by an HTTP request: proof that you’ve removed delivery details successfully.

Controllers, console commands: where to put them? Infrastructure/Web Infrastructure/Cli Controllers as service Console command as service Embrace the framework!

Request service: don’t depend on it (coupling to delivery mechanism: the web) The security context: don’t depend on it (coupling to the PHP session which is only available for web requests) Current user (from security context): same, don’t fetch it, provide it. Request stack service: don’t depend on it, instead, pass the service to any other service that might be interested as an argument. For context-specific things: use a param converter. Controller is the entry point for your application, before that everything is handled by the framework. The rest is handled by dependency injection and as we’ve learnt we shouldn’t consider the request to be a proper service. Example code: the current user as a controller argument. Example/blog post? Decoupling your domain user from the current user Rule: think about the same scenario, but without an HTTP request, cookies and a server-side session! (Infrastructure) event listeners to respond to specific details about the web request or console command.

use Sensio\Bundle\FrameworkExtraBundle\Configuration\ParamConverter;
use Sensio\Bundle\FrameworkExtraBundle\Request\ParamConverter\ParamConverterInte\
rface;
use Symfony\Component\DependencyInjection\ContainerInterface;
use Symfony\Component\HttpFoundation\Request;
use Symfony\Component\Security\Core\User\UserInterface;

class InjectCurrentUserIntoController implements ParamConverterInterface
{
    /**
     * @var ContainerInterface
     */
    private $container;

    public function __construct(ContainerInterface $container)
    {
        $this->container = $container;
    }

    public function apply(Request $request, ParamConverter $configuration)
    {
        // use 'security.context' for Symfony < 2.6
        $currentUser = $this->container
            ->get('security.token_storage')
            ->getToken()
            ->getUser();

        $request->attributes->set(
            $configuration->getName(),
            $currentUser
        );
    }

    public function supports(ParamConverter $configuration)
    {
        return $configuration->getClass() === UserInterface::class;
    }
}
services:
    inject_current_user_into_controller:
        class: AppBundle\InjectCurrentUserIntoController
        arguments:
            - @service_container
        tags:
            - { name: request.param_converter, converter: current_user }
/**
 * @Route("/admin")
 */
public function adminAction(Request $request, UserInterface $user)
{
    return $this->render('default/index.html.twig', array(
        'user' => $user
    ));
}

Persistence

Move files to DomainModel directory, including annotations. Possibly divide directory into what belongs to certain aggregates.

doctrine:
    ...

    orm:
        ...
        mappings:
            domain:
                type: annotation
                dir: %kernel.root_dir%/../src/Domain
                prefix: Domain

Model-first, not field/table first! Model classes should hide Doctrine-specific stuff (public interface is persistence-mechanism independent). Under the hood you’ll comply to Doctrine’s standards anyway, so you might just as well use annotations for configuration. Besides, that will keep the configuration close the code that it’s about.

Hexagonal: output port

namespace Domain;

interface TaskRepository
{
    public function add(Task $task);

    public function taskWithId($id);
}
use Domain\Task;
use Doctrine\Common\Persistence\ManagerRegistry;
use Domain\TaskNotFound;
use Domain\TaskRepository;

class DoctrineORMTaskRepository implements TaskRepository
{
    /**
     * @var ManagerRegistry
     */
    private $managerRegistry;

    public function __construct(ManagerRegistry $managerRegistry)
    {
        $this->managerRegistry = $managerRegistry;
    }

    public function add(Task $task)
    {
        $entityManager = $this->managerRegistry->getManagerForClass(Task::class);
        $entityManager->persist($task);
        $entityManager->flush();
    }

    public function taskWithId($id)
    {
        $entityManager = $this->managerRegistry->getManagerForClass(Task::class);
        $task = $entityManager->find(Task::class, $id);
        if (!$task instanceof Task) {
            throw new TaskNotFound($id);
        }

        return $task;
    }
}
namespace Domain;

class TaskNotFound extends \RuntimeException
{
}

// not DomainException!
     doctrine_orm_task_repository:
        class: Infrastructure\Persistence\DoctrineORMTaskRepository
        arguments:
            - @doctrine

    task_repository:
        alias: doctrine_orm_task_repository

Instead of DoctrineFixtures and Sqlite etc., use an abstract repository Application service testing: replace the abstract repository Example YAML service definition Functional testing: test the concrete repository implementation with an actual database (don’t test Doctrine though)

# only load these in the test environment
services:
    in_memory_task_repository:
        class: Infrastructure\Persistence\InMemoryTaskRepository

    task_repository:
        alias: in_memory_task_repository

Fixtures: in your test case, just add a number of objects to the repository. You don’t need to write all the equivalent code (like complex queries, etc.), since usually it only matters if any object of the given type will be returned.

Coupling to auto-generated IDs: use UUIDs (e.g. e827d662-3f84-4ac3-9564-31db6cefe1de) to determine ids yourself. Also easier with asynchronicity.

composer require ramsey/uuid
composer require ramsey/uuid-doctrine
use Doctrine\ORM\Mapping as ORM;
use Ramsey\Uuid\Uuid;

/**
 * @ORM\Entity()
 */
class Task
{
    /**
     * @ORM\Id()
     * @ORM\Column(type="uuid")
     * @ORM\GeneratedValue(strategy="NONE")
     */
    private $id;

    /**
     * @param UuidInterface $uuid
     */
    public function __construct(UuidInterface $uuid)
    {
        $this->id = $uuid;
    }

    /**
     * @return Uuid
     */
    public function id()
    {
        return $this->id;
    }
}
doctrine:
    dbal:
        ...
        types:
            uuid: Ramsey\Uuid\Doctrine\UuidType

Don’t use Doctrine events to respond to change: coupling to delivery mechanism Instead: use domain events, respond to them.

class Task
{
    ...

    private $events = [];

    /**
     * @param UuidInterface $uuid
     */
    public function __construct(UuidInterface $uuid)
    {
        $this->id = $uuid;

        $this->events[] = new TaskCreated($this);
    }

    ...

    public function events()
    {
        return $this->events;
    }
}
$task = new Task(Uuid::uuid4());

$this->get('task_repository')->add($task);

foreach ($task->events() as $domainEvent) {
    $this->get('event_dispatcher')->dispatch(
        DomainEvent::EVENT_NAME,
        new DomainEvent($domainEvent)
    );
}
use Domain\TaskCreated;
use Psr\Log\LoggerInterface;
use Symfony\Component\EventDispatcher\EventSubscriberInterface;

class LogTaskCreatedEvent implements EventSubscriberInterface
{
    /**
     * @var LoggerInterface
     */
    private $logger;

    public function __construct(LoggerInterface $logger)
    {
        $this->logger = $logger;
    }

    public static function getSubscribedEvents()
    {
        return [
            DomainEvent::EVENT_NAME => 'handleDomainEvent'
        ];
    }

    public function handleDomainEvent(DomainEvent $event)
    {
        $taskCreated = $event->event();
        if (!$taskCreated instanceof TaskCreated) {
            return;
        }

        $this->logger->debug(
            'Task created',
            [
                'id' => (string) $taskCreated->task()->id()
            ]
        );
    }
}
    log_task_created_event:
        class: Infrastructure\EventDispatching\LogTaskCreatedEvent
        arguments:
            - @logger
        tags:
            - { name: kernel.event_subscriber }
DEBUG - Task created
Context: {"id":"52d3674c-20f5-4489-ac6b-d7dfa5c26297"}

Again: SimpleBus has something for this

View models Queries

class Task
{
    ...

    /**
     * @ORM\Column(type="boolean")
     */
    private $finished = false;

    /**
     * @ORM\OneToMany(targetEntity="Collaborator")
     */
    private $collaborators;

    ...

    public function finished()
    {
        $this->finished = true;
    }

    /**
     * @return Collaborator[]
     */
    public function collaborators()
    {
        return $this->collaborators;
    }
}
<h1>Task {{ task.id }}</h1>

{% if task.finished %}
    <p>This task is finished</p>
{% endif %}

<h2>Collaborators</h2>
<ul>
{% for task.collaborators as collaborator %}
    <li>{{ collaborator.name }}</li>
{% endfor %}
</ul>

Entities in templates? Dangerous: fetching relations, not optimised for querying, bad for security reasons, risk for change (entities have public methods which allow for changes), objects instead of raw values Separate view models (simple DTOs) - create based on actual entities. Use ViewModelRepositories and/or ViewModelFactories - PseudoCQRS

use Doctrine\Common\Persistence\ManagerRegistry;
use Domain\TaskNotFound;
use Domain\TaskRepository;

class DoctrineORMTaskRepository implements TaskRepository
{
    /**
     * @var ManagerRegistry
     */
    private $managerRegistry;

    public function __construct(ManagerRegistry $managerRegistry)
    {
        $this->managerRegistry = $managerRegistry;
    }

    public function add(Task $task)
    {
        $entityManager = $this->managerRegistry
            ->getManagerForClass(Task::class);
        $entityManager->persist($task);
        $entityManager->flush();
    }

    public function taskWithId($id)
    {
        $task = $this->managerRegistry
            ->getRepository(Task::class)
            ->find($id);

        if (!$task instanceof Task) {
            throw new TaskNotFound($id);
        }

        return $task;
    }

    public function all()
    {
        return $this->managerRegistry
            ->getRepository(Task::class)
            ->findAll();
    }
}
namespace Infrastructure\ViewModel;

use Domain\Task;
use Domain\TaskRepository;
use ViewModel\TaskSummaryRepository;

class TaskSummaryFromTaskWriteModelRepository implements TaskSummaryRepository
{
    /**
     * @var TaskRepository
     */
    private $taskRepository;

    public function __construct(TaskRepository $taskRepository)
    {
        $this->taskRepository = $taskRepository;
    }

    public function all()
    {
        foreach ($this->taskRepository->all() as $task) {
            yield Task::createTaskSummaryViewModelFrom($task);
        }
    }
}
namespace ViewModel;

class TaskSummary
{
    public $id;
    public $finished;
}
class Task
{
    ...

    /**
     * @param Task $task
     * @return TaskSummary
     */
    public static function createTaskSummaryViewModelFrom(Task $task)
    {
        $taskSummary = new TaskSummary();
        $taskSummary->id = (string) $task->id;
        $taskSummary->finished = $task->finished;

        return $taskSummary;
    }
}

This is the first step towards CQRS with separate read and write models and projections.

V Configuration conventions

10. Application configuration setup

The Symfony Standard Edition recommends the use of a parameters.yml file to contain configuration values that are specific for an instance of the project, be it on a production server or on the developer’s computer. The configuration values defined in config.yml contain placeholders, pointing to parameters in parameters.yml, which are loaded in the imports section:

imports:
    - { resource: parameters.yml }

doctrine:
    dbal:
        user:     %database_user%
        password: %database_password%

In parameters.yml you should then define:

parameters:
    database_user:     matthias
    database_password: cookies

Or you will get a nasty exception when you run the application:

You have requested a non-existent parameter "database_user".

This set-up soon gets impractical. The more dependencies you have in your project, the more configuration values will differ between you and your colleagues (let alone between your machine and the production server). So the list in parameters.yml gets longer and longer and every time a team member adds another parameter to the list, you will get another ParameterNotFoundException after pulling the changes from the repository.

This method is also very inflexible when you want to change the behavior of some bundles without making these changes permanent. Chances are you will accidentally commit things like:

# in config.yml
swiftmailer:
    # deliver all emails sent by the application to this address:
    delivery_address: matthiasnoback@gmail.com

Of course you can define another parameter for it:

# in config.yml
swiftmailer:
    delivery_address: %developer_email_address%

# in parameters.yml
parameters:
    developer_email_address: matthiasnoback@gmail.com

But your colleagues may not want to change the default behavior, they just want everything to keep working when they update the project.

10.1 Use local configuration files

The best solution for this problem is to use configuration files which only exist locally. For each environment you can create a local_{env}.yml configuration file. You can then load these local configuration files after loading the original config_{env}.yml files. To accomplish this, modify app/AppKernel.php:

public function registerContainerConfiguration(LoaderInterface $loader)
{
    $loader->load(__DIR__.'/config/config_'.$this->getEnvironment().'.yml');

    $localFile = __DIR__.'/config/local_'.$this->getEnvironment().'.yml';

    if (is_file($localFile)) {
        $loader->load($localFile);
    }
}

Because the local configuration file will be loaded after the general configuration file, each developer can override any part of the project configuration. When a local configuration file does not exist, the application won’t fail to load. This means that deploying your application does not require an extra step.

Don’t forget to add the local_{env}.yml files to your .gitignore file. It would also be a good idea to add .dist versions of the local_{env}.yml files, containing some useful suggestions for teammates. Make sure all of these suggestions are commented out:

# local_dev.yml.dist
imports:
    - { resource: config_dev.yml }

#swiftmailer:
#    delivery_address: your-mail-address@host

And also make sure that the local_{env}.yml.dist files will be committed to your repository.

Keep parameters.yml

It would still be a good idea to use parameters for some required configuration values. For example, without a database username and password, your whole application will not work, so it makes sense to ask for these values when installing or updating the project.

If you have a Symfony project with Symfony version less than 2.3, you should install ParameterHandler (create by Incenteev) in your project. From then on, after installing or updating using Composer the ParameterHandler compares the contents of parameters.yml.dist with the contents of your current parameters.yml and asks you to supply the missing values.

Add a default_parameters.yml

Though this configuration setup is becoming pretty flexible already, there are situations where you want to use yet another extra way of configuring things. Consider this MongoDB configuration:

doctrine_mongodb:
    connections:
        default:
            server: mongodb://%mongo_host%:%mongo_port%

You don’t want to copy this entire hierarchy to your local_dev.yml file to get things up and running on your own machine. You want to be able to use the parameters %mongo_host% and %mongo_port%. Yet in the default setup of your application these parameters may not vary much. On most developer machines, the host and the port for MongoDB are the same. For this situation, add a default_parameters.yml, which contains these parameters, so when making a fresh install of the project, developers don’t have to provide these values. Import this file before importing parameters.yml:

# in config.yml
imports:
    - { resource: default_parameters.yml }
    - { resource: parameters.yml }

# ...

In default_parameters.yml you could now add:

# in default_parameters.yml
parameters:
    mongo_host: localhost
    mongo_port: 27017

And in case your personal MongoDB database is on a different host or port, you can override just these values in parameters.yml:

# in parameters.yml
parameters:
    mongo_port: 71072

When you are using the ParameterHandler mentioned above, in combination with both a default_parameters.yml and a parameters.yml, make sure to add these options to composer.json:

"extra": {
    "incenteev-parameters": {
        "file": "app/config/parameters.yml",
        "keep-outdated": true
    }
}

This way, extra parameters in parameters.yml that override values from default_parameters.yml and are not mentioned in parameters.yml.dist will not be automatically removed from parameters.yml.

