일반 An alternative to the deep copy technique

Use serialization to make deep copies and avoid extensive manual editing or extending of classes

By Dave Miller, JavaWorld.com, 08/06/99

Implementing a deep copy of an object can be a learning experience -- you learn that you don't want to do it! If the object in question refers to other complex objects, which in turn refer to others, then this task can be daunting indeed. Traditionally, each class in the object must be individually inspected and edited to implement the Cloneable interface and override its clone() method in order to make a deep copy of itself as well as its contained objects. This article describes a simple technique to use in place of this time-consuming conventional deep copy.
 

The concept of deep copy

In order to understand what a deep copy is, let's first look at the concept of shallow copying.

In a previous JavaWorld article, "How to avoid traps and correctly override methods from java.lang.Object," Mark Roulo explains how to clone objects as well as how to achieve shallow copying instead of deep copying. To summarize briefly here, a shallow copy occurs when an object is copied without its contained objects. To illustrate, Figure 1 shows an object, obj1, that contains two objects, containedObj1 and containedObj2.

If a shallow copy is performed on obj1, then it is copied but its contained objects are not, as shown in Figure 2.

A deep copy occurs when an object is copied along with the objects to which it refers. Figure 3 shows obj1 after a deep copy has been performed on it. Not only has obj1 been copied, but the objects contained within it have been copied as well.

If either of these contained objects themselves contain objects, then, in a deep copy, those objects are copied as well, and so on until the entire graph is traversed and copied. Each object is responsible for cloning itself via its clone() method. The default clone() method, inherited from Object, makes a shallow copy of the object. To achieve a deep copy, extra logic must be added that explicitly calls all contained objects' clone() methods, which in turn call their contained objects' clone() methods, and so on. Getting this correct can be difficult and time consuming, and is rarely fun. To make things even more complicated, if an object can't be modified directly and its clone() method produces a shallow copy, then the class must be extended, the clone() method overridden, and this new class used in place of the old. (For example, Vector does not contain the logic necessary for a deep copy.) And if you want to write code that defers until runtime the question of whether to make a deep or shallow copy an object, you're in for an even more complicated situation. In this case, there must be two copy functions for each object: one for a deep copy and one for a shallow. Finally, even if the object being deep copied contains multiple references to another object, the latter object should still only be copied once. This prevents the proliferation of objects, and heads off the special situation in which a circular reference produces an infinite loop of copies.

Serialization

Back in January of 1998, JavaWorld initiated its JavaBeans column by Mark Johnson with an article on serialization, "Do it the 'Nescafe' way -- with freeze-dried JavaBeans." To summarize, serialization is the ability to turn a graph of objects (including the degenerate case of a single object) into an array of bytes that can be turned back into an equivalent graph of objects. An object is said to be serializable if it or one of its ancestors implements java.io.Serializable or java.io.Externalizable. A serializable object can be serialized by passing it to the writeObject() method of an ObjectOutputStream object. This writes out the object's primitive data types, arrays, strings, and other object references. The writeObject() method is then called on the referred objects to serialize them as well. Further, each of these objects have their references and objects serialized; this process goes on and on until the entire graph is traversed and serialized. Does this sound familiar? This functionality can be used to achieve a deep copy.

Deep copy using serialization

The steps for making a deep copy using serialization are:

1. Ensure that all classes in the object's graph are serializable.
    
2. Create input and output streams.
    
3. Use the input and output streams to create object input and object output streams.
    
4. Pass the object that you want to copy to the object output stream.
    
5. Read the new object from the object input stream and cast it back to the class of the object you sent.

I have written a class called ObjectCloner that implements steps two through five. The line marked "A" sets up a ByteArrayOutputStream which is used to create the ObjectOutputStream on line B. Line C is where the magic is done. The writeObject() method recursively traverses the object's graph, generates a new object in byte form, and sends it to the ByteArrayOutputStream. Line D ensures the whole object has been sent. The code on line E then creates a ByteArrayInputStream and populates it with the contents of the ByteArrayOutputStream. Line F instantiates an ObjectInputStream using the ByteArrayInputStream created on line E and the object is deserialized and returned to the calling method on line G. Here's the code:

import java.io.*;
import java.util.*;
import java.awt.*;
public class ObjectCloner
{
   // so that nobody can accidentally create an ObjectCloner object
   private ObjectCloner(){}
   // returns a deep copy of an object
   static public Object deepCopy(Object oldObj) throws Exception
   {
      ObjectOutputStream oos = null;
      ObjectInputStream ois = null;
      try
      {
         ByteArrayOutputStream bos = 
               new ByteArrayOutputStream(); // A
         oos = new ObjectOutputStream(bos); // B
         // serialize and pass the object
         oos.writeObject(oldObj);   // C
         oos.flush();               // D
         ByteArrayInputStream bin = 
               new ByteArrayInputStream(bos.toByteArray()); // E
         ois = new ObjectInputStream(bin);                  // F
         // return the new object
         return ois.readObject(); // G
      }
      catch(Exception e)
      {
         System.out.println("Exception in ObjectCloner = " + e);
         throw(e);
      }
      finally
      {
         oos.close();
         ois.close();
      }
   }
   
}

All a developer with access to ObjectCloner is left to do before running this code is ensure that all classes in the object's graph are serializable. In most cases, this should have been done already; if not, it ought to be relatively easy to do with access to the source code. Most of the classes in the JDK are serializable; only the ones that are platform-dependent, such as FileDescriptor, are not. Also, any classes you get from a third-party vendor that are JavaBean-compliant are by definition serializable. Of course, if you extend a class that is serializable, then the new class is also serializable. With all of these serializable classes floating around, chances are that the only ones you may need to serialize are your own, and this is a piece of cake compared to going through each class and overwriting clone() to do a deep copy.

