Computer Science Design Patterns/Factory method

{{Computer Science Design Patterns/Page The factory pattern is a design pattern used to promote encapsulation of data representation.
 * mode={{{mode|reading}}}
 * page={{{page|{{SUBPAGENAME}}}}}
 * stage=25
 * stage date=29 Jan, 2014
 * stage comment=Add more illustrations.
 * previous=Facade
 * next=Flyweight
 * content=
 * Problem : We want to decide at run time what object is to be created based on some configuration or application parameter. When we write the code we do not know what class should be instantiated.
 * Solution : Define an interface for creating an object, but let subclasses decide which class to instantiate. Factory Method lets a class defer instantiation to subclasses. Its primary purpose is to provide a way for users to retrieve an instance with a known compile-time type, but whose runtime type may actually be different. In other words, a factory method that is supposed to return an instance of the class Foo may return an instance of the class Foo, or it may return an instance of the class Bar, so long as Bar inherits from Foo. The reason for this is that it strengthens the boundary between implementor and client, hiding the true representation of the data (see Abstraction Barrier) from the user, thereby allowing the implementor to change this representation at anytime without affecting the client, as long as the client facing interface doesn't change.

Basic Implementation of the Factory Pattern
The general template for implementing the factory pattern is to provide a primary user facing class with static methods which the user can use to get instances with that type. Constructors are then made private/protected from the user, forcing them to use the static factory methods to get objects. The following Java code shows a very simple implementation of the factory pattern for type Foo. With this code, it would be impossible for a client of the code to use the new operator to get an instance of the class, as is traditionally done: because the constructor is marked private. Instead, the client would have to use the factory method: It should be noted that even within a programming language community, there is no general consensus as to the naming convention of a factory method. Some suggest naming the method with the name of the class, similar to a normal constructor, but starting with a lowercase. Others say that this is confusing, and suggest using an accessor type syntax, like the getInstance style used above, though others complain that this may incorrectly imply a singleton implementation. Likewise, some offer newInstance, but this is criticized as being misleading in certain situations where a strictly new instance may not actually be returned (once again, refer to the singleton pattern). As such, we will not attempt to follow any particularly rigid standard here, we will simply try to use a name that makes the most sense for our current purposes.