11. Configuration conventions

Most parts of the application configuration can be loaded from files in many different formats. The application configuration can be provided using plain PHP, Yaml, INI and XML files. The same (except for the INI format) is true for the routing and validation configuration. But these two also add annotations as an option, and so does Doctrine when it comes to entity and document mapping metadata.

You already have the application structure in order (see the previous chapter), now you have to make choices for your team (maybe settle on this topic together) about the other configuration formats you are going to use in your application. Make sure you document these conventions in an easily accessible place. Then reinforce them as part of your code review process.

In the next sections I’ll show you my own conventions for Symfony projects. I used to be very strict about using XML for everything (service definitions, routing, mapping, etc.). I liked, and still like, the strictness of XML. By using the schema files provided by the framework you can now be sure that what you’re typing is correct. However, reading XML files is pretty hard. I found that it’s much easier to quickly scan and then modify Yaml configuration files then it is to modify XML files. This has actually brought me to favor Yaml over XML these days.

As for annotations: only when I write reusable code, I use XML configuration files instead of annotations. This brings the greatest amount of flexibility to the users of my code. When the code is not (meant to be) reusable, I go for annotations, since it keeps the metadata close to the code.

11.1 Routing

Routing as well as template configuration is preferably defined inside the controllers themselves, using annotations:

use Sensio\Bundle\FrameworkExtraBundle\Configuration\Route;
use Sensio\Bundle\FrameworkExtraBundle\Configuration\Method;
use Sensio\Bundle\FrameworkExtraBundle\Configuration\Template;

/**
 * @Route("/account")
 */
class AccountController
{
    /**
     * @Route("/new")
     * @Method({"GET","POST"})
     * @Template
     */
    public function newAction()
    {
        return array();
    }
}

Each bundle should have a Resources/config/routing.yml file, which loads each controller as a resource:

MatthiasAccountBundle_AccountController:
    resource: "@MatthiasAccountBundle/Controller/AccountController.php"
    type: annotation

MatthiasAccountBundle_CredentialsController:
    resource: "@MatthiasAccountBundle/Controller/CredentialsController.php"
    type: annotation

# ...

Though less explicit, you may also choose to load the routing for an entire controller directory at once:

MatthiasAccountBundleControllers:
    resource: "@MatthiasAccountBundle/Controller/"
    type: annotation

The application routing configuration in /app/config/routing.yml should have references to the routing.yml files of the active bundles:

MatthiasAccountBundle:
    resource: "@MatthiasAccountBundle/Resources/config/routing.yml"

Choosing Route Names

Route names often become one big mess. We have account_new, account_index, accounts, account_list, etc. When an action name changes, the old route name keeps popping up. And when another controller also has a route named accounts, it will override the other already existing route, or be overridden itself.

The solution is very simple: follow this pattern (it may sound familiar):

{name of the bundle without "Bundle"}.{name of the controller without
"Controller"}.{name of the action without "Action"}

For example, the AccountController above has an action newAction. The controller is part of the MatthiasAccountBundle, so its route name should be matthias_account.account.new:

/**
 * @Route("/new", name="matthias_account.account.new")
 */

Remember: when you change a route name, make sure you do a project-wide search-and-replace for occurrences of the route name.

11.2 Services

Define your services using Yaml files.

services:
    matthias_account.account_controller:
        class: Matthias\AccountBundle\Controller\AccountController

When choosing ids for services, use only underscores and lowercase characters. Optionally apply namespacing using dots. The service id for the service of class MailManager inside the MatthiasAccountBundle may be matthias_account.mail_manager.

In case your services are divided into several service files, which are all grouped by a common theme (like metadata, form, etc.), you could insert an extra namespace in the service ids. For instance the service id for the form type of class CreateAccountType in the MatthiasAccountBundle, defined in the form.yml service definition file, may get the id matthias_account.form.create_account_type.

11.3 Mapping metadata

The preferred way of configuring your entities or documents is to use annotations:

/**
 * @ORM\Entity
 * @ORM\Table(name="accounts")
 */
class Account
{
    /**
     * @ORM\Id
     * @ORM\Column(type="secure_random_id")
     */
    private $id;
}

When it comes to (potentially) reusable code, you should not use any annotations. Read more about this in Storage-agnostic models.

VI Security

12. Introduction

12.1 Symfony and security

The Symfony Standard Edition comes with the Security Component, the SecurityBundle and (until version 2.3) the SecurityExtraBundle. When configured for your application (using security.yml), you can have the following for free:

  • One or more secured areas in your application, protected by for instance a login form or HTTP authentication.
  • Total freedom in retrieving user data.
  • Several ways for hashing user passwords.
  • A way to log out.
  • Security-related events that are dispatched through the application’s event dispatcher.
  • Authorization of users based on their roles.
  • Configurable session storage handlers (i.e. you can define where session data should be stored).
  • Access control list (ACL) functionality, which you can use to assign very specific rights (like edit, view, delete, etc.) to specific users concerning specific objects.

All the things in the list above are all implemented very well, but the many different concepts (like: firewall, authentication listener, authentication provider, exception handler, entry point, ACL, ACE, security identity, object identity, etc.) can lead to much confusion. This is why, when you are building any application in which there is a difference between authenticated users and guest users, or where there are access rights on the object level, you should really read about Security & Symfony. There are some good sources on this subject:

  • The chapter Security on symfony.com which gives you the general outline of authentication and authorization management in a Symfony application.
  • The documentation for the Security Component (written by myself) describes in more detail all the parts of the component that play a role in securing your application.

The good thing is, there is much to read. Afterwards you should compare Symfony’s feature set with your own wish list or (“should list”). Because, the bad thing is that there are also many things missing. Some examples of what’s missing or things that you have to create/check yourself:

  • Input sanitizing (almost nothing is done automatically for you, so this will be a larger topic later on).
  • Automatic (inactivity/time-based) session invalidation.
  • Monitoring/preventing session hijacking.
  • Preventing brute-force attacks on login forms.
  • Storing/managing user types, roles and groups dynamically.
  • Showing friendly “access denied” pages.
  • Enforcing strong passwords.
  • Preventing sensitive information from being cached by browsers.

To be able to judge what should really be done, and exactly which solutions help best to prevent the “evil guys” from spoiling the fun, you should also read about PHP security (server configuration amongst many things) and (web) application security in general. This is a subject about which many people have widely diverging opinions, so after understanding the main problems, you will also need to be able to judge the advice given by many different people. I would like to point out one very good source to you, namely the book-in-progress Survive The Deep End: PHP Security, by Pádraic Brady.

When it comes to all the extra security-related measures that are not bundled by default with Symfony, I would like to point you to my own website, where I have posted some articles about security enhancements for your application.

12.2 Goals: prevention and confinement

There are two ways in which you can enhance the security of your application: first you can try to prevent a bad thing from happening, second you can make arrangements so that when it happened anyway, things won’t get too far out of hand. The OWASP Secure Coding Practices - Quick Reference Guide guide puts this as follows:

A threat agent interacts with a system, which may have a vulnerability that can be exploited in order to cause an impact.

Your Symfony application is a system which almost certainly has vulnerabilities (unless it just displays “Hello world!”) and these can be used to break through the thresholds you have placed around your (sensitive) data. Though you should do everything you reasonably can to prevent this from happening, you need to think about what could happen in the case of a security breach. After all, this could be a matter of a leaked password, or even a brute-forced entry to the database server. What does the “hacker” get, and how much time will it take him to get more? And exactly which data has been compromised? Should you inform users about the attack? Which users? Can you track this down? And in the case of a credentials database, could the hashed passwords be “unhashed” using brute-force tactics in a reasonable amount of time?

For example, when your application has a comment system which allows users to directly write a comment on a page, a malevolent user may try to add a JavaScript snippet to a comment. When you don’t have output escaping enabled for these comments, the snippet will just end up on the page as if it were application code. This would be a serious vulnerability of your application. A threat agent could inject a piece of JavaScript that examines document.cookie to find out the ID of the user’s session as it is stored in the session cookie. They may then even hijack the user’s session in their own browser.

Minimize impact

The JavaScript injection exploit will have quite a big impact, since taking over (any) user’s sessions is a very dangerous thing, even more so when the user is an administrator. To minimize the impact you should take several security measures. First of all you should configure PHP to mark the session cookie as HTTP-only. This makes it impossible to read from or write to the session cookie in JavaScript. Then you should find a way to detect hijacked sessions and reject them. There are many more options here, and they all reduce the chance of a security breach, but equally important: they also reduce the impact.

Reflection

It is clear that you should try to find and remove vulnerabilities and that you should also minimize the impact of a possible (even an improbable) attack. But what’s also important is that you know what you are trying to achieve by a certain security measure. It helps you with determining the right solution, when you know what the problem is that you are trying to solve. Be very reflective about your own ways. Sometimes it will occur to you that you are taking measures that won’t help at all, since you already have a stronger security measure in place. And sometimes you will notice that one very insecure part could indirectly give a threat agent access to another, secure part of the system.

Before diving in…

There is so much you can do when it comes to (web) application security - you probably won’t be able to implement it all. You also won’t need to. Talk with your team members about the necessary security measurements, and some quick wins. But also talk with your manager about the time budget and the importance of security for a specific project (or the organization). Security should be a team effort, and a secure application comes for the biggest part from awareness and discipline. So make sure that everybody on your team thinks alike when it comes to securing the application that you’re creating.

In the following chapters I will show you ways in which you can prevent bad things from happening to your Symfony application, but also how you can apply some damage control here and there.

13. Authentication and sessions

As I mentioned earlier, Symfony already takes care of many cases concerning logging in and out, and also has settings for many of the desired behaviors, available in config.yml and security.yml. For instance you can alter some security-related PHP settings directly from within config.yml:

framework:
    session:
        # session cookie name
        name: matthias_session

        # session cookie should not be accessible using JavaScript
        cookie_httponly: true

        # session data should expire in n seconds (when not used)
        gc_maxlifetime: 3600

        # expired session data will be garbage collected with a 1:10 chance
        gc_probability: 1
        gc_divisor: 10

Symfony takes care of migrating the session (effectively changing the session ID) when you are logging in, to prevent the old, unauthenticated session from getting extra rights (should this session be compromised at any time). Symfony also takes care of invalidating the authenticated session after logging out, to prevent session hijacking. This works just fine by default, but could be configured explicitly in security.yml:

security:
    # after authentication, the session will be migrated
    session_fixation_strategy:  migrate

    firewalls:
        secured_area:
            logout:
                # the authenticated session will be unavailable afterwards
                invalidate_session: true

13.1 Invalidating sessions

Session hijacking

All nice and well, but there are some things missing. In the first place, when you would be able to track down the session id of an authenticated user and use it in your own session cookie to take over the authenticated session of a user, no one would notice. You would also not be hindered in your attempt. This means, you have to track changes in the “signature” of the user. For instance, you could check the client’s IP address, or their User Agent. Changes in these client characteristics should at the very least be logged, so that you can monitor suspicious behavior. But you may also require a user to re-authenticate and thereby confirm that the changed signature was not intended as a threat to your application.

namespace Matthias\SecurityBundle\EventListener;

use Matthias\SecurityBundle\Exception\UserSignatureChangedException;
use Symfony\Component\HttpKernel\Event\GetResponseEvent;
use Symfony\Component\HttpKernel\HttpKernelInterface;

class UserSignatureListener
{
    public function onKernelRequest(GetResponseEvent $event)
    {
        if ($event->getRequestType()
                !== HttpKernelInterface::MASTER_REQUEST) {
            return;
        }

        $request = $event->getRequest();

        $clientIp = $request->getClientIp();

        $userAgent = $request->headers->get('user-agent');

        if (...) {
            throw new UserSignatureChangedException();
        }
    }
}

Using the service definition below, you could register the event listener:

<service id="matthias_security.user_signature_listener"
  class="Matthias\SecurityBundle\EventListener\UserSignatureListener">
    <tag
      name="kernel.event_listener"
      event="kernel.request"
      priority="100"
      method="onKernelRequest" />
</service>

The UserSignatureChangedException that will be thrown in the request listener should of course be handled by an exception listener (listen to KernelEvents::EXCEPTION), which sets the appropriate response object, for instance a RedirectResponse to a page where the user can re-authenticate.

Long-running sessions

Say, an authenticated user of your application does nothing for a while, but they keep the browser open. The session data is not being requested for some time, and its lifetime is almost expired. Then, the user refreshes the page. The session cookie containing the session ID is set only to expire when the browser is closed (cookie_lifetime = 0). The cookie will still be valid, as well as the session data (its lifetime almost expired), so the user can continue with their session as if nothing happened. This way there is nothing that prevents a user from having an eternal session.

You may want to invalidate these long-running sessions, based on the date they were last used, or when they were first created. Symfony has no built-in way to do this, but it is very easy to implement some custom session invalidation yourself. The session has a so-called MetadataBag containing Unix timestamps for the time the session was first created and the time its data was last changed.

use Symfony\Component\HttpKernel\Event\GetResponseEvent;
use Symfony\Component\HttpKernel\HttpKernelInterface;

class SessionAgeListener
{
    public function onKernelRequest(GetResponseEvent $event)
    {
        if ($event->getRequestType()
                !== HttpKernelInterface::MASTER_REQUEST) {
            return;
        }

        $session = $event->getRequest()->getSession();
        $metadataBag = $session->getMetadataBag();

        $lastUsed = $metadataBag->getLastUsed();
        if ($lastUsed === null) {
            // the session was created just now
            return;
        }

        $createdAt = $metadataBag->getCreated();

        $now = time();

        // $now, $lastUsed and $createdAt are Unix timestamps

        // if a session is being revived after too many seconds:
        $session->invalidate();

        // create some nice response to let the user know this happened:
        $event->setResponse(...);
    }
}

The service definition below activates the SessionAgeListener:

<service id="matthias_security.verify_session_listener"
  class="Matthias\SecurityBundle\EventListener\SessionAgeListener">
    <tag
        name="kernel.event_listener"
        event="kernel.request"
        priority="100"
        method="onKernelRequest" />
</service>

14. Controller design

When it comes to controller design, I recommend creating small controller classes. First of all this means grouping logically related actions together in one controller class, but also creating only small actions, with no real logic inside and at most two possible execution paths. This may mean that you end up with many controllers, but this method has many advantages, which are also related to security:

  1. When you need to change some user-related behavior, you can very quickly find your way from URL, to route name, to controller.
  2. When some controllers need special attention from a security point of view, you can easily spot them and concentrate the protection logic in one place.

