24 August 2004

I often get questions that surround one of the design choices in the JUnit testing framework - the decision to make a new object for each test method run. Enough to warrant a quick bliki entry. (However I feel almost compelled to point out that my writing about JUnit does not mean that that I don't think that other forms of testing are important. There are lots of useful testing activities, and although JUnit and its cousins are valuable for many of them it isn't the solution for everything. For more blogging on testing I suggest you look at the blogs of Brett Pettichord, Brian Marick, and James Bach. You should also not assume that my writing about xUnit testing implies suggests the unimportance of refactoring, use cases, or flossing.)

Consider the following little Java test class

import junit.framework.*;
import java.util.*;

public class Tester extends TestCase {
  public Tester(String name) {super(name);}
  private List list = new ArrayList();
  public void testFirst() {
    assertEquals(1, list.size());
  public void testSecond() {
    assertEquals(0, list.size());

Some people may not realize this, but both tests pass - and will pass in whichever order they are run. This is the case because to run this JUnit creates two instances of Tester, one for each testXXX method. The list field is thus freshly initialized for test method run. Now some people think this is a bug in JUnit, but it isn't - it's a conscious design decision. (For more on this kind of thing watch out for Kent's new book.)

The basic design of JUnit has its origins in a testing framework that Kent Beck built in Smalltalk. (Actually to call it a framework was a bit of a misnomer - Kent never shipped it out as a framework. He preferred people to build it themselves since it would only take an hour to two - that way they wouldn't be afraid to change it when they wanted something different.) One of the key principles in JUnit is that of isolation - that is no test should ever do anything that would cause other tests to fail.

Isolation provides several advantages.

  • Any combination of tests can be run in any order with the same results.
  • You never have a situation where you're trying to figure out why one test failed and the cause is due to the way another test is written.
  • If one test fails, you don't have to worry about it leaving debris that will cause other tests to fail. That helps prevent cascading errors that hides the real bug.

Now JUnit provides other mechanisms that support isolation - in particular the setUp and tearDown methods that are run at the beginning and end of each test method. To use this for my simple example you do this.

  public void setUp() {
    list = new ArrayList();

Most of the time you don't need to use tearDown since the setUp can do any reinitialization that you need.

You could isolate your test methods by having all the state be in local variables and not use fields at all. However this would mean duplicating your setUp code in every test - and you know how much I despise duplication.

Critics of the JUnit approach have argued that since you have setUp and tearDown you don't need a fresh object each time. You can just make sure you reinitialize all your fields in those methods. Fans of the JUnit approach argue that this may be true, but many people initialize in fields, and you might as well provide this greater degree of isolation. After all an important part of framework design is to make it easy to the right thing (isolation) and hard (but not impossible) to do things that cause problems. After all what's the cost of doing it?

The main argument about the cost of the JUnit approach is based on the extra objects that are created, both the JUnit test cases and all the other objects created in the setup and field initializers. Most of the time I think this argument is bogus. There's a lot of fear about creating lots of objects, but most of the time it isn't justified - it's based on an outdated mental model of how object allocation and collection work. Certainly there are environments where object creation could be an issue and Java was one them in its early days. However modern Java can create objects with virtually no overhead, it's no longer an issue. (It wasn't in Smalltalk for longer, which is why Kent and Erich didn't worry about it.) So most of the time just don't worry about creating objects.

That said, mostly doesn't mean alwaysly. One good example of an object you don't want to create frequently is a database connection. This does make sense to share, but sharing across all the test methods in one test case class isn't enough - you'll want to share it across much more than that. A cheap and nasty way to do this is with static variables. Generally it's wise to shy away from statics, but often they're fine in the context of a test run - although I still prefer to shun them. JUnit actually provides a very flexible mechanism for sharing test fixture objects - the TestSetup decorator. This allows you to setup some common state for any test suite, which gives you a lot of more flexibility about sharing state across groups of tests - much more so than just sharing across the methods in a single test case class.

Perhaps the biggest problem with TestSetup is that finding information on it is so hard that that I almost expected to see "beware of the leopard" in the documentation. And there is a leopard around - if you use TestSetup you are breaking isolation and broken isolation tends to lead to awkward to find bugs. Don't use it unless you really, really need it. (But if you do this forum thread gives you some hints on using it, as does J.B. Rainsberger's new book)

(All of this may make you wonder why each test method isn't in its own class. Indeed the earliest forms of JUnit did do that, using an inner class that subclassed the test case with the fixture. While this was a more obvious design, it made it harder to write the tests. So they went for the more obscure use of the pluggable selector pattern.)

A second objection to the JUnit approach is that it isn't intuitive - in that the mechanism it uses to pull this off is tricky to understand. I sympathize with this, the Pluggable Selector pattern isn't well known, and a design style that uses unfamiliar patterns is often uncomfortable. On the whole I like the JUnit approach because I think the isolation and ease of test writing outweighs the esoteric implementation.

But good people disagree with me. Cedric Beust's TestNG doesn't do this, perhaps more surprisingly the popular NUnit implementation doesn't do it (although Jim now regrets that decision). The following NUnit test causes a failure.

  public class ServerTester
    private IList list = new ArrayList();
    public void first() {
      Assert.AreEqual(1, list.Count);
    public void second() {
      Assert.AreEqual(0, list.Count);

If you're using a framework that works in this style, I strongly recommend you initialize all your instance variables in a setup method. That way you'll keep your tests isolated, and avoid some debug induced hair removal.

I don't happen to agree with reusing the test case instance - but I don't think those that made that decision have a single-digit IQ, have some complex financial killing in mind, or are embarking on some strange behavior with their lower torsos. They called the design trade-off differently - and I think life is better when we can respectfully disagree over the fluid nature of software design.