Factory Pattern Implementation of the Alphabet
That's great, you know how to implement a real simple factory pattern, but what good did it do you? Users are asking for something that fits into type Foo, and they're getting an instance of the class Foo, how is that different from just calling the constructor? Well it's not, except that you're putting another function call on the stack (which is a bad thing). But that's only for the above case. We'll now discuss a more useful use of the factory pattern. Consider a simple type called Letter, representing a letter in the alphabet, which has the following client facing interface (i.e., public instance methods): We could implement this easily enough without using the factory method, which might start out something like this: Fairly simple, but notice we haven't implemented the last two methods yet. We can still do it pretty easily. The first might look like this: Now that's not too bad, but we still need to do isConsonant. Fortunately, we at least know in this case that if it's a vowel, it's not a consonant, and vice versa, so our last method could simply be: So what's the problem here? Basically, every time you call either of these methods, your program has to do all that checking. Granted, this isn't a real heavy burden for the Java Runtime Environment, but you can imagine a much more complex, much more time consuming operation. Wouldn't it be great if we could avoid doing this every time we call the method? Let's say, for instance, we could do it once when we create the object, and then not have to do it again. Well sure, we can do that. Here's an implementation that'll do that for us, and we still don't have to use the factory method: Notice how we moved the lengthy operation into the constructor, and stored the result. OK, so now we're all fine and dandy, no? Sure, but let's say you came up with a new idea, a different implementation: you want to split this type into two classes, one class to handle the vowels, and one to handle the consonants. Great, they can both be subclasses of the Letter class, and the user will never know the difference, right? Wrong. How is the client supposed to get at these new classes? They've got code that works perfectly well for them by calling new Letter('a') and new Letter('Z'). Now you're going to make them go through all their code and change these to new Vowel('a') and new Consonant('Z')? They probably won't be too happy with that. If only you could get new instances of both classes from one method! Well you can, just use a static method in the Letter class, it'll do the same one-time checking as the constructor did, and will return an appropriate instance of the right class. And what do you know, it's a factory method! But that still doesn't do your client much good, they still need to go through and change all the new Letters into Letter.getLetter. Well, sad to say, it's too late to help them at all, unless you just give up your new implementation. But that illustrates the reason for using the factory method right off the bat. One of the key components of good object oriented programming is that you never know exactly where your code will go in the future. By making good use of the abstraction barrier and using encapsulation-friendly programming patterns, such as the factory pattern, you can better prepare yourself&mdash;and your client&mdash;for future changes to the specific implementation. In particular, it allows you to use a "big hammer" kind of approach to get something done in a perhaps-less-than-ideal but rapid manner in order to meet deadlines or move ahead with testing,. You can then go back later and refine the implementation&mdash;the data representation and algorithms&mdash;to be faster, smaller, or what-have-you, and as long as you maintained the abstraction barrier between implementor and client and properly encapsulated your implementation, then you can change it without requiring the client to change any of their code. Well now that I'm sure you're a raving advocate for the factory method, let's take a look at how we would implement it for our Letter type: Several things to note here.
 * First, you'll notice the top level class Letter is abstract. This is fine because you'll notice that it doesn't actually do anything except define the interface and provide a top level container for the two other classes. However, it's also important (not just OK) to make this abstract because we don't want people trying to instantiate the Letter class directly. Of course we could solve this problem by making a private constructor, but making the class abstract instead is cleaner, and makes it more obvious that the Letter class is not meant to be instantiated. It also, as mentioned, allows us to define the user facing interface that the work horse classes need to implement.
 * The two nested classes we created are called local classes, which is basically the same as an inner class except that local classes are static, and inner classes are not. They have to be static so that our static factory method can create them. If they were non static (i.e., dynamic) then they could only be accessed through an instance of the Letter class, which we can never have because Letter is abstract. Also note that (in Java, anyway) the fields for inner and local classes typically use the "i" (for inner) prefix, as opposed to the "f" (for field) prefix used by top level classes. This is simply a naming convention used by many Java programmers and doesn't actually effect the program.
 * The two nested classes that implement the Letter data type do not actually have to be local/inner. They could just have easily been top level classes that extend the abstract Letter class. However, this is contrary to the point of the factory pattern, which is encapsulation. Top level classes can't be private in Java, because that doesn't make any sense (what are they private to?) and the whole point is that no client has to (or should, really) know how the type is implemented. Making these classes top level allows clients to potentially stumble across them, and worse yet, instantiate them, by-passing the factory pattern all together.
 * Lastly, this is not very good code. There's a lot of ways we can make it better to really illustrate the power of the factory pattern. I'll discuss these refactorings briefly, and then show another, more polished, version of the above code which includes a lot of them.

Refactoring the Factory Pattern
Notice that both of the local classes do the same thing in a few places. This is redundant code which is not only more work to write, but it's also highly discouraged in object oriented programming (partially because it takes more work to write, but mostly because it's harder to maintain and prone to errors, e.g., you find a bug in the code and change it in one spot, but forget to in another.) Below is a list of redundancies in the above code: In addition, as we discovered above, the isVowel and isConsonant just happen to always return the opposite of each other for a given instance. However, since this is something of a peculiarity for this particular example, we won't worry about it. The lesson you would learn from us doing that will already be covered in the refactoring of the getLetter method. OK, so we have redundant code in two classes. If you're familiar with abstracting processes, then this is probably a familiar scenario to you. Often, having redundant code in two different classes makes them prime candidates for abstraction, meaning that a new abstract class is created to implement the redundant code, and the two classes simply extend this new abstract class instead of implementing the redundant code. Well what do you know? We already have an abstract super class that our redundant classes have in common. All we have to do is make the super class implement the redundant code, and the other classes will automatically inherit this implementation, as long as we don't override it. So that works fine for the getLetter method, we can move both the method and the iTheLetter field up to the abstract parent class. But what about the constructors? Well our constructor takes an argument, so we won't automatically inherit it, that's just the way java works. But we can use the super keyword to automatically delegate to the super classes constructor. In other words, we'll implement the constructor in the super class, since that's where the field is anyway, and the other two classes will delegate to this method in their own constructors. For our example, this doesn't save much work, we're replacing a one line assignment with a one line call to super, but in theory, there could be hundred of lines of code in the constructors, and moving it up could be a great help. At this point, you might be a little worried about putting a constructor in the Letter class. Didn't I already say not to do that? I thought we didn't want people trying to instantiate Letter directly? Don't worry, the class is still abstract. Even if there's a concrete constructor, Java won't let you instantiate an abstract class, because it's abstract, it could have method that are accessible but undefined, and it wouldn't know what to do if such a method was invoked. So putting the constructor in is fine. After making the above refactorings, our code now looks like this: Note that we made our abstracted field protected. This isn't strictly necessary in this case, we could have left it private, because the subclasses don't actually need to access it at all. In general, I prefer to make things protected instead of private, since, as I mentioned, you can never really be sure where a project will go in the future, and you may not want to restrict future implementors (including yourself) unnecessarily. However, many people prefer to default to private and only use protected when they know it's necessary. A major reason for this is the rather peculiar and somewhat unexpected meaning of protected in Java, which allows not only subclasses, but anything in the same package to access it. This is a bit of a digression, but I think it's a fairly important debate that a good Java programmer should be aware of.
 * The field iTheLetter
 * The method getLetter
 * The constructor of each inner class does the same thing.