To illustrate the second point, let’s say you have one controller concerned with personal settings, but it relates to data like strictly personal account information, payment information like a credit card number and public profile data, like a nickname or an avatar. I would recommend splitting the pages concerned with viewing and editing this data over multiple controllers, each with a simple and recognizable name. Inside the controllers, each action should also have a recognizable name, which can be longer than you think, like AccountController::editAccountInformation (instead of just “edit”) and PaymentController::listCreditCards. This allows you to align your personal alertness level with the level required by the specific code you are going to work on.

Furthermore it should be clear by looking at the code, which controllers are used for modifying the state of the application, i.e. persist or remove some data. You should explicitly define the HTTP methods allowed for a given action. In this sense, pages that can be requested using the GET method are innocent, and pages to which users can POST data, are not.

use Sensio\Bundle\FrameworkExtraBundle\Configuration\Method;

/**
 * @Method("GET")
 */
public function viewPublicProfileAction()
{
}

/**
 * @Method("GET")
 */
public function editPublicProfileAction()
{
}

/**
 * @Method("POST")
 */
public function updatePublicProfileAction()
{
}

Of course, it is already bad when someone can view an edit form without the given object being an object they can modify, but it is worse if they can do real modifications to the object.

14.1 Secure actions

There are many ways in which you can secure actions. The action code itself will sometimes throw an AccessDeniedException, and of course, you should check that a certain object belongs to someone, or can be modified by someone (either because the owner is stored as a property of an object, or because the correct rights are registered using ACL). But you should also implement role management in some way. Roles are very cheap when using Symfony. Just think of a new one and it already exists. When existing roles should include the new role, add it to the “role hierarchy” as defined in security.yml.

It is very important to start adding a list of roles for any controller which requires a user to be logged in. There are some good reasons for this. First of all, when it comes to security you should adhere to the principle of “least privilege”. This means that by default an authenticated user can do nothing at all (maybe only change their password). You, the system, or an administrator will have to give him extra rights first. You can use all kinds of expressions concerning roles above an action, using the @PreAuthorize annotation, but in most situations I have encountered, a simple @Secure suffices:

use JMS\SecurityExtraBundle\Annotation\Secure;

/**
 * @Secure("ROLE_PAGE_EDITOR")
 */
public function editPageAction()
{
}

Often there are different types of users who should have access to this editPageAction, for instance a user with ROLE_ADMINISTRATOR but also someone with ROLE_CONTENT_MANAGER. The solution here is not to add extra roles to the @Secure annotation (like @Secure({"ROLE_ADMINISTRATOR", "ROLE_CONTENT_MANAGER"})) but to solve this by changing the role hierarchy:

security:
    role_hierarchy:
        ROLE_ADMINISTRATOR: [ROLE_PAGE_EDITOR]
        ROLE_CONTENT_MANAGER: [ROLE_PAGE_EDITOR]

When it comes to roles, remember:

  • Roles are volatile: you are free to create and almost free to remove them (basically they are just strings)
  • It works best when a role describes the user having the role (role names ideally end with “MANAGER”, “EDITOR”, “MODERATOR”, etc.)

14.2 Putting controllers behind the firewall

Now that you have many controllers all inside something like a SettingsBundle, you should import them all in this bundle’s routing.yml file:

SettingsBundleControllers:
    resource: "@SettingsBundle/Controller/"
    type: annotation

This entire file can be imported at once from the applications’ routing.yml file. This enables you to define a prefix for all routes in the SettingsBundle together:

SettingsBundle:
    resource: "@SettingsBundle/Resources/config/routing.yml"
    type: yaml
    prefix: /settings

Using this prefix, you can easily define a set of required roles for any URI starting with /settings:

security:
    access_control:
        - { path: ^/settings, roles: [IS_AUTHENTICATED_FULLY] }

The line above means that you need to be fully authenticated to do anything on settings-related pages (when you are logged in just by a “remember me” cookie, this would not suffice to grant you access). This is a good thing, and you now also have a single point of control for this security measure.

15. Input validation

Any input your application receives from the outside world should be checked and processed before actually being used or stored in any way.

15.1 Safe forms

Symfony has several important tools that you can use in the fight for clean input: the Form Component and the Validator Component. Most Symfony developers already trust the Form Component with their lives:

if ($request->isMethod('POST')) {
    $form->bind($request);
    if ($form->isValid()) {
        // persist!
    }
}

This is mostly a good thing since the Form and Validator Component both are quite good, and they can be trusted to do what they say. But: you should not trust them blindly (what can you trust blindly?).

HTML5 validation

Back in the days when we used to write PHP applications by hand from start to end, we were taught to validate form input. For instance we did not want users to leave the “name” field blank. But now we can build a form in a form (which represents the field) and we set it’s required option to true. When you open the form in the browser and try to submit it with no values inserted, it will show you a nice error message (which is in fact a feature of HTML5) instead of immediately submitting the form. However, if you disable this client-side validation, by adding a novalidate HTML5 attribute to the form tag, you will notice that by default there is no server-side validation at all. So, the first thing you need to do when working with forms (and maybe for the dev environment only) is disabling the HTML5 validation:

<form{% if app.debug %} novalidate="true"{% endif %}>

Now you can test the server-side validation of your form.

Validation constraints

Usually you have some domain object, let’s say an entity, that you want to create or modify using a form. You would then set the form’s data_class option to the class name of this entity. In the buildForm method you can instruct the FormBuilder to add certain fields corresponding to attributes of the entity. Now to validate the form data, you need to be very secure in adding constraints to the attributes (in my opinion preferably using annotations). This is nothing but the usual, and it’s very well documented:

namespace Matthias\AccountBundle\Entity;

use Symfony\Component\Validator\Constraints as Assert;

class User
{
    /**
     * @Assert\Email()
     * @Assert\NotBlank()
     */
    private $emailAddress;
}

Forms without an entity

When you have data which does not have a one-to-one correspondence to an entity, the Forms documentation tells you that you can just provide no data class and in fact build the form inline inside your controller:

public function contactAction(Request $request)
{
    $defaultData = array('message' => 'Type your message here');
    $form = $this->createFormBuilder($defaultData)
        ->add('name', 'text')
        ->add('email', 'email')
        ->add('message', 'textarea')
        ->getForm();

    $form->handleRequest($request);

    if ($form->isValid()) {
        // data is an array with "name", "email", and "message" keys
        $data = $form->getData();
    }
}

This to me sounds like a bad idea, for two reasons: first of all, you should always define a form using a custom form type (extending from AbstractType), like this:

use Symfony\Component\Form\AbstractType;
use Symfony\Component\Form\FormBuilderInterface;

class ContactFormType extends AbstractType
{
    public function buildForm(
        FormBuilderInterface $builder,
        array $options
    ) {
        $builder
            ->add('name', 'text')
            ->add('email', 'email')
            ->add('message', 'textarea')
        ;
    }

    public function getName()
    {
        return 'contact_form';
    }
}

This will put all the logic related to the form in one place, making it reusable and much better maintainable.

But you should also always provide a data_class option. The data from the contact form will not be persisted, but still it has a certain coherent structure and also a very clear meaning. So choose a suitable name for the data object of your form, like ContactDetails, create a class with the right fields and add some assertions to each field to ensure the object contains consistent data:

use Symfony\Component\Validator\Constraints as Assert;

class ContactDetails
{
    /**
     * @Assert\NotBlank
     */
    private $name;

    /**
      * @Assert\NotBlank
      * @Assert\Email
      */
    private $email;

    /**
     * @Assert\NotBlank
     */
    private $message;

    // add getters and setters for all fields
}

Now set this class as the data_class option of the ContactFormType:

use Symfony\Component\OptionsResolver\OptionsResolverInterface;

class ContactFormType extends AbstractType
{
    public function setDefaultOptions(OptionsResolverInterface $resolver)
    {
        $resolver->setDefaults(array(
            'data_class' => 'LifeOnline\ContactBundle\Model\ContactDetails'
        ));
    }
}

Your data will be in much better shape since it is now encapsulated, validated and totally under control. You can:

  • Prevent bad data from entering the attributes by filtering the data in setters:
    class ContactDetails
      {
          public function setMessage($message)
          {
              $this->message = strip_tags($message);
          }
      }
    
  • Provide default values a lot easier (without having to know the structure of the expected array):
    public function contactAction(Request $request)
      {
          $contactDetails = ContactDetails::createForUser($this->getUser());
    
          $form = $this->createForm(new ContactFormType(), $contactDetails);
    
          // ...
      }
    
  • Work with an object of a known type, after binding and validating the form:
    public function contactAction(Request $request)
      {
          $form = $this->createForm(new ContactFormType());
    
          if ($request->isMethod('POST')) {
              $form->bind($request);
              if ($form->isValid()) {
                  $contactDetails = $form->getData();
    
                  // $contactDetails is an instance of ContactDetails
              }
          }
      }
    

15.2 Validate values from Request

The Request object is in part just a wrapper for PHP’s superglobals. In /web/app.php you can see how the Request object gets created:

$request = Request::createFromGlobals();

Inside this method you find this simple line of code:

$request = new static($_GET, $_POST, array(), $_COOKIE, $_FILES, $_SERVER);

So even though it may feel much more secure to fetch request data using

$request->query->get('page')

it is in fact no more secure than just $_GET['page']. Though within the Symfony context it is always better to use the methods on the Request class, you should still be wary about these values, validate them and force them into the right format. Strict typing, and checking for allowed values and ranges is quite necessary.

Request attributes

Route parameters

First of all, when a URI pattern contains wildcard values, like id in /comment/{id}, and a certain request matches this pattern, the RouterListener will make sure that all the parameters are copied over to the request object as attributes. These request attributes will be used by the ControllerResolver to collect arguments for the controller based on its parameter names and type hints. See Events leading to a response for a more detailed explanation of this process.

Since most controller arguments are being copied more or less directly from the URI of the request, you need to be extra careful when it comes to handling them. First of all you have to think about requirements for route parameters (like id). Requirements are to be defined as regular expressions. The default requirement for route parameters is [^/]+: they may contain any character except a slash. So you must always define your own requirements, like \w+, which means at least one “word” character (i.e. a-z, A-Z or 0-9, or underscore) or \d+, which means at least one “number” character (i.e. 0-9). If you don’t know regular expressions yet, you should really learn them (as you may have noticed they are also used when working with paths defined in security.yml). Below you will find some examples of route parameters with requirements:

/**
  * id can only be a number:
  * @Route("/comment/{id}", requirements={"id"="\d+"})
  */

/**
 * alias can be at least one lowercase letter, a number or a dash:
 * @Route("/user/{alias}", requirements={"alias"="[a-z0-9\-]+"})
 */

/**
 * type can only be commercial or non-commercial:
 * @Route("/create-account/{type}", requirements={
 *   "type"="commercial|non-commercial"
 * })
 */

It is important to verify the correctness of these values in this early stage, otherwise incorrect values might result as arguments of method calls to services, where they may cause InvalidArgumentExceptions, or worse: unexplainable errors.

Query or request parameters

There are also situations where you want to be more flexible concerning the input for your controller: you may want to support query parameters in your URI, like ?page=1, or request parameters sent using the content of the request (e.g. POST data). But these values are by no means trustworthy. Query and POST data can both easily be modified to not be what you expect it to be. Therefore, when you retrieve this data from the Request object:

  • Let validation be handled by the Form and Validator Component by using Form::bind() and Form::isValid()
  • Or:
    • Cast the values to the type that you expect (in most situations either a string or an integer) and
    • Validate these values yourself by calling the validateValue() method of the validator service. You could think of validating the range of a value (“should be at least 1”), or one of multiple options (“should be any of ‘commercial’, ‘non-commercial’”).

For instance:

use Symfony\Component\Validator\Constraints\Choice;

public function createAccountAction()
{
    $userTypeConstraint = new Choice(array(
        'choices' => array('commercial', 'non-commercial')
    );

    $errorList = $this->get('validator')->validateValue(
        $request->query->get('userType'),
        $userTypeConstraint
    );

    if (count($errorList) == 0) {
        // the provided user type is valid
    }
}
Use the ParamFetcher

There is one tool that I would like to mention here. It is part of the FOSRestBundle which provides many tools for easily creating a REST-like webservice. It provides a very nice way of pre-validating query parameters by first adding some configuration in the form of an annotation for each query parameter that you need. For example, this is how you can configure a page query parameter which should consist only of digits, being “1” by default:

use FOS\RestBundle\Request\ParamFetcher;
use FOS\RestBundle\Controller\Annotations\QueryParam;

/**
 * @QueryParam(
 *   name="page",
 *   requirements="\d+",
 *   default="1",
 *   description="Page number"
 * )
 */
public function listAction(ParamFetcher $paramFetcher)
{
    $page = $paramFetcher->get('page');
}

It works almost exactly the same for POST data without a corresponding form:

use FOS\RestBundle\Request\ParamFetcher;
use FOS\RestBundle\Controller\Annotations\RequestParam;

/**
 * @RequestParam(name="username", requirements="\w+")
 */
public function deleteAccountAction(ParamFetcher $paramFetcher)
{
    ...
}

Even though your application will not have a REST API, you can still just install the FOSRestBundle bundle without many side-effects. Just make sure you disable all built-in listeners except the ParamFetcherListener:

# in /app/config/config.yml
fos_rest:
    param_fetcher_listener: true

15.3 Sanitizing HTML

You likely have forms in your application allowing users to enter some kind of rich text, with bold, italic or underlined text, and maybe you allow users to add their own link tags. Well, this is of course a very dangerous thing when it comes to security. You should take very good care of restricting users in their use of HTML tags but also in them adding HTML attributes, which can actually be more dangerous, especially when they contain JavaScript. And don’t think that alert('Hi!'); is the worst thing a user can do. Even mismatching tags can spoil things - think about what a simple </div> in some random spot could do in most sites.

