Custom Implementations

Many programmers will never need to implement their own collections classes. You can go pretty far using the implementations described in the previous sections of this appendix. Someday, however, you might want to write your own implementation of a core collection interface.

Reasons to Write Your Own Implementation

The following list of kinds of collections you might implement is not intended to be exhaustive.

• Persistent: All of the built-in collection implementations reside in main memory and vanish when the VM exits. If you want a collection that will still be present the next time the VM starts, you can implement it by building a veneer over an external database. Such a collection might conceivably be concurrently accessible by multiple VMs, because it resides outside the VM.
• Application specific: This is a very broad category. One example is an unmodifiable Map containing real-time telemetry data. The keys might represent locations, and the values could be read from sensors at these locations in response to the get operation.
• Highly concurrent: The built-in collections are not designed to support high concurrency. The synchronization wrappers (and the legacy implementations) lock the entire collection every time it's accessed. Suppose that you're building a server and need a Map implementation that can be accessed by many threads concurrently. It is reasonably straightforward to build a hash table that locks each bucket separately, allowing multiple threads to access the table concurrently, assuming that they're accessing keys that hash to different buckets.
• High performance, special purpose: Many data structures take advantage of restricted usage to offer better performance than is possible with general-purpose implementations. For example, consider a Set whose elements are restricted to a small, fixed universe. Such a Set can be represented as a bit-vector, which offers blinding fast performance as well as low memory usage. Another example concerns a List containing long runs of identical element values. Such lists, which occur frequently in text processing, can be run-length encoded; runs can be represented as a single object containing the repeated element and the number of consecutive repetitions. This example is interesting because it trades off two aspects of performance: It requires far less space but more time than an ArrayList.
• High performance, general purpose: The engineers who designed the Collections Framework tried to provide the best general-purpose implementations for each interface, but many, many data structures could have been used, and new ones are invented every day. Maybe you can come up with something faster!
• Enhanced functionality: Suppose that you need a Map or a Set implementation that offers constant time access and insertion-order iteration. This combination can be achieved with a hash table, all of whose elements are further joined, in insertion order, into a doubly linked list. Alternatively, suppose that you need an efficient bag implementation (also known as a multiset): a Collection that offers constant time access while allowing duplicate elements. It's reasonably straightforward to implement such a collection atop a HashMap.
• Convenience: You may want additional convenience implementations beyond those offered by the Java platform. For instance, you may have a frequent need for immutable Map objects representing a single key-value mapping or List objects representing a contiguous range of Integers.
• Adapter: Suppose that you are using a legacy API that has its own ad hoc collections API. You can write an adapter implementation that permits these collections to operate in the Java Collections Framework. An adapter implementation is a thin veneer that wraps objects of one type and makes them behave like objects of another type, by translating operations on the latter type into operations on the former.

How to Write a Custom Implementation

Writing a custom implementation is surprisingly easy with the aid of the abstract implementations furnished by the Java platform. Abstract implementations, skeletal implementations of the core collection interfaces, are designed expressly to facilitate custom implementations. We'll start with an example, an implementation of Arrays.asList:

public static List asList(Object[] a) { return new ArrayList(a); } private static class ArrayList extends AbstractList implements java.io.Serializable { private Object[] a; ArrayList(Object[] array) { a = array; } public Object get(int index) { return a[index]; } public Object set(int index, Object element) { Object oldValue = a[index]; a[index] = element; return oldValue; } public int size() { return a.length; } }

Believe it or not, this is almost exactly the implementation contained in the Java 2 SDK. It's that simple! You provide a constructor and the get, set, and size methods, and AbstractList does all the rest. You get the ListIterator, bulk operations, search operations, hash code computation, comparison, and string representation for free.

Suppose that you want to make the implementation a bit faster. The API documentation for the abstract implementations describes precisely how each method is implemented, so you'll know which methods to override in order to get the performance you want. The performance of the preceding implementation is fine, but it can be improved a bit. In particular, the toArray method iterates over the List, copying one element at a time. Given the internal representation, it's a lot faster and more sensible just to clone the array:

public Object[] toArray() { return (Object[]) a.clone(); }

With the addition of this override and a similar one for toArray(Object[]), this implementation is exactly the one found in the Java 2 platform. In the interests of full disclosure, it's a bit tougher to use the other abstract implementations, because they require you to write your own iterator, but it's still not that difficult.

The abstract implementations can be summarized as follows:

• AbstractCollection: A Collection, such as a bag, that is neither a Set nor a List. At a minimum, you must provide the iterator and the size method.
• AbstractSet: A Set. Its use is identical to AbstractCollection.
• AbstractList: A List backed by a random-access data store, such as an array. At a minimum, you must provide the positional access methods (get(int) and, optionally, set(int), remove(int), and add(int)) and the size method. The abstract class takes care of listIterator (and iterator).
• AbstractSequentialList: A List backed by a sequential-access data store, such as a linked list. At a minimum, you must provide the listIterator and the size methods. The abstract class takes care of the positional access methods. (This is the opposite of AbstractList.)
• AbstractMap: A Map. At a minimum, you must provide the entrySet view. This is typically implemented with the AbstractSet class. If the Map is modifiable, you must also provide the put method.

The process of writing a custom implementation follows.

1. Choose the appropriate abstract implementation class from the preceding list.
2. Provide implementations for all the class's abstract methods. If your custom collection is to be modifiable, you'll have to override one or more concrete methods as well. The API documentation for the abstract implementation class will tell you which methods to override.
3. Test and, if necessary, debug the implementation. You now have a working custom collection implementation!
4. If you're concerned about performance, read the abstract implementation class's API documentation for all the methods whose implementations you're inheriting. If any of them seem too slow, override them. If you override any methods, be sure to measure the performance of the method before and after the override! How much effort you put into tweaking the performance should be a function of how much use the implementation will get and how performance-critical the use. (Often this step is best omitted.)