The Factory Pattern and Parametric Polymorphism
The version of the Java Virtual Machine 5.0 has introduced something called Parametric Polymorphism, which goes by many other names in other languages, including "generic typing" in C++. In order to really understand the rest of this section, you should read that section first. But basically, this means that you can introduce additional parameters into a class—parameters which are set at instantiation—that define the types of certain elements in the class, for instance fields or method return values. This is a very powerful tool which allows programmers to avoid a lot of those nasty instanceofs and narrowing castes. However, the implementation of this device in the JVM does not promote the use of the Factory pattern, and in the two do not play well together. This is because Java does not allow methods to be parameterized the way types are, so you cannot dynamically parameterize an instance through a method, only through use of the new operator. As an example, imagine a type Foo which is parameterized with a single type which we'll call T. In java we would write this class like this: Now we can have instances of Foo parameterized by all sorts of types, for instance: But let's say we want to use the factory pattern for Foo. How do we do that? You could create a different factory method for each type you want to parameterize over, for instance: But what about something like the ArrayList class (in the java.util package)? In the java standard libraries released with 5.0, ArrayList is parameterized to define the type of the object stored in it. We certainly don't want to restrict what kinds of types it can be parameterized with by having to write a factory method for each type. This is often the case with parameterized types: you don't know what types users will want to parameterize with, and you don't want to restrict them, so the factory pattern won't work for that. You are allowed to instantiate a parameterized type in a generic form, meaning you don't specify the parameter at all, you just instantiate it the way you would have before 5.0. But that forces you to give up the parameterization. This is how you do it with generics:

Examples
In Java, a class that implements  is a factory of statements. By calling the  method, you create a statement for which you only know the interface. The factory chooses the right instance class for you.

Cost
This pattern is not so expensive when it is implemented at the right time. It can be more expensive if you have to refactor an existing code.

Creation
Its implementation is easy and there is no additional cost (it is not more expensive than an implementation without this pattern).

Maintenance
There is no additional cost nor additional constraint.

Removal
This pattern can be easily removed as automatic refactoring operations can easily remove its existence.

Good practices

 * Name the factory class with the instanciated class in prefix and the factory term in suffix, to indicate the use of the pattern to the other developers.
 * Avoid to persist the factory result (in database) into the factory class, to respect the SOLID principles and allow the creation of ephemeral objects, for example into the automatic tests.

Implementation
This C++11 implementation is based on the pre C++98 implementation in the book.

The program output is like:

In Common Lisp, factory methods are not really needed, because classes and class names are first class values.

Example with pizza
Output: Price of HamMushroom is 850 Price of Deluxe is 1050 Price of Hawaiian is 1150

Another example with image

 * Abstract creator: Interface to create the Product.
 * Concrete creator : a class to create specific Product.
 * Factory class : a class to return a specific concrete creator at runtime to create the product.

This example in JavaScript uses Firebug console to output information. Output Price of HamAndMushroom is 8.50 Price of Deluxe is 10.50 Price of Hawaiian is 11.50

Usage example:

Factories are not really needed for this example in Perl, and this may be written more concisely: Output Price of HamAndMushroom is 8.50 Price of Deluxe is 10.50 Price of Hawaiian is 11.50

As in Perl, Common Lisp and other dynamic languages, factories of the above sort aren't really necessary, since classes are first-class objects and can be passed around directly, leading to this more natural version: Note in the above that the classes themselves are simply used as values, which pizza_class iterates over. The class gets created simply by treating it as a function. In this case, if pizza_class holds a class, then pizza_class creates a new object of that class. Another way of writing the final clause, which sticks more closely to the original example and uses strings instead of class objects, is as follows: In this case, the correct class name is constructed as a string by adding  'Pizza' , and eval is called to turn it into a class object.

}}