Blogadda

Blogadda
Visit BlogAdda.com to discover Indian blogs

Thursday, June 11, 2009

Multiple Inheritance in Java and Delegate Factory pattern

When Sun was designing Java, it omitted multiple inheritance - or more precisely multiple implementation inheritance - on purpose. Yet multiple inheritance can be useful, particularly when the potential ancestors of a class have orthogonal concerns. This article presents a utility class that not only allows multiple inheritance to be simulated, but also has other far-reaching applications.

Have you ever found yourself wanting to write something similar to:

public class Employee extends Person, Employment {
// detail omitted
}

Here, Person is a concrete class that represents a person, while Employment is another concrete class that represents the details of a person who is employed. If you could only put them together, you would have everything necessary to define and implement an Employee class. Except in Java - you can't. Inheriting implementation from more than one superclass - multiple implementation inheritance - is not a feature of the language. Java allows a class to have a single superclass and no more.

On the other hand, a class can implement multiple interfaces. In other words, Java supports multiple interface inheritance. Suppose the PersonLike interface is:

public interface PersonLike {
String getName();
int getAge();
}

and the EmployeeLike interface is:

public interface EmployeeLike {
float getSalary();
java.util.Date getHireDate();
}

If Person implements the Person-Like interface, and Employment implements an EmployeeLike interface, it's perfectly acceptable to write:

public class Employee implements PersonLike, EmployeeLike {
// detail omitted
}

Here there is no explicit superclass. Since we are allowed to specify at most one superclass, we could also write:

public class Employee extends Person implements PersonLike, EmployeeLike {
// detail omitted
}

We would need to write the implementation of EmployeeLike, but the implementation of PersonLike is taken care of through the Person superclass. Alternatively we might write:

public class Employee extends Employment implements PersonLike, EmployeeLike{
// detail omitted
}

This is the opposite situation: the EmployeeLike interface is taken care of through the Employment superclass, but we do need to write an implementation for PersonLike.

Java does not support multiple implementation inheritance, but does support multiple interface inheritance. When you read or overhear someone remark that Java does not support multiple inheritance, what is actually meant is that it does not support multiple implementation inheritance.

Stay Adaptable
Suppose then that you have the concrete implementations Person, which implements the PersonLike interface, and Employment, which implements the EmployeeLike interface. Although only one can be selected to be the superclass, it would be useful to somehow exploit the other implementation.

The easiest way to do this in Java is by applying the (Object) Adapter pattern. If we make Person the superclass, we can use Employment using an object adapter held within the employee:

public class Employee extends Person implements PersonLike, EmployeeLike {
private EmployeeLike employment = new
Employment();
public float getSalary() { return
employment.getSalary(); }
public java.util.Date getHireDate() { return employment.getHireDate(); }
}

For each method of EmployeeLike, the employee delegates to the object adapter. This helps motivate the decision as to whether Person or Employment should be the superclass; choose the one with the most methods as the superclass so there will be less manual delegation code to write when dealing with the other interface.

The Adapter pattern is a fine way to support multiple interface inheritance while exploiting two concrete implementations. Indeed, it's more often the case that an anonymous inner class is used as the object adapter, allowing customization of behavior with respect to the context (of being embedded within a subclass).

However, writing that delegation code is tedious, especially if both interfaces to be implemented have many methods in them. In many cases, we can get Java to do the delegation to the would-be superclass(es) automatically.

Enter Dynamic Proxies
Dynamic proxies were introduced into Java in J2SE v1.3. Part of the java.lang.reflect package, they allow Java to synthesize a class at runtime. The methods supported by this synthesized class are specified by the interface (or interfaces) that it implements. The implementation is taken care of through an invocation handler (java.lang.reflect.InvocationHandler) that is handed an object representing the method being invoked (java.lang. reflect.Method). As you can see, dynamic proxies use heavy doses of the Java Reflection API.