An easy way to find out if you have any nonserializable classes in an object's graph is to assume that they are all serializable and run ObjectCloner's deepCopy() method on it. If there is an object whose class is not serializable, then a java.io.NotSerializableException will be thrown, telling you which class caused the problem.

A quick implementation example is shown below. It creates a simple object, v1, which is a Vector that contains a Point. This object is then printed out to show its contents. The original object, v1, is then copied to a new object, vNew, which is printed to show that it contains the same value as v1. Next, the contents of v1 are changed, and finally both v1 and vNew are printed so that their values can be compared.

import java.util.*;
import java.awt.*;
public class Driver1
{
   static public void main(String[] args)
   {
      try
      {
         // get the method from the command line
         String meth;
         if((args.length == 1) && 
           ((args[0].equals("deep")) || (args[0].equals("shallow"))))
         {
            meth = args[0];
         }
         else
         {
            System.out.println("Usage: java Driver1 [deep, shallow]");
            return;
         }
         // create original object
         Vector v1 = new Vector();
         Point p1 = new Point(1,1);
         v1.addElement(p1);
         // see what it is
         System.out.println("Original = " + v1);
         Vector vNew = null;
         if(meth.equals("deep"))
         {
            // deep copy
            vNew = (Vector)(ObjectCloner.deepCopy(v1));  // A
         }
         else if(meth.equals("shallow"))
         {
            // shallow copy
            vNew = (Vector)v1.clone();                  // B
         }
         // verify it is the same
         System.out.println("New      = " + vNew);
         // change the original object's contents
         p1.x = 2;
         p1.y = 2;
         // see what is in each one now
         System.out.println("Original = " + v1);
         System.out.println("New      = " + vNew);
      }
      catch(Exception e)
      {
         System.out.println("Exception in main = " +  e);
      }
   }
}

To invoke the deep copy (line A), execute java.exe Driver1 deep. When the deep copy runs, we get the following printout:

Original = [java.awt.Point[x=1,y=1]]
New      = [java.awt.Point[x=1,y=1]]
Original = [java.awt.Point[x=2,y=2]]
New      = [java.awt.Point[x=1,y=1]]

This shows that when the original Point, p1, was changed, the new Point created as a result of the deep copy remained unaffected, since the entire graph was copied. For comparison, invoke the shallow copy (line B) by executing java.exe Driver1 shallow. When the shallow copy runs, we get the following printout:

Original = [java.awt.Point[x=1,y=1]]
New      = [java.awt.Point[x=1,y=1]]
Original = [java.awt.Point[x=2,y=2]]
New      = [java.awt.Point[x=2,y=2]]

This shows that when the original Point was changed, the new Point was changed as well. This is due to the fact that the shallow copy makes copies only of the references, and not of the objects to which they refer. This is a very simple example, but I think it illustrates the, um, point.

Implementation issues

Now that I've preached about all of the virtues of deep copy using serialization, let's look at some things to watch out for.

The first problematic case is a class that is not serializable and that cannot be edited. This could happen, for example, if you're using a third-party class that doesn't come with the source code. In this case you can extend it, make the extended class implement Serializable, add any (or all) necessary constructors that just call the associated superconstructor, and use this new class everywhere you did the old one (here is an example of this).

This may seem like a lot of work, but, unless the original class's clone() method implements deep copy, you will be doing something similar in order to override its clone() method anyway.

The next issue is the runtime speed of this technique. As you can imagine, creating a socket, serializing an object, passing it through the socket, and then deserializing it is slow compared to calling methods in existing objects. Here is some source code that measures the time it takes to do both deep copy methods (via serialization and clone()) on some simple classes, and produces benchmarks for different numbers of iterations. The results, shown in milliseconds, are in the table below:

As you can see, there is a large difference in performance. If the code you are writing is performance-critical, then you may have to bite the bullet and hand-code a deep copy. If you have a complex graph and are given one day to implement a deep copy, and the code will be run as a batch job at one in the morning on Sundays, then this technique gives you another option to consider.

Another issue is dealing with the case of a class whose objects' instances within a virtual machine must be controlled. This is a special case of the Singleton pattern, in which a class has only one object within a VM. As discussed above, when you serialize an object, you create a totally new object that will not be unique. To get around this default behavior you can use the readResolve() method to force the stream to return an appropriate object rather than the one that was serialized. In this particular case, the appropriate object is the same one that was serialized. Here is an example of how to implement the readResolve() method. You can find out more about readResolve() as well as other serialization details at Sun's Web site dedicated to the Java Object Serialization Specification (see Resources).

One last gotcha to watch out for is the case of transient variables. If a variable is marked as transient, then it will not be serialized, and therefore it and its graph will not be copied. Instead, the value of the transient variable in the new object will be the Java language defaults (null, false, and zero). There will be no compiletime or runtime errors, which can result in behavior that is hard to debug. Just being aware of this can save a lot of time.

The deep copy technique can save a programmer many hours of work but can cause the problems described above. As always, be sure to weigh the advantages and disadvantages before deciding which method to use.

Conclusion

Implementing deep copy of a complex object graph can be a difficult task. The technique shown above is a simple alternative to the conventional procedure of overwriting the clone() method for every object in the graph.

About the author

Dave Miller is a senior architect with the consulting firm Javelin Technology, where he works on Java and Internet applications. He has worked for companies such as Hughes, IBM, Nortel, and MCIWorldcom on object-oriented projects, and has worked exclusively with Java for the past three years.