But first, some background is probably in order, as this is a jargon-heavy post.

Aspect Oriented Programming

I was first introduced to AOP in 2006 via an April 2006 php|architect article by Dmitry Sheiko. That article detailed adding calls at various places in a method where you might want to hook into functionality -- for instance, to log, cache, etc. Expanding on this, I considered other possibilities: manipulating incoming arguments, validation, ACL checks, implementing write-through caching strategies, and more. The approach is daunting, however; typical, naive implementations lead to a lot of boiler-plate code:

interface Listener
{
    public function notify($signal, $argv = null);
}

class Foo
{
    protected $listeners;

    public function attach($signal, Listener $listener)
    {
        $this->listeners[$signal][] = $listener;
    }

    public function doSomething($arg1, $arg2)
    {
        foreach ($this->listeners as $listener) {
            $listener->notify('preDoSomething', func_get_args());
        }
        
        // do some work
        
        foreach ($this->listeners as $listener) {
            $listener->notify('postDoSomething', $result);
        }
    }
}

The article didn't go into any real details on how you might short-circuit the filters or handle return values from them, and aspect handling itself was not detailed completely. As such, the code begins to add up, particularly if many classes and/or methods implement the functionality.

Intercepting Filters

A similar concept to AOP is the idea of Intercepting Filters. Like AOP, the idea is to separate cross-cutting concerns such as logging, debugging, and more from the actual logic the component exposes. The difference is that typically Intercepting Filters are language independent, have a standard implementation in a given framework, and can be re-used. The approach gwoo used in his post falls more under this category.

Lithium, a PHP 5.3 framework and the reference for gwoo's article, has a very intriguing approach. Instead of calling the filters explicitly within the body of the code, they suggest that the body of the code simply becomes one of the filters, via a closure:

Dispatcher::applyFilter('run', function ($self, $params, $chain) {
    // do something...
    return $chain->next($self, $params, $chain);
});

In Lithium, each filter is responsible for calling the next (each filter receives the chain as its third and final argument); as soon as one doesn't call next(), execution is stopped, and the result returned (or at least that's how I read the source). You can call the chain either before or after the code you want to execute in each filter; placement will determine whether it's a pre- or a post-filter. The approach answers a number of the concerns I outlined previously -- namely, standardization of approach, and the ability to short-circuit execution.

The above example defines a filter that will run when the run() method of the Dispatcher class is executed. $self will typically be the object instance, $params an array of the parameters passed to the method, and $chain as described above. The method itself will execute any filters -- typically with something like this:

use lithium/core/Object as BaseObject;

class Foo extends BaseObject
{
    public function doSomething($with, $these, $args)
    {
        $params = compact('with', 'these', 'args');
        
        return $this->_filter(__METHOD__, $params, function ($self, $params) {
            // do the actual work here
            return $result;
        });
    }
}

(The _filter() method is defined in lithium\core\Object, and basically passes a local, static chain of filters to Lithium's Filter class for execution. applyFilter() from the previous example statically adds a callback under the named method to the chain.)

This solution is elegant -- but I see some limitations:

• First and foremost, I'm not particularly fond of the filtering functionality being via static methods on a single class; it introduces a hard-coded, hidden dependency. This means you cannot provide alternate filtering functionality without extending the class consuming the filters, nor without either extending the base filters implementation should you wish to provide a compatible API (for instance, to introduce an implementation that understands priorities).
  
  Additionally, the easiest way to implement filtering in Lithium is by extending the lithium\core\Object class -- I could find no examples elsewhere in the documentation that showed how you would compose the Filters implementation in your own objects. As such, the easiest way to compose filters is now via inheritance, which seems to be counter-productive to the whole rationale behind filtering, to my thinking.
• Second, the approach of making the body of the calling method a closure makes it difficult to create non-public helper methods. Inside the filter, you're no longer in the scope of the object, losing the semantics that tie the various metadata and functionality of the object together. (The Lithium docs provide illustrations of how to accomplish this, but they require extra work, and a keen understanding of how references work in PHP.)
• Third, it's sometimes useful to have access to the return results of all the filters (not just the last executed); you may want to aggregate them in some way, or branch logic based on the various returns.
• Fourth, it's sometimes useful to have multiple call points within the main code. As an example, for many caching strategies, you'd check first to see if you have a cache hit, and return immediatly if found; otherwise, you'd execute the code, and cache the result prior to returning it. This might be possible in Lithium with constructs like this:

  
  Filters::add('SomeClass::doSomething', function ($method, $self, $params) {
      if (null !== ($content = cache_hit($params))) {
          return $content;
      }
      $content = Filters::next($method, $self, $params);
      return $content;
  });
  

  However, if you have several filters such as this, the order then becomes paramount, and that introduces new complexities.
  