This then is the key to simulating multiple implementation inheritance within Java. We can write a custom InvocationHandler that is constructed with a set of classes; these represent the superclasses of the subclass to be synthesized. The interface(s) of our subclass will be the union of the interfaces implemented by these superclasses. Our InvocationHandler will instantiate instances of these superclasses so that it can delegate to them. We then arrange it so that the invocation handler, on being given a method to be invoked, will reflectively invoke the method on the appropriate superclass object instance. (There must be one; remember the subclass's interface is derived from the superclass's, so at least one superclass must be able to handle the method invocation.)

To make things simple, we can make our InvocationHandler implementation also return the proxy. In other words, the invocation handler can act as a factory, able to return instance(s) of the synthesized subclass that will delegate to the superclass instances. We call our invocation handler implementation DelegatorFactory for this reason:

// imports omitted
public final class DelegatorFactory
implements InvocationHandler {
public Object getObject() {
return Proxy.newProxyInstance(
this.getClass().getClassLoader(),
getSupportedInterfaces(),
this);
}
}
// code omitted
}

The supported interfaces of the resultant object are derived from the superclasses provided in the DelegatorFactory's constructor:

// imports omitted
public final class DelegatorFactory implements InvocationHandler {
public DelegatorFactory(final Class[]
ancestors) {
// implementation omitted
}
// code omitted
}

There is more to DelegatorFactory as we shall soon see, but we now have enough to simulate multiple implementation inheritance. Going back to the question first posed, instead of:

public class Employee extends Person, Employment {
// detail omitted
}

followed (presumably) by:

Employee employee = new Employee();

We can instead write:

Object employee =
new DelgatorFactory(
new Class[] {
Person.class,
Employee.class
}
).getObject();

Although the syntax is somewhat different, the same essential information is being provided. That is, the concrete implementations are provided in Person and in Employment. This object will use the implementation of Person if invoked as a PersonLike, and the implementation of Employment if invoked as an EmployeeLike:

((PersonLike)employee).getAge();
((EmployableLike)employee).getHireDate();

How Convenient
In the above example, the casts are necessary because the getObject() method of DelegatorFactory can only return a reference of type java.lang.Object. But the clunkiness arises because our original aim of defining the Employee class with two concrete superclasses actually does something else as well:

public class Employee extends Person, Employment {
// detail omitted
}

Not only does this indicate that the implementation of Employee should be based on that of its superclasses, it also defines Employee as a type. In other words, it's then possible to write:

Employee employee;

What is missing in our dynamic proxy solution is this definition of type. Let's first do that in the usual way. As shown in Figure 2, we don't need to use a class though; an interface is sufficient.

As code, this is simply:

public interface Employee extends PersonLike, EmployeeLike { }

There is no detail omitted here; this is our complete definition. Note that Employee is now an interface and not a class. The following will not work, however:

Employee employee =
(Employee)
new DelegatorFactory(
new Class[] {
Person.class,
Employment.class
}
).getObject();

This is because the only interfaces implemented by the dynamic proxy returned by getObject() are PersonLike and EmployableLike. No matter that logically the Employee interface does not require any additional implementation from our dynamically created object; Employee is not an interface that we can cast to. However, DelegatorFactory does provide an alternative constructor:

Employee employee =
(Employee)
new DelegatorFactory(
new Class[] {
Person.class,
Employment.class
},
Employee.class
).getObject();

Note the new second argument (Employee.class) to the constructor. Casting the object returned from getObject() to Employee will now work. Behind the scenes, the Delegator- Factory simply adds this interface to the set of those to be implemented by the dynamic proxy. Note that Delegator Factory takes this interface object on trust: there is no validation that the interface doesn't introduce any new methods that are not already present in the interfaces of the superclasses.

Initializing the Superclasses
In "regular" Java, if a superclass does not provide a no-arg constructor, it's necessary for the subclass to correctly initialize the superclass using constructor chaining. Normally this is done by including the superclass's constructor's argument(s) in the subclass's constructor's argument(s), and then passing them up the class hierarchy using super().

The facilities shown in Delegator-Factory thus far do not support this. The DelegatorFactory is given a list of superclasses, and then instantiates an instance of each (to delegate to) using java.lang.Class.newInstance(). This requires a public no-arg constructor to exist.

