One very important (yet often overlooked) design guideline which I advocate is the Principle of least knowledge.

The Principle of Least knowledge, also known as The law of Demeter, or more precisely, the Law of Demeter for Functions/Methods (LoD-F) is a design principle which provides guidelines for designing a system with minimal dependencies. It is typically summarized as “Only talk to your immediate friends.”

What this means is a client should only have knowledge of an objects members, and not have access to properties and methods of other objects via the members. To put it in simple terms you should only have access to the members of the object, and nothing beyond that. Think if it like this: if you use more than 1 dot you are violating the principle.

Consider the following: We have three classes: ClassA, ClassB and ClassC. ClassA has an instance member of type ClassB. ClassB has an instance member of type ClassC. This can be designed in such a way which allows direct access all the way down the dependency chain to ClassC or beyond, as in the following example:

The above example is quite common, however it violates The Principle of Least Knowledge as it creates multiple dependencies, thus reducing maintainability as should the internal structure of ClassA need to change so would all instances of ClassA.

Now keep in mind that in all software development there are trade-offs to some degree. Sometimes performance trumps maintainability or vice-versa, other times readability trumps both. A perfect example of where you would not want to use The Principle of Least Knowledge is in a Cairngorm ModelLocator implementation. The Cairngorm ModelLocator violates the Principle of least knowledge for good reason - it simply would not be practical to write wrapper methods for every object on the ModelLocator. This is the main drawback of the Principle of least Knowledge; the need to create wrapper methods for each object, which are more formally known as Demeter Transmogrifiers.

The goal of good software design is to minimize dependencies, and by carefully following the guidelines provided by The Principle of Least Knowledge this becomes much easier to accomplish.

When defining a constructor or method with more than three parameters it is considered a best practice to create a parameters object for holding the values which are required. This especially holds true when some of the parameters are optional (as they will contain default values) and also when two or more consecutive parameters are of the same type (as developers may accidentally transpose these parameters - which will not result in compile time or runtime errors, making debugging a nightmare!)

The most appropriate solution for such cases is to break up the method into separate methods, however when this is not feasible creating a parameters object will improve both code readability and provide a cleaner design which leaves less room for unexpected errors.

Consider the following method:

The parameters defined in this method are typical of what you will often see, however it is much cleaner and reliable to create a parameters object for holding these parameters. In addition, a parameters object allows for a much easier client implementation as default values need not be reassigned for all parameters which proceed a parameter which you need to specify a value other than the default value. For instance, if a method accepts 5 parameters and the first two parameters are required but the last three are optional, if you you need to specify a value for the last parameter you will need to re-assign the default values for the proceeding parameters, when in fact you really only need to specify three parameters.

The following example demonstrates creating a parameter object which can be passed to “someMethod”:

When invoking “someMethod” clients can now simply instantiate an instance of the parameter object, set values only for what is needed and pass it in as follows:

So if you would like to improved code readability and provide proactive exception precautions, consider utilizing parameter objects for methods which require more than three parameters.

The topic of Singletons in ActionScript 3.0 has been coming up again lately and it has been very interesting to see all of the unique solutions the community has come up with. In particular I like the idea of having Singleton metadata which would allow the compiler do all of the work for us.

Personally I feel the Singleton pattern is extremely useful. It is arguably one of the most common design patterns around today. Practically every management centric API requires a Singleton one way or another. Some developers claim that the Singleton pattern is nothing more than a euphemism for a global variable, to some extent this is true, however the intent of a Singleton is clearly different.

As useful as the Singleton pattern is my biggest complaint about Singletons has always been the actual construction code required to create / protect the Singleton instance. This extra code often becomes quite verbose and it is annoying to have to sift through all of the Singleton code when working with the actual class implementation code. It would also be nice to not have to constantly re-write the Singleton construction and implementation code every time a Singleton is needed.

So is there a way around all of this? Yes!

I developed a simple class called SingletonMonitor which singleton classes can extend to allow the omission of all Singleton specific construction code. All that is needed is to have the class which is to be a Singleton extend SingletonMonitor. That’s it. No more getInstance(), inner classes, type checking and so on is needed. As a best practice I recommend that you define the Singleton instance in the class itself in order to improve code readability.

An example demonstrating how to use the SingletonMonitor can be seen in the following:

As you can see no Singleton construction code is needed. Additionally, by extending SingletonMonitor you are clearly stating that the class is intended to be a Singleton.