First of all: don’t rely on things like Markdown to solve your HTML injection problem. The specification of Markdown clearly says that any HTML in the source text should be left intact and added “as is” to the generated output. Second: don’t rely on the second argument of strip_tags to allow some tags since this also allows any attribute for these tags. Third: don’t dream up your own regular expressions for allowing some tags and attributes. There will always be hacks that circumvent your handwritten rules. Instead, use HTMLPurifier and configure it properly for specific situations allowing rich (HTML) text.

There is an excellent bundle which integrates HTMLPurifier with any Symfony application: the ExerciseHTMLPurifierBundle. It defines HTMLPurifier services for each configuration set you define in config.yml. It also supports automatic filtering of form values.

Automatic sanitizing

The thing is: with every Symfony application you have to manually take care of sanitizing input. It would be great if all request attributes would be automatically filtered according to some rules, or when every form or entity would contain only known-to-be-clean values. Unfortunately there is no such thing in the open source world that I am aware of right now. A prototype of the desired functionality can be found in the DMSFilterBundle, but in its current state this bundle does not handle every situation well. It filters only form input for the root objects of forms. It also does not filter values manually set upon entities, documents, etc.

16. Output escaping

16.1 Twig

Symfony developers are told that:

If you’re using Twig, output escaping is on by default and you’re protected.

Everyone will be very happy after reading this, except a security-minded person, who is very skeptical about remarks like this. There is no such thing as automatic output escaping that works in all situations. The default output escaping for Symfony assumes that you are rendering in an HTML context, at the element level. Also Twig follows many special (though secure rules) for auto-escaping, which you should know about. Read about all of them in the Twig documentation for developers.

Know your escaping context

Most important is this assumption about the default context being html. This will likely not be the only context in your application. You need to find out which variables are being printed where and in what context. Twig supports these contexts out of the box:

html
When rendering variables at the HTML element level
html_attr
When rendering variables inside HTML attributes
js
When rendering variables inside JavaScript code
css
When rendering variables inside a stylesheet
url
When rendering a part of a URL (like a query parameter)

When you have determined the context and it is something other than html, escape as appropriate using the Twig escape filter:

{{ someVariable|escape('js') }}

Escaping function output

You always have to be at guard when it comes to auto-escaping. Especially when you create your own Twig functions, (or filters). The output of your functions will also be automatically escaped in the current context (usually html):

class MyTwigExtension extends \Twig_Extension
{
    public function getFunctions()
    {
        return array(
            \Twig_SimpleFunction('important', function($thing) {
                return sprintf(
                    '<strong>%s</strong> is important to me',
                    $thing
                );
            })
        );
    }
}

The result of this function will be auto-escaped, so:

{{ important('My family') }}

Will result in:

&lt;strong&gt;My family&lt;/strong&gt; is important to me

Except when you add an is_safe option when defining the function in your Twig extension:

class MyTwigExtension extends \Twig_Extension
{
    public function getFunctions()
    {
        return array(
            \Twig_SimpleFunction('important', function($thing) {
                ...
            }, array(
                'is_safe' => array('html')
            )
        );
    }
}

Now, the output will not be escaped. But, so will the input!

Escaping function arguments

When we have marked the output of a function as safe but we nevertheless include function arguments in this same output, we end up with a totally unsafe function. So:

{{ important('<script>alert("Security")</script>') }}

Will result in:

<strong><script>alert("Security")</script></strong> is important to me

Now of course you would not type such a thing in your own templates like this, but most of the time you just pass some user-supplied value as an argument to Twig functions (and remember that even most of the things in your database originate from user input). And such a value can be anything, assuming it may not have been correctly sanitized.

Well, there is an easy solution: add a pre_escape option:

class MyTwigExtension extends \Twig_Extension
{
    public function getFunctions()
    {
        return array(
            \Twig_SimpleFunction('important', function($thing) {
                ...
            }, array(
                'is_safe' => array('html'),
                'pre_escape' => 'html'
            )
        );
    }
}

Now all arguments provided to the function will first be escaped for the given context.

Be wary of the raw filter

You can override any auto-escaping functionality by ending an expression with the raw filter:

{{ message|raw }}

This would escape nothing, and when the message variable contains some HTML, it will be left as-is. This means that you have to be sure that everything has been done to prevent bad things from happening. Which in most cases means: you must be sure that this variable has gone through some kind of sanitizing process (like purifying it using HTMLPurifier, see above). Or (and preferably and) that the source of the value can be trusted - for example an imported piece of source code, written by a developer. Not, however, the value of a textarea field submitted by an “administrator”. If the security of your authentication system has by any means be compromised, someone may enter some very malevolent text in that specific field and immediately affect many users of your application. I am not saying that you should be scared, but rather cautious.

17. Being secretive

In many cases hacking a system means getting to know as much as possible about its inner workings, to know what its sensitive and thus exploitable parts are. Therefore you need to be careful about what you display to outsiders.

17.1 Mask authentication errors

The Symfony documentation shows a very bad example of application security, where literal exception messages are shown above an authentication form (error is an Exception object):

{% if error %}
    <div>{{ error.message }}</div>
{% endif %}

Directly afterwards we are encouraged to “use the exception message wisely”, since it may contain sensitive information. It definitely will, as I have noticed on many occasions. Sometimes even raw database errors, including entire schema definitions can end up in the template when you display the exception message directly.

Luckily all exceptions that extend from AuthenticationException have a method getMessageKey. When calling this method you get a cleaner exception message, with no (sensitive) details about the exceptional situation itself. Just simple messages like “Invalid credentials.” or “Authentication request could not be processed due to a system problem.”.

So the correct Twig template around a login form would display errors like this:

{% if error %}
    <div>{{ error.messageKey }}</div>
{% endif %}

17.2 Prevent exceptions from showing up

When there are places in your application where you want to use exceptions for showing an error message to the user, there are three possible strategies. Either use something like described above (using a getMessageKey() method, which contains a message that is more user-friendly and also more vague) or wrap lower-level exceptions in more high-level, general exceptions and set a new message that is safe for display. You should also be wary of general exceptions occurring at deeper levels of abstraction, which you mistakenly take for one of your own exceptions. You might also mark some exceptions as safe, by filtering them:

$error = null;

try {
    // do something dangerous
    ...
} catch (SomeSpecificException $exception) {
    $error = $exception;
} catch (\Exception $previous) {
    // unexpected, and unknown exception
    $error = new GenericException('An error occurred', null, $previous);
}

// $error can be safely used now

The problem with this approach is that Symfony’s default exception handler (which also logs exceptions when logging is enabled) will not be notified of the truly unexpected exception. As it is most likely a symptom of some problem on a deeper level of the application code, this may not be desired behavior. Therefore, in most situations it suffices to catch specific, known exceptions and let any other exception just bubble up.

17.3 Customize error pages

When such a “normal” exception, indicating a system failure, bubbles up to the standard exception handler, you must be sure to prevent anything from being displayed in the response to the user. Any system information leaked to a user may be used to determine a strategy for compromising the system. You should therefore create custom error pages which blend in well with the rest of your application. Never show any generated information about the problem, especially no stack trace. Problems will be logged automatically and can also be sent to you by email, either by configuring Monolog to do so, or by using some other logging tool which sends you alerts when something goes wrong. The latter may be a better choice, since maybe the application has become so unstable that it can not be trusted to do so itself.

17.4 Be vague about user data

Authentication problems are mostly covered by the Symfony Security Component. There are however many bad interaction design decisions that may accidentally disclose sensitive information about users of your system. When your system leaks information about which other users there are, your system is susceptible to something called “harvesting”. These are some examples of interaction design that leak information about users.

  1. A user can reset their password on the “Forgot password” page which contains an “Email” text field. When the user provides their email address and submits the form, the system finds the account record of the corresponding user and generates a “password reset link”, then sends a mail to the user containing this link. After submitting the form, the page says: “We have sent a password reset mail to the given email address”.

    The problem is: when someone tries to collect email addresses of users, they will now know that a user with the provided email address actually exists in the system.

  2. Same situation as above, but now the message is: “If the provided email address belongs to one of our users, we have sent a password reset mail to it.”

    This seems secure, however: the request duration is notably longer when an email address is of a known user of the system, than when the email address is unknown. Someone who is harvesting user email addresses may infer based on the length of the request whether an email address is known or unknown to the system. This is called a timing attack.

  3. When the user is logged in they are allowed to change their email address using a simple form with one field: “Email”. After submitting the form the system changes the user’s email address or it shows a message to the user saying: “The provided email address is already in use”.

    Again: we now know the email address of a user, even though this is sensitive data.

The best solution to prevent harvesting is to be vague: don’t say if there was a match, if you have sent an email, etc. You also have to make sure that it can not be guessed if any of these things has been done by measuring the duration of the request.

The solution to the third problem would be:

  1. Allow the user to change their email address, but don’t persist the change immediately.
  2. Send an email to the new address, with a link allowing the user to confirm the requested change.
  3. After confirming the change, update the account information with the new email address.

VII Using annotations

18. Introduction

The first release of Symfony2 came bundled with the Doctrine Common library. This library contains some tools used by all the second-generation Doctrine projects, like the Doctrine ORM, Doctrine MongoDB ODM, etc. The library offers shared functionality and utility classes, like an event manager, persistence- related interfaces, some useful base classes and: an annotation reader.

The annotation reader was initially only used by Doctrine ORM to parse annotations inside entity classes. Doctrine ORM (or Doctrine2) introduced annotations as a way to specify how a simple PHP class should be mapped to a relational database table:

use Doctrine\ORM\Mapping as ORM;

/**
 * @ORM\Entity
 */
class User
{
    /**
     * @ORM\Column(type="string")
     */
    private $name;
}

This “syntactic sugar” was soon recognized to be a really useful alternative for the existing solutions, like writing down your mapping configuration in a Yaml or XML file. The creators of Symfony2 also embraced annotations and started adding them to their components as an alternative way of loading configuration. For instance, the Validator component supports annotations to configure validation rules for classes. And the Routing component has annotations for linking a route to a class and one of its methods.

Annotations: Domain-specific languages

Annotations have some very interesting characteristics. They are not written in the PHP language: @Route("/admin/", name="admin_index") can not be parsed by the PHP interpreter. Nevertheless, annotations have a “runtime” aspect, since they are parsed by an annotation reader from within a running PHP application. Their mere presence has some very real effects on the way the application runs.

Furthermore, annotations are easy to write. Once you know all the attributes an annotation supports, you know all there is to using it in your application. The funny thing is: you don’t even need to be a PHP programmer to use an annotation, since you are not required to write any PHP code.

Finally, annotations are always domain-specific. They are introduced as a means to convey some high-level concept from a specific domain, like routing, templating, persistence, etc. Annotations help you abstract from the details and instead think in higher-level concepts.

These characteristics add up to the following conclusion: we should consider annotations as a form of domain-specific languages (DSL). A DSL is a small language that is used to convert higher-level concepts into lower-level details. It is implemented in a general purpose language like PHP, but can have its own syntax (in this case “annotation syntax”).

Each domain specific set of annotations forms a domain-specific language to express the details underlying the high-level domain concepts. For example, the annotations provided by Doctrine ORM form a little language to express inside an entity class at a high level what should happen inside the relational database at a low level to make it ready for storing entities in it. This means that when you use Doctrine annotations to configure the mapping for an entity, you don’t have to worry about implementation details like: what is the best way to define a boolean column, or what was the syntax for making a column’s value “required” again?

When I first started to work with Symfony2 I saw everyone using annotations. I felt some hesitation to use them too. It seemed to me quite dangerous to use comments to influence the flow of an application. Then after a couple of months I got used to working with annotations and as it happens I never encountered a “dangerous situation” caused by an annotation after all.

Instead, I learned more about annotations: what they really are and how they are used by Symfony, Doctrine, etc. Soon I started creating my own annotations and using them for many different purposes inside the Symfony applications I’ve been working on. In this chapter I’d like to share my findings with you. You will learn about the inner workings of annotations and how to create them yourself. In the last chapter of this part I show you some Symfony-specific ways in which you can use annotations to influence the application’s flow.

19. An annotation is a simple value object

You will always find annotations inside so-called doc blocks, which are basically comments but they start with /**. Regular annotations start with @ and serve as documentation for the code that immediately follows it, like this:

/**
 * @param int $id The ID of the user
 * @return array  Template variables
 */
public function editAction($id)
{
    ...
}

When parsing a doc block, the Doctrine annotation reader will skip all regular annotations, like @param and @return, based on a pre-configured list. When an annotation is not on that list, the annotation reader will assume that the annotation is a class name. For example, the @Route annotation of the UserController::editAction() in the code sample below is actually a class name: Sensio\Bundle\FrameworkExtraBundle\Configuration\Route, which has been imported by the use statement:

use Sensio\Bundle\FrameworkExtraBundle\Configuration\Route;

class UserController
{
    /**
     * @Route("/users/{id}/edit", name="user.edit")
     */
    public function editAction($id)
    {
        ...
    }
}

The annotation reader will try to create an instance of that class and fill it with the data between the brackets ("/users/{id}/edit", name="user.edit"). The resulting Route object can then be used by the routing loader to add an extra route to the RouteCollection.

In every Symfony application an instance of the Doctrine annotation reader is already available as the annotation_reader service:

// get the service container
$container = ...