If the would-be superclass does not offer a public no-arg constructor, the DelegatorFactory should be instantiated using a different constructor that takes preinstantiated superclass instances:

Person person = new Person("joe", 28);
Employment employment =
new Employment(someCalendar.getTime(),
30000);
Employee employee =
(Employee)
new DelegatorFactory(
new Object[] {
person, employment
},
Employee.class
).getObject();

If the would-be superclass does not have a public constructor, or is abstract, a custom subclass (probably an anonymous inner class) should be instantiated and used instead.

Dealing with Diamonds
Typically, multiple implementation inheritance is used when the superclasses have orthogonal concerns. Certainly this is the case with PersonLike and EmployeeLike, and each method is unambiguous as to which ancestor it relates to.

However, sometimes there may be a common super-interface in the interfaces implemented by the "superclasses." For example, suppose we have the concrete class, Car, which implements Driveable, the Boat class, which implements Sailable, and both Driveable and Sailable extend from Steerable. Since we want to use both Car and Boat to define a new subclass, we will also introduce a convenience interface, AmphibiousCar (see Figure 3).

The steer() method of Steerable is used to alter the bearing (0 to 359 degrees) of the steerable object. The getBearing() method, of course, should return this bearing.

For simplicity, the drive() method of Driveable and the sail() method of Sailable return a suitable string indicating the current bearing. Invoking drive() might return a string such as:

driving at bearing 30 degrees.

From what we currently know, we would create an amphibious car object using:

AmphibiousCar ac =
(AmphibiousCar)
new DelegatorFactory(
Class[] {
Car.class, Boat.class
}).getObject();

What happens if we invoke the steer() method on our new amphibious car ac? Should the invocation handler delegate to the Car superclass object or the Boat? The default behavior is to delegate to the first matching object. Hence, we will get:

ac.steer(30);
System.out.println(ac.drive());
// prints "driving at bearing 30 degrees"
System.out.println(ac.sail());
// prints "sailing at bearing 0 degrees"

The Boat superclass component of our class never knew that the bearing had changed.

It's this kind of problem that persuaded the Java language designers to exclude multiple implementation inheritance. This is too large an area to cover in this article, but what we have here is an example of part of the so-called "diamond" problem, where there is a common ancestor. You can see the diamond in the interfaces: Steerable, Driveable, Sailable, and Amphibious-Car.

The DelegatorFactory utility deals with the diamond problem by allowing you to specify the invocation behavior to the delegate superclasses as a pluggable strategy (an example of the Strategy pattern). The strategy is defined by the InvocationStrategy interface. The default strategy (InvokeFirstOnlyStrategy) is to invoke the first ancestor superclass that can handle the method. However, in the case of the diamond, what is required is that both ancestors need to handle the method. The InvokeAllStrategy handles this. If the method being invoked has a nonvoid return type, the return value from the first ancestor is returned. The two strategies are shown in Figure 4.

The invocation strategy can either be set after the DelegatorFactory has been instantiated, or can be set through (yet another) overloaded constructor. Hence our amphibious car should be created using:

AmphibiousCar ac =
(AmphibiousCar)
new DelegatorFactory(
Class[] {
Car.class, Boat.class
},
new InvokeAllStrategy()
).getObject();

This time, we get:

ac.steer(30);
System.out.println(ac.drive());
// prints "driving at bearing 30 degrees"
System.out.println(ac.sail());
// prints "sailing at bearing 30 degrees"

The InvokeFirstOnlyStrategy and InvokeAllStrategy are not the only strategies available (indeed we shall see one more shortly); however, they should work for most situations.

If a custom invocation strategy is required, it can be written by implementing the InvocationStrategy interface:

public interface InvocationStrategy {
Object invoke(final List ancestors,
final Method method,
final Object[] args)
throws Throwable
}

The ancestors parameter is an immutable list of the object instances representing the superclass. The method parameter represents the Method being invoked, and the args parameter contains the arguments to that Method. A typical invocation strategy would likely call method.invoke(S) somewhere within its implementation, with the first argument (the object upon which to invoke the method) being one of the ancestors.

We shall look at some applications of custom invocation strategies shortly. For now, though, an adaptation of InvokeAllStrategy might be to return the average return value of all ancestors, not just the return value of the first one.