So how is this accomplished? It’s pretty simple…

When a derived class is instantiated the SingletonMonitor constructor is invoked, the constructor parses the current stack trace in order to determine the derived class’ name (hence the hack). The name of the derived class is then used as a key in the SingletonMonitor hash table. When a subsequent instantiation of the class is made SingletonMonitor checks the name of the class and if it has previously been defined in the hash an exception is thrown. I originally developed this using introspection to determine the fully qualified name of the class, however the preferred implementation is to have the class eagerly instantiated at compiled time (i.e. constant), thus the stack trace is not available.

Admittedly this is a bit of a hack, but so are the alternatives, otherwise this issue would never have been a topic of discussion in the first place.

However what I like about this new approach is that the Singleton is being managed outside of it’s concrete implementation, and that is the goal of this post; to present an alternative means of managing Singleton construction. Through this the Singleton construction and management code can be omitted as it is being handled by a completely separate object.

So the SingletonMonitor was the first example of how Singleton management can be achieved. The second example demonstrates the Singleton management approach implemented via composition as opposed to inheritance - which is my preferred mechanism of Singleton Management. In addition to creating the SingletonMonitor which utilizes inheritance I also created a SingletonManager which utilizes composition. The SingletonManager is the implementation I recommend and prefer.

An example demonstrating how to use the SingletonManager can be seen in the following:

The SingletonManager requires the class constructor to invoke SingletonManager.validateInstance and pass in a reference of the instance. This is automated in the SingletonMonitor as the class name is determined automatically which is convenient, however the readability of the SingletonManager is preferred as it clearly states intent. Additionally the SingletonManager guarantees the correct type is resolved.

So this is a new way of thinking about Singletons; to provide a management system from which Singleton construction and protection can consistently provided.

To be honest, this was really just an experiment I have been playing around with for some time now that I thought I should share, I am not sure if I would use the SingletonMonitor in production code as the parsing of the stack trace just feels a bit to much like a hack. However I will most likely utilize the SingeltonManager moving forward as it is a great way to abstract Singleton construction and protection from the class implementation.

My hope is that there will be a true solution available as we move forward, but for the time being if you would like to create Singletons without the need for all of the Singleton implementation code feel free to extend SingletonMonitor for management or SingletonManager for compositional management.

My original post on Cairngorm and IResponder had accidentally been deleted while updating my moderations queue, and many of you have contacted me stating that the post is no longer available so I will re-iterate what I mentioned in that post.

For some time now I have been entertaining the notion of abstracting IResponder implementations from Command classes into separate, discreet classes which are intended to handle asynchronous service invocation responses. There has been some talk in the Cairngorm community recently regarding Cairngorm Commands and IResponder implementations so I thought I would share my thoughts on the subject.

Typically most Cairngorm Events are handled by an associated Command. The Command handles the Event by updating a model on the ModelLocator, and / or instantiating a Business Delegate to manage invoking a middle-tier service.

At this point one could argue that the Command has finished doing it’s job - handling the Event. Let’s clarify by taking a look at a formal definition of the Command Pattern:

“The Command pattern is a design pattern in which objects are used to represent actions. A command object encapsulates an action and its parameters.”

From this we can deduce (in the context of a Cairngorm Event) that the Event is the “action” and the Command is the object which encapsulates the handling of the Event (action). The actual handling of the Service response could be considered a separate concern which is outside of the direct concern of the Event and Command, thus requiring an additional object to handle the service response.

However for many developers it is (by design) typical to simply have the Command implement IResponder and also handle the response from the service in addition to the actual Event. This makes sense from a convenience perspective, but not necessarily from a design perspective.

What I have been experimenting with is pretty simple and straightforward. It involves having a completely separate object implement IResponder and handle the service response directly.

Consider the following example in which a specific use-case requires an account log in. The following classes would be required: LoginEvent, LoginCommand, LoginResponder and LoginDelegate. Utilizing a separate class as the responder is very simple and straightforward and would be implemented as follows:

So far nothing different here, the above Event is just like any other CairngormEvent. Now let’s take a look at the Command implementation.

Based on the above example, the only method which must be implemented is execute(), as defined by ICommand. The body of the execute() implementation instantiates an instance of LoginResponder and LoginDelegate, the LoginResponder instance is passed to the LoginDelegate as the IResponder instance.

As can be seen in the following example, the Business Delegate implementation is the same as any other typical Cairngorm Delegate:

The IResponder implementation would be as follows:

That’s pretty much it. Clean, simple, and yes, more code, however this design supports clean separation of concerns and promotes encapsulation and code reusability.

At the end of the day it really comes down to personal preference. For me, I always prefer to have more classes which encapsulate very specific tasks and responsibilities. As long as you have a clear and concise, but most of all, consistent design you usually can’t go wrong.