$reader = $container->get('annotation_reader');

To parse the doc block of the UserController::editAction() we need to create a reflection object for that method first:

$method = new \ReflectionMethod('UserController', 'editAction');

Then we can ask the annotation reader to parse the annotations of the reflected method:

$annotations = $reader->getMethodAnnotations($method);

The result of parsing the method annotations would be an array with one Route object. The object properties contain the attributes that were provided with the @Route annotation:

print_r($annotations);

/*
Array
(
    [0] => Sensio\Bundle\FrameworkExtraBundle\Configuration\Route Object
        (
            ...
            [path]
                => /users/{id}/edit
            [name]
                => user.edit
            ...
        )
)
*/

Each annotation class (like the Route class) needs to have an @Annotation annotation itself in order to be recognized by the annotation reader as a valid annotation:

/**
 * @Annotation
 */
class MyAnnotation
{
}

With the @Annotation annotation in place, the MyAnnotation class can be used as a real annotation for a class, a method or a property:

/**
 * @MyAnnotation
 */
class SomeClass
{
    /**
     * @MyAnnotation
     */
    private $someProperty;

    /**
     * @MyAnnotation
     */
    public function someFunction()
    {
    }

    /**
     * @MyAnnotation(@MyAnnotation)
     */
    public function otherFunction()
    {
    }
}

It is even possible to use @MyAnnotation inside another @MyAnnotation, as you can see inside the doc block of SomeClass::otherFunction().

When we ask the annotation reader to parse the class annotations for SomeClass, it will return an array with one object: an instance of MyAnnotation:

// get the annotation reader
$reader = ...

$class = new \ReflectionClass('SomeClass');
$annotations = $reader->getClassAnnotations($class);

print_r($annotations);

/*
Array(
    [0] => MyAnnotation object
)
*/

19.1 Adding attributes to your annotation

To be able to define your own annotations is already very nice, but they become more useful when you store some data inside them. This data can be provided by the user when they add your annotation to one of their classes, methods, etc.

The annotation parser supports many different kinds of syntaxes for populating the attributes of an annotation. It accepts strings, numbers, boolean values, arrays and objects (which are themselves annotations). For example:

/**
 * @MyAnnotation(
 *   "some string",
 *   "hashMap" = {
 *     "key" = "value"
 *   },
 *   "booleanValue" = true,
 *   "nestedAnnotation" = @MyAnnotation
 * )
 */

Any logical combination of types is possible. For instance, it’s possible to put each scalar or object type inside a hash map.

When the user has provided some data for the annotation, like in the example above, the annotation reader needs to pass that data to the annotation object it creates. There are two different strategies which the annotation parser can apply.

Passing the attributes via the constructor

First, the annotation parser will look inside the MyAnnotation class for the presence of a constructor. If it finds one, it passes all the attributes as the first argument of the constructor when it creates an instance of the MyAnnotation class:

/**
 * @Annotation
 */
class MyAnnotation
{
    public function __construct(array $attributes)
    {
        // "some string"
        $value = $attributes['value'];

        // array('key' => 'value', ...)
        $hashMap = $attributes['hashMap'];

        // true
        $booleanValue = $attributes['booleanValue'];

        // an instance of MyAnnotation
        $nestedAnnotation = $attributes['nestedAnnotation'];
    }
}

From then on you can do anything you like with those values. You would probably validate them and store them in private properties.

Populating public properties with the provided attributes

If there is no constructor (or the constructor has no first argument), the parser will try to copy the attributes that were provided into public properties of the MyAnnotation class:

/**
 * @Annotation
 */
class MyAnnotation
{
    public $value;
    public $hashMap;
    public $booleanValue;
    public $nestedAnnotation;
}

Using this second strategy you have no chance to validate the attributes before they are copied into the public properties. Luckily it is possible to add some basic validation rules directly inside the annotation class itself.

Validation using @Attributes

We can add an @Attributes annotation to the MyAnnotation class. It accepts an array of @Attribute annotations that can be used to describe each supported attribute: the type of value that is expected, and whether or not the attribute is required.

/**
 * @Annotation
 * @Attributes({
 *   @Attribute("value", type="string", required=true),
 *   @Attribute("hashMap", type="array<string>", required=false),
 *   @Attribute("booleanValue", type="boolean"),
 *   @Attribute("nestedAnnotation", type="MyAnnotation")
 * })
 */
class MyAnnotation
{
    public $value;
    public $hashMap;
    public $booleanValue;
    public $nestedAnnotation;
}

By default attributes are not required. When an optional attribute has not been provided, its value will be null. When an attribute is of type array then the value that is provided will be converted to an array automatically, so "some string value" becomes array("some string value").

Validation using @var and @Required

The validation options provided by using the @Attributes annotation are very useful. But if you don’t like the fact that the rules for each property are not directly above that property’s definition, you can also choose not to use @Attributes and instead add type declarations for each property:

/**
 * @Annotation
 */
class MyAnnotation
{
    /**
     * @var string
     * @Required
     */
    public $value;

    /**
     * @var array<string>
     */
    public $hashMap;

    /**
     * @var boolean
     */
    public $booleanValue;

    /**
     * @var MyAnnotation
     */
    public $nestedAnnotation;
}

To mark an attribute of the annotation as required, add @Required to the doc block above the corresponding property.

There is one nice option that adds a little extra to the validation process: the @Enum annotation. You can use it to define which values are allowed for a given attribute. This annotation has to be in the doc block of the relevant property. It works in combination with the @Attributes annotation as well as with the @var validation:

/**
 * @Annotation
 */
class MyAnnotation
{
    /**
     * @Enum({"yes", "no"})
     */
    $answer;
}

Now @MyAnnotation(answer="yes") would be valid, while MyAnnotation(answer="unsure") would trigger an error.

19.2 Limiting the use of an annotation

Different kinds of annotations have different use cases. For instance, the @Entity annotation from the Doctrine ORM does only make sense for classes, not for methods. On the other hand, the @Template annotation from the SensioFrameworkExtraBundle should only be used for methods, not for classes. Likewise, some annotations can only be applied to properties, like @Type from the JMS Serializer. And some annotations can only be applied inside another annotation as one of its attributes, like the @Attribute annotation that can only occur inside an @Attributes annotation.

These different use cases are called “targets” and we can configure them ourselves per annotation class, by adding a @Target annotation to it:

/**
 * @Annotation
 * @Target("CLASS")
 */
class MyAnnotation
{
}

The available targets are: CLASS, METHOD, PROPERTY, ANNOTATION and ALL (which is the default target for annotations if you don’t specify it). If an annotation has multiple targets, you can provide them as an array of strings:

/**
 * @Annotation
 * @Target({"CLASS", "METHOD"})
 */
class MyAnnotation
{
}

20. Valid use cases for annotations

You can do all sorts of things with annotations, but in my experience there are just a few valid use-cases for them.

20.1 Loading configuration

Most often annotations are used as a way to load static configuration for classes. These are some good examples of libraries that use annotations in this way:

In all of these libraries annotations are used as just a particular source of configuration. They all offer other, equivalent ways to accomplish the same thing, for instance loading some XML, Yaml or PHP configuration file. The result of loading configuration from all of these different sources will be used for the job at hand (preparing a database schema, validating or serializing an object, generating routes, etc.). In a way, annotations used for loading configuration are nothing special - they are just one of many ways.

An important characteristic of all these libraries that use annotations as a medium for configuration is that they only parse the relevant annotations once, and combine the results with data retrieved from other sources. The resulting data will be cached in order to make the second run much faster. After the first run, it does not matter anymore that the data originates from annotations, XML files, etc. The data has been combined and unified into one format.

Annotations and coupling

Some people complain that annotations increase class coupling. There are arguments for and against this position: annotations can be considered as “just comments”, so they would definitely not increase class coupling, as class coupling is about how a class itself (not its comments) is coupled to other classes. However, when a doc block is being parsed, annotations are taken to be class names. So in this sense your class becomes coupled to the annotation classes it uses.

This special kind of coupling becomes very obvious when you have a class that has annotations for Doctrine ORM and for the JMS Serializer:

use JMS\Serializer\Annotation as Serialize;
use Doctrine\ORM\Mapping as ORM:

class Something
{
    /**
     * @ORM\Column(type="string")
     * @Serialize\Type("string")
     */
    private $property;
}

Let’s say that the project in which you want to use this class does not have the JMS Serializer as a dependency. But it does have Doctrine ORM and you want to use the mapping metadata to be able to store instances of Something in your relational database.

Well, as soon as Doctrine starts to read the annotations, it will fail, showing us the following error:

[Semantical Error] The annotation "@Serialize\Type" in [...] does not
exist, or could not be auto-loaded.

This is a very obvious sign that the Something class is indeed coupled to all of its annotations. Errors like these are the reason why you should not use annotations in a class when you want it to be reusable. When used outside the project, chances are it will fail because it misses a dependency.

Using annotations in classes that are only used inside one project seems fine to me (as discussed in another part of this book called Configuration conventions. There is no problem with that, since the class doesn’t need to work in a different context.

20.2 Controlling application flow

Instead of being parsed just once, as part of some configuration loading process, annotations can also be used for influencing the flow on an application. The best examples are from the SensioFrameworkExtraBundle which has annotations that control the way the HttpKernel eventually produces a response for a particular request. As explained in the first part of this book there are many points during that process at which event listeners are allowed to change the outcome. The SensioFrameworkExtraBundle contains multiple event listeners which will modify the Request object, create a Response object if none was returned from the controller, or modify a given Response object, based on the annotations that were provided in the doc block of a controller.

For example, you can use the @Template annotation above an action method in a controller class to indicate that the result of executing that action should be used as template variables:

use Sensio\Bundle\FrameworkExtraBundle\Configuration\Template;

class UserController
{
    /**
     * @Template("MatthiasAdminBundle:User:edit.html.twig")
     */
    public function editAction(User $user)
    {
        return array(
            'user' => $user,
        );
    }
}

To make this work the SensioFrameworkExtraBundle registers an event listener which intercepts the kernel.controller event and collects @Template annotations for the current controller. It then tries to figure out which template file should be rendered and stores this file name in a request attribute called _template. When the controller has been executed and the return value is not a proper Response object, another event listener intercepts the kernel.view event and renders the template file mentioned in the request attribute _template using the return value from the controller as template variables.

21. Using annotations in your Symfony application

As a PHP developer you are likely to write some library every once in a while that needs to load configuration from different sources, one of which might be annotations. If that is the case, then I’d like to recommend the jms/metadata package to you. It provides some good tools that help you combine configuration (or “metadata”) from different sources and store it as one object in a file cache.

I’ve written a few articles on my blog about using this library to collect metadata using annotations, to add an alternative driver for collecting metadata, and finally to cache the metadata properly.

In this chapter I choose not to discuss this subject in more detail, since the use of annotations for collecting metadata is not specific to Symfony applications. In fact, any application or library could make use of annotations in this way.

We will instead explore some ways in which you can control the application flow using annotations (the second use case). This subject was also partly covered on my blog some time ago in an article about preventing controller execution using annotations. In the following sections I will describe a variety of other options that you have when combining annotations with kernel events.

21.1 Responding to Request attributes: the @Referrer annotation

Browsers have the useful habit to add a Referer header to a request (unless it is a direct request, like when you type a URL by hand or select one from your bookmarks). The name of the header actually contains a spelling error (it should have one extra “r”: Referrer), but there’s nothing we can do about that, except for making no spelling mistakes ourselves.

The Referer header contains the full URL of the page that was visited by the client before they requested the current URL. While handling a request you can use it to redirect a client back to where they came from. Or you could store the previous URL to later analyze which sites refer to your site.

In this particular example I’d like to use the Referer header to apply certain rules. For instance, some actions inside my application’s controllers should only be executed when the user comes from a certain other page, and some of them are only available when the previous URL has the same domain as the current URL. Though this would not be a good security measure, it can be something that’s nice to have. More importantly it is a good example of how you can influence the application flow based on an attribute of the current Request object.

Based on the scenario I described above, I should be able to write something like this above my actions:

class SomeController
{
    /**
     * @Referrer(pattern="^/demo", sameDomain=true)
     */
    public function specialAction()
    {
        ...
    }
}

This would trigger some kind of validation mechanism which takes the Referer header and checks if its path matches the given pattern and if the referring URL has the same domain as the current URL.

The corresponding annotation class could look something like this:

namespace Matthias\ReferrerBundle\Annotation;

/**
 * @Annotation
 * @Attributes({
 *   @Attribute("pattern", type="string"),
 *   @Attribute("sameDomain", type="boolean")
 * })
 */
class Referrer
{
    public $pattern = '.*';
    public $sameDomain = false;
}

None of the attributes are marked as required. Instead I chose to add default values for the class properties $pattern and $sameDomain.

We want the referrer validation process to be triggered by an annotation that belongs to an action method of a controller class. What would be a good moment to do so? Well, as you may remember from the first part of this book: after the controller has been fully determined and resolved, the kernel dispatches a kernel.controller event. Event listeners are then allowed to do anything that’s needed, given the current controller. This seems like a perfect fit. So in this case we should create an event listener that listens to this particular kernel.controller event. When the listener is notified of such an event, it can inspect the current controller and see if it has a @Referrer annotation.

The ReferrerListener below will do the trick. We’ll make sure that its onKernelController() method will be triggered when a kernel.controller event is dispatched by the kernel. I assume that you know how to register this event listener.

namespace Matthias\ReferrerBundle\EventListener;

use Doctrine\Common\Annotations\Reader;
use Symfony\Component\HttpKernel\Event\FilterControllerEvent;
use Matthias\ReferrerBundle\Annotation\Referrer;
use Symfony\Component\HttpFoundation\Request;

class ReferrerListener
{
    private $annotationReader;

    public function __construct(Reader $annotationReader)
    {
        $this->annotationReader = $annotationReader;
    }

    public function onKernelController(FilterControllerEvent $event)
    {
        // the current controller callable
        $controller = $event->getController();

        if (!is_array($controller)) {
            // we only know how to handle a callable that looks like
            // array($controllerObject, 'nameOfActionMethod')
            return;
        }

        // the annotation reader needs a reflection object like this
        $action = new \ReflectionMethod($controller[0], $controller[1]);

        $referrerAnnotation = $this
            ->annotationReader
            ->getMethodAnnotation(
                $action,
                'Matthias\ReferrerBundle\Annotation\Referrer'
            );

        // $referrerAnnotation is either an instance of Referrer or null
        if (!($referrerAnnotation instanceof Referrer)) {
            return;
        }

        $this->validateReferrer($event->getRequest(), $referrerAnnotation);
    }

    private function validateReferrer(
        Request $request,
        Referrer $configuration
    ) {
        $actualReferrer = $request->headers->get('referer');

        $pattern = $configuration->pattern;
        $sameDomain = $configuration->sameDomain;

        // do anything you like
        // maybe throw an exception if you don't like it
    }
}

Make sure that the service you create for this event listener will receive the annotation_reader service as its first constructor argument.