Implicit Diamonds
In the previous diamond example, the Steerable interface is explicitly a super-interface of both Driveable and Sailable. What if the super-interface has not been explicitly factored out, though?

For example, in the original PersonLike and EmployeeLike example, what if each provided a foo() method, returning a string. Not imaginative, but never mind. Let's construct our employee and use an InvokeAllStrategy:

Employee employee = (Employee)
new DelegatorFactory(new Class[]{Person.class, Employment.class},
Employee.class,
new InvokeAllStrategy())
.getObject();

Now let us invoke foo():

employee.foo(); // what will happen?

Should the Person's implementation be called, that of Employment, or both? Although you might wish that both would be called (by virtue of our installed strategy), the sad truth is that only Person's implementation would be called. This is because the dynamic proxy has no way of knowing which interface to match foo() to, so it simply matches it to the first interface listed. (It's a java.lang.reflect.Method that is passed to the DelegatorFactory, not the string literal "foo()". Methods are associated with a specific declaring class/interface.) In terms of the DelegatorFactory's implementation, this means the first superclass listed in its constructor.

Note also that the compile time type does not matter. Neither of the following will change the outcome:

((PersonLike)employee).foo(); ((EmployeeLike)employee).foo();

In fact, it would be possible to modify DelegatorFactory to make Invoke-AllStrategy effective in this case, but that would involve parsing on the Method.getName() rather than the method. However, this has deliberately not been done. We'd rather you factored out the super-interface and made the diamond explicit. In the above example, add a FooLike (or Fooable) interface and make both PersonLike and EmployLike extend from it.

Other Applications
The issue raised by diamonds (implicit or otherwise) is that of how to deal with more than one implementation of a given method within an interface. However, it's interesting to turn this on its head.

In aircraft and other safety-critical environments, it's common to implement subsystems in triplicate. For example, there may be three different navigational systems, possibly with each implemented by different subcontractors. Each of these would be able to respond to the request, "Where is the location of the aircraft?"

Other systems within the aircraft interact with the navigational subsystem through a broker. This accepts the request on behalf of the navigational subsystem, and then forwards the request onto each implementation. Assuming there are no bugs in any of those implementations, they should all respond with the same data (within some delta of acceptable variance).

If there is a bug in one of the implementations, it may produce a response that differs wildly from the other two implementations. In this case, the broker disregards that response completely and uses the responses of the other implementations that agree with each other.

The design of DelegatorFactory and its pluggable invocation strategies make it easy to implement such a broker. Imagine a Calculator interface that defines a single method add(int, int):int. We can then have three implementations of this interface, as shown in Figure 5.

The FastCalculator uses regular integer arithmetic. The OneByOne- Calculator rather long-windedly performs its arithmetic by incrementing the first operand one-by-one in a loop. Both of these implementations are correct, just different. The final BrokenCalculator is just that; it actually performs a subtraction, not an addition.

The InvokeSafelyStrategy invocation strategy requires at least three ancestors that implement each method invoked. It will invoke the method on all ancestors, and then look to see that there is precisely one response that is most popular. Here is how to create a safe calculator that will ignore the incorrect implementation within the BrokenCalculator:

DelegatorFactory dfInvokeSafely =
new DelegatorFactory(
new Class[] {
BrokenCalculator.class,
OneByOneCalculator.class,
FastCalculator.class
},
Calculator.class,
new InvokeSafelyStrategy()
);
Calculator safeCalculator =
(Calculator)dfInvokeSafely.getObject();
assertEquals(7, safeCalculator.add(3,4));

Note that the InvokeSafelyStrategy is not all that intelligent. It stores the return values from each ancestor within a HashSet, so it relies on an accurate implementation of equals() and hashCode(). If the actual return type were a float (wrapped within a Float object), a more sophisticated invocation strategy would most likely be required. In general, this strategy will work only with well-defined value objects that can intrinsically deal with any rounding and other such errors.