  Another example would be with façade methods, where you may wish to introduce filters before and after each method call:

  
      public function doSomeWorkflow($message)
      {
          $this->somePrivateMethod($message);
          $this->nextPrivateMethod($message);
          $this->lastPrivateMethod($message);
      }
  

  
  
  (I can already hear Nate saying "make those all filters!" or "filter each method!" -- but that's the problem with simple examples - they can't always express the nuances of a use case.)
• Fifth, it's useful to be able to attach callbacks that are not aware of the chain. For instance, you may have code you've already written that works perfectly fine in a standalone situation -- e.g., a logger -- and you simply want to add it to the chain. In the Lithium paradigm, you'd need to curry the calls in, instead of simply using the existing method:

  
  // This:
  SomeClass::applyFilter('doSomething', function ($self, $params, $chain) use ($log) {
      $log->info($params['message'];
      $chain->next($self, $params, $chain);
  });
  // VS:
  SomeClass::signals()->connect('doSomething', $log, 'info');
  

  Related to this, I personally dislike aggregating the filter parameters into a single associative array. I don't like having to test for the existence of parameters, and would much rather PHP tell me if I'm missing required parameters or if any fail typehints. That said, doing so provides a consistent API when filtering.

All in all, however, the approach Lithium provides is very good; it just doesn't completely suit my tastes or use cases.

Signal Slots

Interestingly, the capabilities I need are not far from what Lithium provides -- in fact, I'd argue that the Intercepting Filters of Lithium are actually probably more akin to another pattern, Signal Slots.

With Signal Slots, your code emits signals (Lithium does this -- it emits the name of the method being called); any handler, or slot (filters in Lithium), connected to the signal is then executed.

As such, you typically have some sort of signal "manager" object (the Filters class in Lithium) that aggregates signals and attached slots; this manager is then composed into the object emitting signals. For those of you familiar with events in JavaScript or other event-driven languages, this should sound quite familiar.

Such an approach looks like this:

class Foo
{
    protected $signals;

    public function signals(SignalSlot $signals = null)
    {
        if (null === $signals) {
            // No argument? make sure we have a signal manager
            if (null === $this->signals) {
                $this->signals = new Signals(); // SignalSlot implementation
            }
        } else {
            // Compose in an instance of a signal manager
            $this->signals = $signals;
        }
        return $this->signals;
    }

    public function doSomething($with, $these, $args)
    {
        $this->signals()->emit('doSomething.pre', $this, $with, $these, $args);
        
        // do some work
        $this->signals()->emit('doSomething.during', $this, $with, $these, $args);

        // do some more work
        // This time, pass the result
        $this->signals()->emit('doSomething.post', $this, $result, $with, $these, $args);
        return $result;
    }
}

$f = new Foo();
$f->signals()->connect('doSomething.pre', $log, 'info');
$f->signals()->connect('doSomething.during', $validator, 'isValid');
$f->signals()->connect('doSomething.post', $indexer, 'index');

Basically, a SignalSlot provides an object in which signals and their attached slots are aggregated. This allows having a single manager for multiple signals (which is similar to how Lithium's Filters class works), while also providing a way to emit multiple signals from a single procedure. Additionally, since it is simply an object, you can compose it in to classes that may emit signals -- without requiring inheritance.

This is the basic approach of the ZF2 SignalSlot implementation, as well as that found in Symfony 2's Event Dispatcher and Zeta Components' SignalSlot component.

Both Symfony 2's Event Dispatcher and ZF2's SignalSlot component build in the ability to short-circuit, Symfony via a notifyUntil() method, and ZF2 via an emitUntil method. With ZF2, each time a signal is emitted, a ResponseCollection is returned by the manager, containing an aggregate of all slot responses. Calling emitUntil() will short-circuit execution of remaining slots if a given slot returns a response that validates against given criteria; at this point, the collection is marked as "stopped", and you can pull the "last" response and return it:

$responses = $this->signals()->emitUntil(function($response) {
    return ($response instanceof SpecificResultType);
}, 'doSomething.pre', $this, $with, $these, $args);
if ($responses->stopped()) {
    return $responses->last();
}

This introduces extra code in the method emitting the signals -- but meets the criteria that no given slot need be aware of the chain.

The Signal Slot approach actually supports paradigms similar to those illustrated in Lithium. For instance, I can make my method body a slot:

class Foo
{
    protected $handlers = array();

    // ... skip signals composition ...
    
    public function doSomething($with, $these, $args)
    {
        $params = compact('with', 'these', 'args');
        
        // connect() returns a signal handler (slot); store it so that we only
        // ever attach it once...
        if (isset($this->handlers[__FUNCTION__])) {
            $this->handlers[__FUNCTION__] = $this->signals()->connect(__FUNCTION__, function($self, $params) {
                // do the work here!
            });
        }
        
        // Emit the signal, and return the last result
        return $this->signals()->emit(__FUNCTION__, $this, $params)->last();
    }
}

Concerns

Using Signal Slots and Intercepting Filters is not without its concerns, nor is any given implementation perfect.