21.2 Prevent controller execution: the @RequiresCredits annotation

In the following example I want to either prevent or allow execution of a given controller based on the amount of credits the current user has. Imagine that our application has some kind of credit system. The user can buy credits and then they are able to visit certain pay-per-view pages. For example: page A costs 100 credits. When a user has 150 credits, after visiting page A they have only 50 credits left. They will not be able to visit that same page again, unless they buy some extra credits.

I image it to look something like this:

class PayPerViewController
{
    /**
     * @RequiresCredits(100)
     */
    public function expensiveAction()
    {
        ...
    }

    /**
     * @RequiresCredits(50)
     */
    public function cheapAction()
    {
        ...
    }
}

First let’s implement the corresponding annotation class:

namespace Matthias\CreditsBundle\Annotation;

/**
 * @Annotation
 * @Attributes({
 *   @Attribute("credits", type="integer", required=true)
 * })
 */
class RequiresCredits
{
    public $credits;
}

Then, just like in the previous example, we need to listen to the kernel.controller event and analyze the current controller and action, looking for a @RequiresCredits annotation to see if credits are required to execute this action:

namespace Matthias\CreditsBundle\EventListener;

use Matthias\CreditsBundle\Annotation\RequiresCredits;
use Matthias\CreditsBundle\Exception\InsufficientCreditsException;
use Doctrine\Common\Annotations\Reader;
use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

class CreditsListener
{
    private $annotationReader;

    public function __construct(Reader $annotationReader)
    {
        $this->annotationReader = $annotationReader;
    }

    public function onKernelController(FilterControllerEvent $event)
    {
        $controller = $event->getController();

        if (!is_array($controller)) {
            return;
        }

        $action = new \ReflectionMethod($controller[0], $controller[1]);

        $annotation = $this
            ->annotationReader
            ->getMethodAnnotation(
                $action,
                'Matthias\CreditsBundle\Annotation\RequiresCredits'
            );

        if (!($annotation instanceof RequiresCredits)) {
            return;
        }

        $amountOfCreditsRequired = $annotation->credits;

        // somehow determine if the user can afford to call this action
        $userCanAffordThis = ...;

        if (!$userCanAffordThis) {
            // now what?
            ...
        }
    }
}

Of course, the calculations that need to be made to calculate if the user can afford to visit this page should be done by a specialized service, to be injected as a constructor argument.

The question arises: what should we do next to prevent the HttpKernel from executing the controller when the user currently can not afford it. There are several different options here:

  1. You can replace the current controller with another one, for instance a controller that renders a page where the user can buy some extra credits.
  2. You can throw an exception, for instance an InsufficientCreditsException.

If you choose the first option then you should replace the current controller with a valid PHP callable. You could create a new controller on the spot, but it would be better to have it injected as a constructor argument:

class CreditsListener
{
    ...
    private $creditsController;

    public function __construct(
        ...
        CreditsController $creditsController
    ) {
        ...
        $this->creditsController = $creditsController;
    }

    public function onKernelController(FilterControllerEvent $event)
    {
        ...

        // replace the current controller with another one
        $event->setController(
            array(
                $this->creditsController,
                'buyAction'
            )
        );
    }
}

If you choose the second option, you could throw a custom exception, like this:

namespace Matthias\CreditsBundle\Exception;

class InsufficientCreditsException extends \RuntimeException
{
    public function __construct($creditsRequired)
    {
        parent::__construct(
            sprintf(
                'User can not afford to pay %d credits',
                $creditsRequired
            )
        );
    }
}

class CreditsListener
{
    ...

    public function onKernelController(FilterControllerEvent $event)
    {
        ...

        throw new InsufficientCreditsException($annotation->credits);
    }
}

Just throwing an exception would result in the standard error page to be shown. Instead, you might want to customize the response for this particular exception. You can accomplish this by registering an event listener for the kernel.exception event. The event listener would receive a GetResponseForExceptionEvent object. It could then check if the exception that triggered the event is an instance of InsufficientCreditsException and if so, render a proper response for that exception, like a page where the user can buy some extra credits. It would look something like this:

use Symfony\Bundle\FrameworkBundle\Templating\EngineInterface;

class CreditsListener
{
    ...
    private $templating;

    public function __construct(
        ...
        EngineInterface $templating // inject the "templating" service here
    ) {
        ...
        $this->templating = $templating;
    }

    public function onKernelException(GetResponseForExceptionEvent $event)
    {
        $exception = $event->getException();
        if (!($exception instanceof InsufficientCreditsException)) {
            return;
        }

        $response = $this
            ->templating
            ->renderResponse(
                'MatthiasCreditsBundle::insufficientCredits.html.twig',
                array(
                    'requiredCredits' => $exception->getRequiredCredits()
                )
            );

        $event->setResponse($response);
    }
}

Instead of returning a standard error page, the HttpKernel will return the Response object provided by the CreditsListener.

21.3 Modify the response: the @DownloadAs annotation

The last example in this chapter is about modifying the Response object based on a controller annotation.

When you want the result of a controller action to be offered to the client as a downloadable file, you have to do something like this inside your action:

use Symfony\Component\HttpFoundation\Response;
use Symfony\Component\HttpFoundation\ResponseHeaderBag;

class SomeController
{
    public function downloadAction()
    {
        $response = new Response('body of the response');

        $dispositionHeader = $response->headers->makeDisposition(
            ResponseHeaderBag::ATTACHMENT,
            'filename.txt'
        );

        $response
            ->headers
            ->set('Content-Disposition', $dispositionHeader);

        return $response;
    }
}

This is rather verbose and it is not good to have this type of difficult-to-read code inside every action that offers something for download.

Verbosity and repetitiveness in controller actions are often a sign that some of its functionality needs to be moved to a dedicated service. But sometimes it is a hint that the functionality might be pushed to an event listener instead, which does its magic at the right moment during the process of handling a request.

In this case it is safe to assume that creating an event listener would be the right thing to do. It would need to listen to the kernel.response event. This event is dispatched just before the Response object is handed back to the front controller, which is the right moment to turn the response into a downloadable file by adding the Content-Disposition header to it:

namespace Matthias\DownloadBundle\EventListener;

use Symfony\Component\HttpFoundation\ResponseHeaderBag;
use Symfony\Component\HttpKernel\Event\FilterResponseEvent;

class DownloadListener
{
    public function onKernelResponse(FilterResponseEvent $event)
    {
        // we still need to determine the filename
        $downloadAsFilename = ...;

        $response = $event->getResponse();

        $dispositionHeader = $response
            ->headers
            ->makeDisposition(
                ResponseHeaderBag::DISPOSITION_ATTACHMENT,
                $downloadAsFilename
            );

        $response
            ->headers
            ->set('Content-Disposition', $dispositionHeader);
    }
}

If we would register this event listener right now, every response to every request would be offered as a downloadable file. This is not what we want. The response should only be a downloadable file if the controller action is marked as “downloadable”. Let’s do that by introducing a @DownloadAs annotation:

class SomeController
{
    /**
     * @DownloadAs("users.csv")
     */
    public function downloadAction()
    {
        ...
    }
}

We could implement the @DownloadAs annotation as follows:

namespace Matthias\DownloadBundle\Annotation;

/**
 * @Annotation
 * @Attributes({
 *   @Attribute("filename", type="string", required=true)
 * })
 */
class DownloadAs
{
    public $filename;
}

Now we need a way to find out if the action that was executed has a @DownloadAs annotation. Unfortunately, inside the onKernelResponse() method we know nothing about the controller that was used to produce the response. This means that we have to hook into the process earlier, at a moment when we do know the controller. Again, the best thing to do would be to listen to the kernel.controller event and use the annotation reader to find out if it has a @DownloadAs annotation.

We can add another method to the DownloadListener class: onKernelController(). It should be notified of the kernel.controller event. We use the annotation reader to look for the @DownloadAs annotation. If such an annotation was found, we store the suggested filename as an attribute of the current Request object:

use Matthias\DownloadBundle\Annotation\DownloadAs;
use Doctrine\Common\Annotations\Reader;
use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

class DownloadListener
{
    private $annotationReader;

    public function __construct(Reader $annotationReader)
    {
        $this->annotationReader = $annotationReader;
    }

    public function onKernelController(FilterControllerEvent $event)
    {
        // this is more or less the same as in previous examples
        $controller = $event->getController();
        if (!is_array($controller)) {
            return;
        }
        $action = new \ReflectionMethod($controller[0], $controller[1]);

        $annotation = $this
            ->annotationReader
            ->getMethodAnnotation(
                $action,
                'Matthias\DownloadBundle\Annotation\DownloadAs'
            );

        if (!($annotation instanceof DownloadAs)) {
            return;
        }

        // store the filename as a request attribute
        $event->getRequest()->attributes->set(
            '_download_as_filename',
            $annotation->filename
        );
    }

    ...
}

Don’t forget to register this class as a service and register it as an event listener for both the kernel.controller and the kernel.response event.

Now when the controller for a request has been determined, the onKernelController() method of the DownloadListener class looks for a relevant @DownloadAs annotation, copies the filename from it and stores that filename as the _download_as_filename request attribute.

The only thing we need to do is to check inside the onFilterResponse() method if that same attribute has been provided. Only if it is, then the response should be available for download (i.e. have an “attachment” disposition), otherwise it will do nothing and return early:

class DownloadListener
{
    ...

    public function onKernelResponse(FilterResponseEvent $event)
    {
        $downloadAsFilename = $event
            ->getRequest()
            ->attributes
            ->get('_download_as_filename');

        if ($downloadAsFilename === null) {
            // the response should not be a downloadable file
            return;
        }

        $response = $event->getResponse();

        // set the content disposition
        ...
    }
}

22. Designing for reusability

Maybe you have noticed these similarities between the different event listeners that we have seen in the previous sections:

  • All of the event listeners listen to the same event, i.e. kernel.controller. Only some of them listen to other events, like kernel.response.
  • All of the event listeners require the Doctrine annotation reader to fetch annotations for a controller action.
  • All of the event listeners only work when the controller is a callable defined as an array of an object and a method name.

This shared logic could easily be abstracted and this is something we should definitely do since it will make each of the event listeners much cleaner.

Let me just give my solution to this problem (like always there are different ways to accomplish the same thing):

namespace Matthias\ControllerAnnotationsBundle\EventListener;

use Doctrine\Common\Annotations\Reader;
use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

abstract class AbstractControllerAnnotationListener
{
    private $annotationReader;

    /**
     * Return the class of the annotation that should be present for
     * the current controller in order for the processAnnotation() method
     * to be called
     *
     * @return string
     */
    abstract protected function getAnnotationClass();

    /**
     * Will only be called if an annotation of the class returned by
     * getAnnotationClass() was found
     */
    abstract protected function processAnnotation(
        $annotation,
        FilterControllerEvent $event
    );

    public function __construct(Reader $annotationReader)
    {
        $this->annotationReader = $annotationReader;
    }

    public function onKernelController(FilterControllerEvent $event)
    {
        $controller = $event->getController();

        if (!is_array($controller)) {
            return;
        }

        $action = new \ReflectionMethod($controller[0], $controller[1]);

        $annotationClass = $this->getAnnotationClass();

        $annotation = $this
            ->annotationReader
            ->getMethodAnnotation(
                $action,
                $annotationClass
            );

        if (!($annotation instanceof $annotationClass)) {
            return;
        }

        $this->processAnnotation($annotation, $event);
    }
}

You can use this abstract class as the parent class for each event listener that acts upon controller annotations. For instance the ReferrerListener could become a lot cleaner once it extends from AbstractControllerAnnotationListener and correctly implements its abstract methods:

namespace Matthias\ReferrerBundle\EventListener;

use Matthias\ControllerAnnotationsBundle\EventListener\
    AbstractControllerAnnotationListener;

use Symfony\Component\HttpKernel\Event\FilterControllerEvent;

class ReferrerListener extends AbstractControllerAnnotationListener
{
    protected function getAnnotationClass()
    {
        return 'Matthias\ReferrerBundle\Annotation\Referrer';
    }

    protected function processAnnotation(
        $annotation,
        FilterControllerEvent $event
    ) {
        $actualReferrer = $event->getRequest()->headers->get('referer');

        $pattern = $annotation->pattern;
        $sameDomain = $annotation->sameDomain;
    }
}

23. Conclusion

Looking back at the examples above I think we may conclude that the use of annotations inside controllers is a great way to influence the application flow.

Controller annotations allow you to write domain-specific code in a higher-level language (the “annotation language”). The example of the @RequiresCredits annotation made this especially clear. It hides a lot of application logic from the user. Such annotations actually form domain-specific languages on their own. They make it easier for others to understand what is going on and allows them to implement the same thing faster, without copying and pasting complicated PHP code. This helps you prevent code duplication and promote reusability.

So next time you need a way to reuse part of a controller, ask yourself: is this a nice and justifiable use case for introducing a controller annotation? It will not take too much time to create a prototype for it if you use the matthiasnoback/symfony-controller-annotation package.

VIII Being a Symfony developer

Many of the developers that I know call themselves “PHP developer”. Some of them might even say that they are a “Symfony developer”. Others will say, “I am a developer”, without revealing their favorite programming language or the only programming language they master.

When I started to work with Symfony, I almost immediately fell somewhat in love with this great framework. Its first version was already so much better than what I was used to. But the second version was just the right thing for my mind - I started learning a lot about it, digging deep into its code, writing documentation for parts that were still undocumented, writing articles about how to accomplish certain things with Symfony, and speaking in public about Symfony and Symfony-related PHP libraries.

After all of this, I would call myself a Symfony developer now. But the best part of the story is: everything that I learned from and while working with Symfony, is generally applicable to software written “for” any other PHP framework. Even when I work with a non-Symfony or a “legacy” PHP application (with no reused code at all), it still pays off to think about ways in which I can use code from the “Symfony ecosystem”, or in fact, from any library that I can pull in using Composer, and make it a better application.

In this part I will demonstrate that being a good Symfony developer is about knowing the framework well, but then writing code that would be beneficent for any PHP project out there and finishing with a tiny layer between this code and the Symfony framework so that most of your code will be reusable even if it is being transferred to a non-Symfony project.