You could easily adapt or refine the InvokeSafelyStrategy into further strategies. For example:

  • A parameterized version of InvokeSafelyStrategy could be used to deal with floats and other return types that would have rounding issues.
  • A background strategy might perform each invocation within a separate thread. Any invocation that had not responded within a certain timeout would be discarded.
  • A high-performance system, on the other hand, might use a strategy that again uses a backgrounding strategy but returns the result of the first one to finish, killing off the rest.
  • A logging strategy might perform some logging and then forward the invocation (typically to a single delegate).
  • A caching strategy would check its cache with respect to the input parameter, and only if the result is unknown would it invoke the delegate (caching the subsequent result).
  • A listener/broadcast strategy could represent a collection of listener objects; notifying all listeners of an event would require notifying only the broadcaster, which would then iterate over all listener objects as required.

    Moreover, there is nothing to prevent multiple invocations from being chained together, (that is, the Decorator pattern). Alternatively, we could imagine a composite strategy (the Composite pattern) that combines a set of strategies together. Either the invocation chain (decorator) or the set of leaf strategies (composite) could be changed at runtime, meaning that we can change the behavior and responsibilities of the object dynamically. This is a fundamentally different paradigm from conventional Java with its static typing. Normally, it's the type/class of the object that determines its behavior, something that cannot be changed once the object is instantiated. Here, though, we have ended up configuring the behavior of objects on an instance-by-instance basis: so-called instance-based programming. In effect, the whole notion of type becomes much less important.

    There are echoes here too of aspect-oriented programming. Most aspect-oriented programming uses compile-time techniques (the term used is "weaving") to add in behavior to classes. The classic example of aspect-oriented programming is to add logging within all method calls. You can easily see, though, that these same features can be incorporated dynamically using invocation strategies; the decorator/composite invocation strategies would allow an arbitrary set of aspects to be added to a class. The difference though is that now the aspects are applied at runtime (and hence can be changed without recompile and redeployment).

    Conclusion
    The DelegatorFactory is simple to use, supporting classic mix-in (orthogonal) multiple-implementation inheritance "out-of-the-box" and - with its pluggable invocation strategy design - allows diamond hierarchies to be easily supported. Moreover, the design also lends itself to other quite unrelated problem spaces; for example, creating safe systems was explored. Taken to its logical conclusion, the approach supports both instance-based programming and aspect-oriented programming.

    Of course, what makes DelegatorFactory work is Java's support for dynamic proxies, and that in turn requires that the ancestor superclasses implement interfaces. This approach won't work for class-based designs (JDOM is an example that comes to mind). But arguably class-based designs should be used only for value objects that should be final anyway. Those situations where multiple inheritance is desired are more likely to occur when working with reference objects.

    One particular case deliberately not supported by DelegatorFactory is when there is a so-called implicit diamond. The solution though is to pull out the methods that appear in both interfaces, and move them into a new super-interface. Then, make sure you use InvokeAllStrategy rather than the default InvokeFirstOnlyStrategy.

    Of course, using a dynamic proxy object will be slower than a hand-crafted solution, principally because reflection is used. However, the difference may not be noticeable in practice. In recent releases of Java, Sun has put much effort in speeding up reflective invocation; as of JDK 1.4.1, it may well be that regular invocation is only twice as fast as reflective invocation (previously this figure was something like 40 times faster).

    Using DelegatorFactory
    The DelegatorFactory utility class and supporting classes described here can be downloaded from www.sys-con.com/java/sourcec.cfm, and are compilable using Ant (v1.5.1 was used to create the build file). A JUnit-based test harness is also provided; JUnit v3.8.1 is required. The motivating examples in this article are based on the JUnit tests, so they should be easy enough to follow.

    To run the tests with JUnit's text-based test runner, use:

    ant test

    Alternatively, you can use JUnit's test runner by running directly:

    ant rebuild
    java -classpath
    dist/halware-util-dynamic-bin.jar;dist/halware-util-dynamic-bin-test.jar
    com.halware.util.dynamic.test.AllTests gui

    (The GUI test runner is not the default since JUnit's classloaders do not understand the Class-Path manifest attribute.)

    I hope you find DelegatorFactory useful. It has been distributed under the GNU Lesser Public License, so you are free to embed it within your own software as required.

  • No comments: