Not all frameworks are alike. One way to distinguish between them is the notion of white box vs. black box.
A white-box framework requires the framework user to understand the internals of the framework to use it effectively. In a white box framework, you usually extend behavior by creating subclasses, taking advantage of inheritance. A white box framework often comes with source code.
A black-box framework does not require a deep understanding of the framework’s implementation in order to use it. Behavior is extended by composing objects together, and delegating behavior between objects.
A framework can be both white-box and black-box at the same time. Your perception of how “transparent” a framework is may depend on non-code aspects such as documentation or tools.
Frameworks tend to change over their lifetime. (See [Johnson & Roberts].) When a framework is new, it tends to be white-box: you change things by subclassing, and you have to peak at source code to get things done. As it evolves, it becomes more black-box, and you find yourself composing structures of smaller objects. Johnson and Roberts point out that frameworks can evolve beyond black-box, perhaps becoming visual programming environments, where programs can be created by interconnecting components selected from a palette. (JavaBeans is an effort in that direction.)
Visual environments and even black-box frameworks sound so much easier to use than white-box frameworks – why would we ever bother creating white-box frameworks in the first place? The basic answer is “cost”. To develop a black-box framework, we need a sense of which objects change the most, so we can know where the flexibility is most needed. To develop a visual environment, we need even more information: we need to know how objects are typically connected together. Discovering this costs time. White-box frameworks are easier to create and have more flexibility.
White-Box Frameworks
The most common sign that a framework is white-box is heavy use of inheritance. When you use the framework by extending a number of abstract (or even concrete) classes, you are dealing with a white-box framework. Inheritance is a closer coupling than composition; an inherited class has more context it must be aware of and maintain. This is visible even in Java’s protection scheme: a subclass has access to the public and protected parts of the class, while a separate object only sees the public parts. Furthermore, a subclass can potentially “break” a superclass even in methods it doesn’t override, for example by changing a protected field in an unexpected way.
What are the effects of an inheritance-based approach?
· We need to understand how the subclass and superclass work together.
· We have access to both the protected and the public parts of the class.
· To provide functionality, we can override existing methods, and implement abstract methods.
· We have access to the parent’s methods (by calling super.method()).
Example: A simple applet
import …
public class MyApplet extends Applet {
public void paint(…) {…} //TBD??
}
Notice how we depend directly on the superclass, even to using its methods freely.
A subclass is coupled to its parents, and we deal with the benefits and costs of that fact.
Black-Box Frameworks
Black-box frameworks are based on composition and delegation rather than inheritance. Delegation is the idea that instead of an object doing something itself, it gives another object the task. When you delegate, the object you delegate to has a protocol or interface it supports, that the main object can rely on.
In black-box frameworks, objects tend to be smaller, and there tend to be more of them. The intelligence of the system comes as much from how these objects are connected together as much as what they do in themselves.
Composition tends to be more flexible than inheritance. Consider an object that uses inheritance versus one that delegates. With inheritance, the object basically has two choices: it can do the work itself, or it can call on the parent class to do it. In a language like Java, the parent class is fixed at compile time.
With delegation, an object can do the work itself (or perhaps in its parent), or it can give the work to another object. This object can be of any useful class (rather than only the parent), but it can also change over time. (You can delegate to an object, change the delegate to be another object, and use the new object as the delegate next time.)
Example: TBD
[TBD compare to cutting grass – children vs. lawn service?]
Converting Inheritance to Composition
In “Designing Reusable Classes,” Johnson and Foote identify this rule:
Rule 12. Send messages to components instead of to self. An inheritance-based framework can be converted into a component-based framework black box structure by replacing over-ridden methods by message sends to components.
Let’s apply their advice to an example. Suppose we’ve got a class that sues the Template Method design pattern like this:
public class ListProcessor {
final public void processList() {
doSetup();
while (hasMore()) {
doProcess1();
}
doTeardown();
}
protected void doSetup() {}
protected void doTeardown() {}
abstract protected boolean hasMore();
abstract protected void doProcess1();
// …
}
with this as a typical subclass:
public class TestProcessor extends ListProcessor {
int i = 0;
protected boolean hasMore() {return i < 5;}
protected void doProcess1() {System.out.println(i);}
}
In the simplest form of the transformation, we can take each method designed to be subclassed, and whenever it is called, replace it by a message to the delegate. We’ll also add some code to set up the delegate.
public class ListProcessor {
ListProcessor delegate;
public ListProcessor(lp) {delegate = lp;}
final public void processList() {
delegate.doSetup();
while (delegate.hasMore()) {
delegate.doProcess1();
}
delegate.doTeardown();
}
protected void doSetup() {}
protected void doTeardown() {}
abstract protected boolean hasMore();
abstract protected void doProcess1();
// …
}
We can extend this like before. [TBD] You might create it like this:
ListProcessor lp = new ListProcessor (new TestProcessor());
Compare how these two situations look at run-time:
[LP] ß [Test] => [:Test]
[LP]odelegate => [:LP] — [:Test]
^
[Test]
So far, this doesn’t seem to be worth the trouble: ListProcessor already has a copy of “processList()”; it doesn’t need the one in Test.