24. Reusable code has low coupling

24.1 Separate company and product code

Your situation as a Symfony developer is most likely:

  • You are working for a company.
  • You have customers (internal or external) for whom you create web applications.

When you start working on a new application, you will put your project’s code in the /src directory. But before you start, add two new directories inside this /src directory: /src/NameOfYourCompany and /src/NameOfTheProduct. Of course, these directory names should also be reflected in the namespaces of the classes you create for your project.

Whenever you start working on a new feature for the application - think of which part you could in theory reuse, and which part is unique for the application you are working on. Even when this reusable code will in practice never be reused, its quality will benefit from this mindset. Start developing classes in the company namespace. Only when you really feel the need to use something that is specific for the project, move it over to the product namespace, or use some kind of extension principle (a subclass, configuration, event listeners, etc.).

Writing reusable code for your company does not mean that it should be open sourced. It just means that you should write it as if it will be open sourced. You don’t have to create side-projects for all the company code right away. You can develop it inside the project you’re currently working on and maybe later make it ready for reuse in another project. Read more about the practical side of this in Dependency management and version control

24.2 Separate library and bundle code

When you notice that you are writing classes and interfaces that have no relation with the entire Symfony framework, or only with some parts of the framework, you should separate your code into “library” code and “bundle” code. The library code is the part of the code that is more or less stand-alone (although it could have some external dependencies). Library code could be reused by a developer who works with the Zend framework, or with the Silex micro-framework (to name just a few). Bundle code replaces or extends some classes from the library code, adding extra Symfony-specific features. It defines a bundle configuration, and it has service definitions, to make instances of library classes available in the application. A bundle in this way just makes the library code available, while requiring just a minimum effort from developers who want to use the library in their Symfony project.

These are the things that belong inside a bundle:

  • Controllers
  • Container extensions
  • Service definitions
  • Compiler passes
  • Event subscribers
  • Container-aware classes that extend more generic classes
  • Form types
  • Routing configuration
  • Other metadata

The list could be much longer. But it could also be much shorter. When you think about it, only the first couple of things on the list are really specific for bundles (i.e. used only in the context of a standard Symfony project). All of the other things could also be used in projects that only make use of specific Symfony Components. For instance form types could also be used in any PHP project with only the Form Component installed.

24.3 Reduce coupling to the framework

You can go quite far into reducing coupling to the Symfony framework or to one of its components, and thereby making your code even more reusable.

Event listeners over event subscribers

For example you should prefer event listeners over event subscribers. Event subscribers are special event listeners, that implement EventSubscriberInterface:

use Symfony\Component\EventDispatcher\EventSubscriberInterface;

class SendConfirmationMailListener implements EventSubscriberInterface
{
    public static function getSubscribedEvents()
    {
        return array(
            AccountEvents::NEW_ACCOUNT_CREATED => 'onNewAccount'
        );
    }

    public function onNewAccount(AccountEvent $event)
    {
        ...
    }
}

You can register event subscribers like this in a very clean way using the service tag kernel.event_subscriber:

<service id="send_confirmation_mail_listener"
  class="SendConfirmationMailListener">
  <tag name="kernel.event_subscriber" />
</service>

There are some problems with this approach:

  1. An event subscriber becomes totally useless when the Symfony EventDispatcher Component is not available, even though there is nothing Symfony-specific about this event listener.
  2. onNewAccount() receives an AccountEvent object, but it is nowhere determined that such an object can only arise from an event with the name AccountEvents::NEW_ACCOUNT_CREATED.

Therefore an event listener which is actually an event subscriber is not good for reusable code. It couples the code to the EventDispatcher Component. It is better to remove the interface, remove the method required by the interface and register the listener methods manually using the service tag kernel.event_listener:

<service id="send_confirmation_mail_listener"
  class="SendConfirmationMailListener">
  <tag name="kernel.event_listener"
    event="new_account_created" method="onNewAccount" />
</service>

Constructor arguments over fetching container parameters

When you know that the service container has just the parameter you need, it can be tempting to inject the entire container to just retrieve this specific parameter:

use Symfony\Component\DependencyInjection\ContainerInterface;

class SendConfirmationMailListener
{
    private $container;

    public function __construct(ContainerInterface $container)
    {
        $this->container = $container;
    }

    private function sendConfirmationMail()
    {
        $mailFrom = $this->container->getParameter('send_mail_from');

        ...
    }
}

The corresponding service definition would be:

<service id="send_confirmation_mail_listener"
  class="SendConfirmationMailListener">
  <argument type="service" id="service_container" />
</service>

This is obviously very bad for the mobility of this service: it can now only function when there is a Symfony service container available in the application. To make matters worse: even if there is such a service container, it is by no means certain that a parameter send_mail_from would be defined.

Therefore, always inject container parameters as constructor arguments, like this:

<service id="send_confirmation_mail_listener"
  class="SendConfirmationMailListener">
  <argument">%send_mail_from%</argument>
</service>

Constructor arguments over fetching container services

Just like fetching parameters from the service container directly, using the service container as a service locator is an equally bad idea.

class SendConfirmationMailListener
{
    private $container;

    public function __construct(ContainerInterface $container)
    {
        $this->container = $container;
    }

    private function sendConfirmationMail()
    {
        $mailer = $this->container->get('mailer');

        ...
    }
}

The reason for doing things like this is usually performance. Say we would inject the mailer directly as a constructor argument:

class SendConfirmationMailListener
{
    private $mailer;

    public function __construct(\Swift_Mailer $mailer)
    {
        $this->mailer = $mailer;
    }
}

Now, this class is an event listener and it listens to the new_account_created event. Whenever the listener gets instantiated, the mailer service will also be initialized, even when only in some cases a mail would be sent. This may take some time (and at least some system resources).

Though this is indeed true, injecting the service container to fetch a service is bad because:

  1. It couples your code to a very specific type of service container (the Symfony service container).
  2. The silent and possibly wrong assumption is that a mailer service exists (and that it is a \Swift_Mailer object).
  3. Your service could fetch many more things from the service container than it should be allowed to.

The thing is: people get inspired to inject the service container from using the standard Symfony Controller class, which implements the ContainerAwareInterface. You basically should not use this interface nor its basic implementation ContainerAware. There are good use cases, but in most situations you really don’t need the entire service container: you need a specific service. And your class should be coupled to this service’s class or preferably its interface, not to the service container.

The performance issue

But what about the performance issue then? Well, since Symfony 2.3 you don’t need to worry about that anymore. You can add the lazy="true" attribute to service definitions for which you want a proxy class to be created. The result is that you can inject the real dependencies into your services, but they are only fully initialized when you call any of their methods for the first time. I created a small extension for this, the LazyServicesBundle that allows you to make services lazy by pointing out (constructor) arguments that should be turned into lazy-loading services.

Framework-agnostic controllers

In Symfony applications controllers are usually the things that are most tightly coupled to the framework. When you always like to minimize coupling to anything (like I do), controllers seem therefore a good candidate to do so. We should however first consider the things that most controllers do when they are being executed:

  • Take something from the request (either a route parameter, or a query attribute).
  • Fetch some service from the service container and do the thing that was requested.
  • Render a template with the variable retrieved from the service…
  • Or add a flash message to the session, to notify the user of the result of the action…
  • And generate the URL of some other page and redirect the user to it.

All of these things should be done in the way the framework expects you to do them. As you can imagine, this will make your controller very much coupled to the framework. And it should be! A Symfony controller would not function very well in another application which does not use the Symfony Router, the Symfony HttpKernel and the Symfony HttpFoundation components. And such an application I would like to call a “Symfony application”.

So when it comes to controller classes you may skip any steps you would normally take to make a class decoupled. Therefore, don’t be reluctant to extend your controller classes from the standard Symfony Controller class. It will be very helpful since it provides many shortcut methods for things you often do inside a controller (e.g. createForm(), generateUrl(), createNotFoundException(), etc.).

Also, you don’t have to define services for your controllers. They don’t need to be reused as a service. However, I sometimes feel like it is a good idea to be able to take full control over the construction process of a controller, which would otherwise be instantiated “magically” using the new operator by the ControllerResolver, after which - also even more magically - the container will be injected, whenever the controller implements ContainerAwareInterface (read more about this in All the things a controller can be).

One last remark before you start happily coupling your controllers to the framework again: you should still try to make controllers as small as possible. When you enter the controller, you should quickly move the focus away from the controller to some kind of service that does the real work. See also Slim controllers for motivation and practical suggestions.

Thin commands

Writing a console command is very easy and generally recommended to provide ways to execute specific tasks for your bundle using the command line. But commands should be treated the same way as a controller gets treated: it should be a thin layer between the input the user provides, the services that are used to execute the task and the output that the user should receive about successes and failures.

Therefore, ideally, the execute() method of a command contains some checks on the input provided, and then immediately it fetches a service from the service container and calls some action on it.

namespace Matthias\BatchProcessBundle\Command;

use Symfony\Bundle\FrameworkBundle\Command\ContainerAwareCommand;
use Symfony\Component\Console\Input\InputInterface;
use Symfony\Component\Console\Output\OutputInterface;

class BatchProcessCommand extends ContainerAwareCommand
{
    protected function configure()
    {
        $this->setName('matthias:batch-process');
    }

    protected function execute(
        InputInterface $input,
        OutputInterface $output
    ) {
        $processor = $this->getContainer()->get('batch_processor');

        foreach (array('main', 'sub', 'special') as $collection) {
            $processor->process($collection);
        }
    }
}

When you feel the need to do more things inside a command, move as many responsibilities as you can to specific services.

This again makes it easier to reuse command code, even with an entirely different library for executing commands. It also helps you to test the code that really “does things” better, since testing code inside a command is much more difficult and much less isolated than testing code in normal PHP classes.

Moving code away from the command class will make it more difficult to generate relevant output for the console. There is a perfect solution for this: use a bit of aspect oriented programming to generate output based on the actual method calls and their return values.

The environment

Symfony has a very nice way of separating and cascading bundle configuration for different environments. These environments are called (by convention) dev, test and prod. This value is however entirely arbitrary. It is not even a class constant somewhere, but it’s just the first argument used when instantiating the kernel in the front controllers in /web:

// in /web/app_dev.php
$kernel = new AppKernel('dev', true);

You should therefore never rely on the environment name or hard-code it inside any part of your code. Things like this must be avoided:

class SendConfirmationMailListener
{
    private $environment;

    public function __construct($environment)
    {
        $this->environment = $environment;
    }

    private function sendConfirmationMail()
    {
        if ($this->environment === 'dev') {
            // always send the mail to matthiasnoback@gmail.com :)
        }
    }
}

Any modification of behavior depending on the environment of the application should take place on the level of your bundle’s configuration and even then it should be implicit, since which environment names a developer uses is up to him, not up to you. In fact, when you copy everything from config_dev.yml to config_prod.yml your production environment will look and feel just like a development environment.

25. Reusable code should be mobile

The main characteristic of mobile code is that you should be able to take the code and easily transfer it to another application. After installing the code in another project, you know that it is mobile when:

  • You can easily fix bugs in the code, even though the code has been duplicated many times.
  • It does not fail because of another library, component, bundle or PHP extension that is not present in the other project.
  • You can easily configure it to work with another database and maybe even another type of database.
  • It does not fail because certain database tables or collections don’t exist.
  • It does not fail because it is placed in an unusual directory inside the project.

25.1 Dependency management and version control

The first two characteristics of mobile code are related to infrastructure: when reused, your code should not merely be duplicated - it should still be under version control so that each bug-fix can be easily distributed over the different copies of the code.

You can do this using a mechanism like externals, or sub-modules (depending on the type of version control software that you use), but the better way is to use Composer. Each piece of code that you want to share between projects should be in its own repository. In case you want to keep the code for yourself (or for your company): host the repository privately, on GitHub, Bitbucket, etc. When you want to make your code open source, GitHub would be the currently fashionable place.

In both cases you need a composer.json file. This file will basically make your code known as something coherent: a package. In composer.json you can define the requirements (PHP version, PHP extensions and other libraries), you can instruct the autoloader on where to find classes and functions, and you can add some information about yourself, the package, and the tags it should have.

As an example: this is a somewhat modified composer.json from the knplabs/gaufrette library:

{
  "name":         "knplabs/gaufrette",
  "type":         "library",
  "description":  "Library that provides a filesystem abstraction layer",
  "keywords":     ["file", "filesystem", "media", "abstraction"],
  "minimum-stability": "dev",
  "homepage":     "http://knplabs.com",
  "license":      "MIT",
  "authors": [
    {
      "name": "KnpLabs Team",
      "homepage": "http://knplabs.com"
    }
  ],
  "require": {
    "php": ">=5.3.2"
  },
  "require-dev": {
    "amazonwebservices/aws-sdk-for-php": "1.5.*",
    "phpspec/phpspec2": "dev-master",
    "rackspace/php-cloudfiles": "*",
    "doctrine/dbal": ">=2.3",
    "dropbox-php/dropbox-php": "*",
    "herzult/php-ssh": "*",
    "phpunit/phpunit": "3.7.*"
  },
  "suggest": {
    "knplabs/knp-gaufrette-bundle": "*",
    "dropbox-php/dropbox-php": "to use the Dropbox adapter",
    "amazonwebservices/aws-sdk-for-php": "to use the Amazon S3 adapter",
    "doctrine/dbal": "to use the Doctrine DBAL adapter",
    "ext-zip": "to use the Zip adapter"
  },
  "autoload": {
    "psr-0": { "Gaufrette": "src/" }
  }
}

The list of requirements should be exhaustive for all the core functionality of your package. When the package contains classes that can be used optionally, which require extra dependencies, mention them under suggest. When you have written tests for these classes (like you should) make sure to list these extra dependencies together with your testing framework under require-dev so that all tests can be executed by anyone and no test will fail or get skipped because of a missing dependency.

Package repositories

Whenever a repository has a composer.json in its root directory, you can submit it as a package on Packagist. This will effectively make your code open source. When you don’t want this to happen, you can also privately host a package repository using Satis. This works great for sharing code over many projects in the same company.

After creating the composer.json based on what your code really needs, and registering it as a package, either using Packagist or Satis, you can install the package in any of your projects by running:

composer.phar require [name-of-package] 0.1.*

Now, your code also complies with the second characteristic of mobile code: after installing it, it will not fail because of missing or incorrect dependencies.