• Zeta Components does a fantastic job of handling signal slots. However, you cannot short-circuit execution, nor introspect return values. It does offer two features neither ZF2 nor Symfony 2 offer (at this time): the ability to statically connect slots to signals, allowing you to wire without having an existing instance, nor even caring what object might emit the signal; and the ability to add a priority to slots, which allows you to alter the execution order.
• Lithium does a nice job of providing good standards (signals are method names; parameters are predictable for all handlers), but at the price of some flexibility (static implementation with no interface for alternate implementations; no ability to re-use existing methods and functions with differing signatures without currying).
• Symfony 2 offers short-ciruiting and flexibility in callbacks, but requires that you create an event object to pass to the Event Dispatcher, making the usage slightly more verbose, and offers no standardization of signal naming.
• ZF2's SignalSlots offer similar benefits (and drawbacks) to Symfony 2's implementation, provides standardization of the signal manager response, allows registering classes that self-register with the signal handler, but lacks static wiring capabilities or prioritization.

On a more abstract level, signal slots and intercepting filters can lead to difficulties in learning and mastering code that use them:

• How are signals named?
• How do you document the parameters available to slots/filters?
• How can those using IDEs discover available signals? and the arguments expected?
• Where does the wiring occur?
• For instance, if any wiring is automated, this can potentially lead to more difficulty in debugging.
• If done manually, when, and where?
• What happens if a slot doesn't receive arguments it needs, or cannot handle the arguments it receives?

In short, while they solve many problems, the implementations also introduce new concerns -- though this will be true of any extension system, in my experience.

Conclusions

I personally am a huge fan of Intercepting Filters and Signal Slots. I think they can make code easier to extend, by providing a standard methodology for introducing cross-cutting concerns without requiring class extension. They can also make code quite expressive -- sometimes at the cost of readability -- by introducing functional programming paradigms.

If you have not investigated these concepts or components before, I highly recommend doing so; I think they play a fundamental role in the next generation of PHP frameworks.

Caveats

I am not an expert, nor well-versed, in all the frameworks listed here, and as such, some of the information may be incorrect or incomplete. I am the author of ZF2's current Signal Slot implementation, and am still working on improvements to it.

Updates

• 2011-01-10 11:35Z-05:00 Cal Evans found the original php|architect article I referenced, and I've revised some of the assessments based on re-reading it, as well as linked to the issue.

To be honest, I was a little worried about the conference -- many see Symfony and ZF as being in competition, and that there would be no cross-pollination. I'm hoping that between Fabien, Stefan, and myself, we helped dispel that myth this week.

The fact of the matter is that no single project can be fully comprehensive, and do everything perfectly. In my examinations of different frameworks, PHP and otherwise, the places where they most differ and which generates the most loyalty amongst users are the MVC approaches and tooling support. In good frameworks, this is just a portion of the code, and the remainder is in support libraries or plugins that extend that functionality.

This is true of both Symfony and Zend Framework. Symfony's development team has chosen to focus on a very specific core of functionality related to the MVC approach, which makes their maintenance job easier, and leads to a stable product. Zend Framework's MVC implementation is offered as a group of separate components, with components such as Zend_Application and Zend_Tool helping to bring cohesion and structure to them.

What this means is that once you've developed the basic infrastructure of your application, the scaffolding, you're now left with decisions about how to implement the actual functionality of the application itself. The problem as I see it is: how do you do that development? Many developers are myopic and will not look beyond the framework they have chosen for for development. This can lead to multiple implementations of the same code, and often leads to incomplete implementations as well.

My feeling is that whenever you find yourself about to write new code, look to see if somebody else has written the code already. Anybody -- don't limit yourself to your framework of choice. If I want to do serious HTML sniffing, validation, and cleanup, I go to HTMLPurifier; if I want a workflow component, I check out eZ Components Workflow; I always check PEAR.

This week, I tried to spread this message within the Symfony community, showing them how easy it is to integrate ZF components within Symfony projects. The integration itself is simple: instantiate the Zend autoloader, and start using ZF classes. This same technique can be used to load PEAR, or eZComponents, or Doctrine 2, etc. The trick is getting out of the "Not Invented Here" syndrome, letting go of your ego, and using other people's code.

(Yes, I know we have code in ZF duplicating functionality in other libraries; in most cases, we try and offer at least a new approach to the problem -- but we could do better.)

Fabien also made an interesting announcement. During a Q&A session with the Symfony core team, he said that Symfony 2 will not write re-invent the wheel when it doesn't need to -- and announced that Symfony 2 will be using Zend_Log and Zend_Cache instead of rewriting the current Symfony components. I find this admirable -- and it's something I'm hoping to do in a few places with Zend Framework 2.0 as well, as I know there are features and code that others have, quite simply, written better.

One last note in this ramble: With the various "2.0" versions of frameworks, most projects are learning from both mistakes made as well as from the usage patters of the developers adopting them. One of those lessons, to my mind, is that no one framework can do it all well and by themselves. I fully expect to see the next generation of frameworks making it trivial to pull features from other frameworks and libraries in order to fill out functionality.