The next step is to introduce a new class for the delegate, restricted to just the capabilities it needs. The methods called on the delegate define its protocol. We could handle this via an abstract class, but I prefer to use a Java interface:
public interface ListProcessAction {
void doSetup();
boolean hasMore();
void doProcess1();
void doTeardown();
}
These are the methods intended to be over-ridden.
Now the main class can use the ListProcessorAction for its delegate. Furthermore, as ListProcessor is no longer intended to be subclassed, it no longer has any need for those protected methods:
We could make our class depend on this interface:
public class ListProcessor implements ListProcessorAction {
ListProcessorAction delegate;
public ListProcessor(ListProcessorAction a) {delegate = a;}
final public void processList() {
delegate.doSetup();
while (delegate.hasMore()) {
delegate.doProcess1();
}
delegate.doTeardown();
}
// …
}
with this concrete implementation of the action:
public class ConcreteAction {
int i = 5;
public void doSetup() {}
public boolean hasMore() {return i < 5;}
public void doProcess1() {System.out.println(i);}
public void doTeardown() {}
}
[TBD: Typically when these involve abstract methods, you might create an abstract class version, which will be extended by the end class. Or if the protocol is small and completely abstract, you don’t need concrete classes.]
The structure looks like this:
[LP] -delegate- [<<interface>> LPA] => [:LP] -delegate- [:CA]
^
[ConcreteAction]
This runtime structure is similar to the previous one, but now ListProcessor and ConcreteAction are separate classes.
We have split one big object, that knew both the algorithm and the individual steps, into two classes, one for each concern. Look at the tradeoffs. In the initial version, everything was in one place. To trace the new version, you have to understand the delegation structure and how it can vary at runtime. When you write a new action, it’s easier to focus on it in isolation, but harder to see how it fits into the big picture.
See how the design has changed: [TBD]
[//TBD]
Step by Step
This is a systematic description of how to convert
[Base] to [Base] –delegate– [<<int>>Action] <- [ConcreteAction]
Cautions: [TBD]
· Calls to super(). (Need to work through ramifications.)
· Recursion. (Need to eliminate or understand.)
This approach follows a re-factoring style, moving in small steps and letting the compiler do the work. (See [Fowler].)
1. Create a new interface.
public interface Action { }
(name it appropriately).
2. Create a new class implementing this interface:
public class ConcreteAction implements Action {}
3. In Base, add a delegate field, and modify each constructor to take an Action as a parameter; use this to set the delegate:
protected Action delegate;
public Base (Action a) {
delegate = a;
// rest as before
}
…
4. Each protected method in Base is presumably called in Base. For each such method:
· Move the signature to Action and change it to “public”
· Move the routine itself to ConcreteAction (and make it public there as well).
· Replace each call to “method()” with “delegate.method()”.
For example,
Base {…
protected method() { /*impl*/.}
… method(); …
}
becomes
Base {… delegate.method(); … }
Action { … public method(); …}
ConcreteAction {… public method() {/*impl*/} … }
Note: You can find the call sites by temporarily changing the method name to “xxx”, and seeing what breaks in base.
Moving protected methods may force you to pull over some private methods as well, or perhaps maintain a reference back to Base. Unfortunately, this is not a fully mechanical process. Similarly, if Base’s methods involve recursion or calls to super(), you will need insight into how the class works and how you want to split it.
5. Check whether methods in Base call any of Base’s public methods. If so,
· Copy the signature to Action
· Copy the method to ConcreteAction
· Replace the method body in Base with “delegate.method()”.
Again, be aware of private methods, recursion, and super().
6. Polish things up: get Base, Action, and ConcreteAction to compile properly.
7. Check out any subclasses of Base. Figure out whether they should remain a subclass of Base, become a subclass of ConcreteAction, or become an independent action implementing the Action interface. (The class may need to be split.) Distribute the subclass’ behavior where it should go. As usual, be careful about recursion and super().
8. Find each call to a constructor for Base or its subclasses. (Let the compiler tell you where they are.) Add a new parameter to the constructor, “new ConcreteAction()” where Base wants it.