25.2 Hard-coded storage layer

Auto-mapped entities

One of the biggest problems with bundles out there is that some of them define entity classes (using for instance annotations) and then put them in the Entity directory inside the bundle. This automatically makes these specific entities available in your project. This may seem nice, but in most situations it is not: you are then not able to choose your own schema, add extra fields, or maybe leave out the entities entirely and redefine the model using, for instance, MongoDB.

When a bundle enforces a specific storage layer like described above, you will notice this as soon as you run a command like doctrine:schema:update: suddenly all kinds of tables are being created, beyond your control basically, since by enabling the bundle in which there are defined, these will automatically be registered. The same problem arises when you run doctrine:migrations:diff. Then a migration script will be generated for creating these tables that you never wanted in the first place.

Storage-agnostic models

The right way to define your bundle’s model is to provide base classes. These base classes contain all the usual getters, setters and other methods for modifying or inspecting the state of an object. They have protected properties (instead of the usual private properties) with no mapping-related annotations at all. These storage-agnostic model classes should reside in the Model directory of your bundle, and they should not contain any code which is specific for their manager (like an EntityManager or a DocumentManager).

namespace Matthias\MessageBundle\Model;

class Message
{
    protected $body;

    public function getBody()
    {
        return $this->body;
    }
}

This allows developers who want to use your bundle to extend from the base class and define any metadata required for mapping the data to some kind of database in their own way:

namespace Matthias\ProjectBundle\Entity;

use Matthias\MessageBundle\Model\Message as BaseMessage;
use Doctrine\ORM\Mapping as ORM;

/**
 * @ORM\Entity
 */
class Message extends BaseMessage
{
    /**
     * @ORM\Column(type="text")
     */
    protected $body;

    // Application specific data:

    /**
     * @ORM\Column(type="integer")
     */
    protected $upVotes;

    /**
     * @ORM\Column(type="integer")
     */
    protected $downVotes;
}

As you can see, this will also allow someone to add extra fields which are not defined in the reusable, generic model. Make sure your bundle has a configuration option for setting the user-defined model class:

matthias_message:
    message_class: Matthias\ProjectBundle\Entity\Message

Object managers

Even when you have defined a storage-agnostic model, it does not mean that your entire bundle needs to be storage-agnostic. The only thing you need to take care of is that the users of your bundle should be able to implement their own storage handling and disable the standard way that you implemented. You can do this by applying the strategy pattern for loading exclusive services, and by making use of aliases.

Also, you should not forget to make the names of entity managers, document managers, etc. configurable. For instance Doctrine ORM allows users to define different entity managers for different entity classes. So even though the default entity manager is the right entity manager in most cases, you should add a configuration key for it in your bundle’s configuration.

25.3 Hard-coded filesystem references

Make sure your bundle does not contain any hard-coded filesystem references outside* the bundle itself. You may assume that anything starting from the root of your bundle is in your control, but the location of the vendor directory, or the web directory, or any other directory should be considered unknown. In fact, they are: when it comes to the directory structure, nothing is fixed when using Symfony. There are only “sensible defaults”.

Any reference you want to make to any file above ./ should be configurable in config.yml as part of your bundle’s configuration. You may of course offer sensible defaults, so your bundle works out-of-the-box when installed in a standard Symfony project. The standard way to accomplish this is to define a scalar node in your configuration tree, with a default value. Then process the configuration in the load() method of your bundle’s extension class and define a parameter for it. This parameter can later be injected, for instance as a constructor argument. See also Define parameters in a container extension.

Using the filesystem

When your bundle requires a location other than the database to store data, like uploaded files, consider making the bundle filesystem-independent. Since not every application has write-access to the hard disk of the server it is hosted on, you may choose to use a filesystem abstraction layer, like Gaufrette. It has adapters for many kinds of filesystems (local or remote), and your services won’t notice anything when you switch between them. There is also a GaufretteBundle which further eases integration of the library with your Symfony bundles.

26. Reusable code should be open for extension

Code that is reusable has to be flexible. A user of your bundle should be able to configure services in a different way than they were configured in the original project you developed it for. They should also be able to replace any service of the bundle with a custom implementation, without the entire system falling apart. In other words, reusable code should adhere to the principle “open for extension, closed for modification”. The user should not be forced to change the original code, but only to replace or extend specific parts in order to better support their needs.

26.1 Configurable behavior

To make your bundle flexible, it should have a rich set of configurable options, with sensible defaults and information about what a certain option means. Only the services that are requested based on these configuration values provided by the developer in for instance config.yml should be fully loaded. See also the part of this book about Patterns of dependency injection. It contains many suggestions on how to make a bundle configurable.

26.2 Everything should be replaceable

By using a compiler pass (see also Patterns of dependency injection) it is possible to replace any previously created service definition with your own service definition, or to change the class or any argument of existing service definitions. This means that when a developer wants to replace part of the functionality of your bundle with their own, there will in theory always be a way to accomplish this. There are still some things that you have to do, to help developers with replacing just what they need to replace, while keeping other parts of the bundle unchanged.

Use lots of interfaces

Define lots of interfaces for the classes that you use, both in your bundle and in the corresponding library code. Interfaces are literally the contracts that are signed between one object and another. A class that implements an interface says: don’t worry about how I do things exactly, this is how you can talk to me.

A traditional example, using only classes:

class Translator
{
    public function trans($id, array $parameters = array())
    {
        ...
    }

    ... lots of other methods
}

class TranslationExtension
{
    private $translator;

    public function __construct(Translator $translator)
    {
        $this->translator = $translator;
    }

    public function trans($id, array $parameters = array())
    {
        return $this->translator->trans($id, $parameters);
    }
}

It has this service definition:

<service id="translator" class="...">
</service>

<service id="twig.extension.trans" class="...">
    <argument type="service" id="translator" />
</service>

When a developer wants to replace the translator, they need to extend from the existing Translator class to satisfy the constructor argument type-hinted Translator:

class MyTranslator extends Translator
{
    ...
}
<service id="my_translator" class="MyTranslator">
</service>

But now MyTranslator inherits everything that is already in the Translator class, even though it is going to do things very differently.

A much better solution would be to define an interface for translators:

interface TranslatorInterface
{
    public function trans($id, array $parameters = array());
}

The only thing a replacement translator has to do, is implement this interface:

class MyTranslator implements TranslatorInterface
{
    public function trans($id, array $parameters = array())
    {
        // do things very differently
        ...
    }
}

Finally, any existing type-hints should be changed from Translator to TranslatorInterface:

class TranslationExtension
{
    public function __construct(TranslatorInterface $translator)
    {
        ...
    }
}

And now nothing stands in the way of replacing the existing translator service with the my_translator service, either removing the existing definition, adding a new definition, by changing the class of the existing definition or by defining an alias from translator to my_translator.

Use the bundle configuration to replace services

As I mentioned before, you could in theory replace everything there is in the service container by something you created yourself. There are many ways to do this. But they are all not so clean and they won’t be good for maintainability. The best way I’ve seen so far is to allow other developers to replace specific services, by providing service ids through the bundle configuration:

# in /app/config/config.yml
matthias_message:
    # point to a specific service that should be used
    message_domain_manager: matthias_project.message_domain_manager

26.3 Add extension points

You can make behavior configurable, or allow others to replace parts of your bundle. But you can also add ways to extend the behavior of your bundle.

Service tags

One way to allow other bundles in an application to extend the behavior of your bundle, is to make use of service tags. You could think of the way you can register form types that can be used in the entire application:

<service id="address_type" class="...">
  <tag name="form.type" alias="address" />
</service>

Using a compiler pass you can find all services with a given tag and do whatever you want with them. See also Service tags for implementation details and more examples.

Events

Events are great for allowing other parts of an application to hook into the execution of your bundle’s code. You could use events to simply notify the system that something has happened:

$eventDispatcher->dispatch(
    'matthias_account.account_created',
    new AccountEvent($account)
);

Or you could use an event object to let other parts of the application modify (filter) some value, like the Symfony kernel itself does, when it dispatches its kernel events (see also Early response).

27. Reusable code should be easy to use

27.1 Add documentation

The Symfony Cookbook article about best practices for creating bundles recommends writing documentation for bundles. The proposed format is Re-Structured Text and the main entry point should be Resources/doc/index.rst. However, many PHP projects I know of also have a README.md file (which is in the Markdown format), describing in a few words what the project is about and how to install it.

Documentation should cover at least:

  • The main concepts in the design of your code.
  • Use cases, illustrated with code and configuration examples.
  • Service tags that can be used to extend functionality.
  • Events that are being dispatched.
  • Any flow-controlling exceptions that can be thrown.

27.2 Throw helpful exceptions

Chances are that someone trying to use your bundle will encounter some problems. Maybe you forgot to make some requirements explicit, maybe you made some other assumptions about how and in what situation the bundle’s services would be used. In order to help your “user” to overcome any problem in this area, you should throw helpful exceptions, which point him in the direction of the solution.

Use specific exception classes

Choose your exception classes well, and preferably add some exception classes yourself:

namespace Matthias\ImportBundle\Exception;

class WrongFormatException extends \RuntimeException
{
}

This allows someone to catch exceptions from which the application may still be able to recover:

try {
    $this->loadFile($file);
} catch (WrongFormatException $exception) {
    // ...
} catch (FileNotFoundException $exception) {
    // ...
}

Set detailed and friendly messages

There is a common misconception that you should not put all the relevant details in the message of an exception. This would be bad for “security”. Well, since you should prepare your production environment to display no error messages at all, there’s no problem in giving full disclosure in an exception message. On the contrary: you should provide all relevant details. So, use sprintf() to construct a nice and friendly message:

throw new \InvalidArgumentException(sprintf(
    'Tag "%s" for service "%s" should have an "alias" attribute',
    $tagName,
    $serviceId
));

In contrast with things like:

throw new \OutOfBoundsException('You are a hacker');

Which would communicate very clearly your opinion about the user, but not help a developer in finding out what they can do to prevent the error.

28. Reusable code should be reliable

28.1 Add enough tests

When your code should be stable and maintainable at any rate, it needs enough tests. What is enough? Maybe you don’t need a 100 percent coverage, but at least the use cases that you say your code supports should be verified to work correctly using unit tests.

Writing the tests will not be such a difficult task, since when you followed my advice in the previous section, your bundle consists of small classes, and many of them have interfaces, so they should already be very test-friendly.

After you have tested the true units of your bundle, all the little classes in there, it is time to test them all together. In other words: write integration tests. In my bundles, this means most of the times manually instantiating some classes, prepare some constructor arguments and create some more objects, and finally run a single method on one of these objects (while keeping track of which is which). This step is important, as you want to find and fix any problems without going back to the browser after many hours of fast and furious coding in your favorite editor.

Test your bundle extension and configuration

Something that is missing in the test suites of many bundles: unit tests for framework-specific classes like the Bundle class and, when applicable its Extension and Configuration class. You can usually skip the Bundle class, but the Extension and Configuration classes have logic that should be tested. After all, you want to make sure that all things inside your bundle are tied together correctly.

To test the bundle’s extension class, you only need a ContainerBuilder instance. Then run the load() method of the extension and provide it with an array of config arrays:

namespace Matthias\AccountBundle\Tests\DependencyInjection;

use Symfony\Component\DependencyInjection\ContainerBuilder;
use Matthias\AccountBundle\DependencyInjection\MatthiasAccountExtension;

class MatthiasAccountExtensionTest extends \PHPUnit_Framework_TestCase
{
    public function testLoadsDoctrineORMServicesWhenEnabled()
    {
        $container = new ContainerBuilder();
        $extension = new MatthiasAccountExtension();

        $configs = array(
            array(
                'storage_engine' => 'doctrine_orm',
            )
        );
        $extension->load($configs, $container);

        $this->assertTrue(
            $container->has('matthias_account.doctrine_orm.storage')
        );
    }
}

Testing the bundle configuration is mainly an exercise in testing the Config Component itself. But since the code in the Configuration class is itself configuration (think one second longer about that…), you need to make sure that this configuration is sound. This is how you can do it:

namespace Matthias\AccountBundle\Tests\DependencyInjection;

use Matthias\AccountBundle\DependencyInjection\Configuration;
use Symfony\Component\Config\Definition\Processor;

class ConfigurationTest extends \PHPUnit_Framework_TestCase
{
    public function testHasSensibleDefaults()
    {
        $configurations = array();
        $processedConfig = $this->processConfiguration($configurations);

        $expectedConfiguration = array(
            'user_types' => array('user', 'administrator'),
            'send_confirmation_mail' => true
        );

        $this->assertSame($expectedConfiguration, $processedConfig);
    }

    private function processConfiguration(array $configValues)
    {
        $configuration = new Configuration();

        $processor = new Processor();

        return $processor->processConfiguration(
            $configuration,
            $configValues
        );
    }
}

Conclusion

Though not from request to response, this has been quite a journey! I wrote the first pages of this book in April of 2013. In June I said it would take just 6 more weeks to finish it. Well, it took three months, but now it is exactly as I had imagined it to be: a book for developers familiar with Symfony, wanting to take an extra step. Now that you’ve finished the book, I hope that you found it interesting. And maybe you can use some of the practices described here in your work as a developer.

Since this is the last page of the book, it’s time for some meta-comments!

I think that the PHP community has taken some huge steps in the last couple of years. There are some great tools and platforms now, like GitHub and Composer, Vagrant and Puppet. PHP developers are becoming aware of the commonality of their individual ideas (for instance as expressed by the PSR-* standards). They have also slowly started to take care of handling version numbers well (e.g. semantic versioning). And they are generally very much concerned about bugs in their open-sourced software (which other people use in production environments). Above all, many PHP developers contribute to open-source projects for free, which generates a good vibe and a great tradition of giving and sharing.

But PHP developers are not used to all of the activities related to working with or even delivering open-source software: defining a manageable list of package dependencies, distributing packages, releasing software to the world, announcing changes to an API, handling bugs reported by strangers, working with continuous integration, etc.

More importantly: developing software for reusability has proven to be quite a difficult task in itself. Luckily, there are many principles you can adhere to. Some of them are described in the last part of this book. And some of them are well known to people writing in other languages than PHP. See for instance the list of package design principles as described by Robert Martin. Ideas like these are not known to everyone in the PHP community, while they should be. If I were to ever write a book again, it would be about that: PHP package design.