Conflicting Class Hierarchies
Java only supports single inheritance, but sometimes you find yourself wanting multiple inheritance. You can use interfaces and delegation to help in many such situations.
Look at java.util.Observable. In some ways, it could be a basis for an implementation of the Observer design pattern. However, the fact that it is a class and not an interface is a flaw.
Suppose you have a Vector, and you’d like it be Observable (perhaps so a listbox widget can watch it for changes). Because both Vector and Observable classes are already classes, you’d like this:
[Vector] <- ObservableVector -> [Observable] // MI => !Java
Suppose Observable were an interface instead. A class can implement as many interfaces as it needs to, so we could do this:
[Vector] <- ObservableVector – – – > [<<int>> Observable] // Legal Java but not JDK 1.x
[TBD – double-check against JDK]
There’s a reason Observable is a class though: it maintains machinery for notification. If we had a convenience implementation, we could delegate to it.
[Vector] <- ObservableVector — — > [<<int>> Observable]
— delegate — [ObservableHelper]
If Java had multiple inheritance, one class could cover both the Vector behavior and the Observable behavior. Without multiple inheritance, we can get the effect by connecting together a pair of classes.
[TBD: How class can we get to the 1.1 Listener event model?]
The Swing library designers faced this problem. Their solution is to ignore Observable, and instead create a ListModel that models the basics of vector handling. In Swing, there is an AbstractListModel, that handles notification, and can delegate to a concrete vector or list class.
Inner Classes for Multiple Inheritance
Sometimes a class implements several interfaces when it would be better served by using Java’s inner classes. Consider this example:
public class MyPanel extends JPanel implements MouseListener {
JButton b1 = new JButton(“First”);
JButton b2 = new JButton(“Second”);
public MyPanel() {
b1.addActionListener(this);
add(b1);
b2.addActionListener(this);
add(b2);
}
public void actionPerformed(ActionEvent e) {
if (e.getSource() == b1) {
System.out.println(“b1 action”);
} else { // assume b2
System.out.println(“b2 action”);
}
}
}
Here, MyPanel acts both like a JPanel and an listener. Notice the code for actionPerformed(): it’s got an ugly “if” statement that’s practically a case statement. Such a construct is a sign that we’re not as OO as we could be, and we can move intelligence around.
We’ll use a pair of inner classes, cleaning up MyPanel a bit. This keeps it from receiving unnecessary notifications, and avoids the need for a test to see which button was clicked.
public class MyPanel extends JPanel implements MouseListener {
JButton b1 = new JButton(“First”);
JButton b2 = new JButton(“Second”);
public class FirstListener implements ActionListener {
public void actionPerformed(ActionEvent e) {
System.out.println(“b1”+e);
}
}
public class SecondListener implements ActionListener {
public void actionPerformed(ActionEvent e) {
System.out.println(“b2” + Date.getTime());
}
}
public MyPanel() {
b1.addActionListener(new FirstListener());
add(b1);
b2.addActionListener(new SecondListener());
add(b2);
}
}
The first version was more in the style of JDK 1.0.2, where the event detection hierarchy had to match the container hierarchy. The second version is more in JDK 1.1 style.
You can carry this a step further to use anonymous inner classes:
public class MyPanel extends JPanel implements MouseListener {
JButton b1 = new JButton(“First”);
JButton b2 = new JButton(“Second”);
public MyPanel() {
b1.addActionListener(new ActionListener() {
public void actionPerformed(ActionEvent e) {
System.out.println(“b1”+e);
}
});
add(b1);
b2.addActionListener(new ActionListener() {
public void actionPerformed(ActionEvent e) {
System.out.println(“b2” + Date.getTime());
}
});
add(b2);
}
}
Many people might find this on the edge of readability – some on the near side and some on the far side.
Decorator and Accumulating Functionality
[TBD]
Interfaces and Plug-Compatibility
[TBD]
Type Generality
Miscellaneous
[TBD]
“Rule 6. The top of the class hierarchy should be abstract.”
Responsibility-Driven Design Tutorial, #21, OOPSLA ’98, R. Wirfs-Brock & Alan McKean [TBD]
Slide 81: When factoring, favor simplicity
From simple to complex:
Parameterizing a method
Simple condition checking & decisions within a method
Delegation to a pluggable object [Java interface – template method]
Classification and inheritance
Dynamic parsing & interpretation
Slide 84: Inheritance and Hot Spots
Configurable Algorithm Design
Tag steps as replaceable
Send messages to self
Template method: Entire algorithm is final.
end Wirfs-Brock