Legacy Code Retreat part one: get it under test

Friday 3 August 2012 at 14:11 BST

It's been a while, but people have convinced me that I need to crank this one out.

A little while ago, Sandro and I ran the UK's first Legacy Code Retreat. Essentially, it works like a regular code retreat, where people come together to solve a problem again and again in a multitude of different ways. The difference with this one is simple: we don't start from scratch. You don't implement Conway's Game of Life over and over again, but instead start with a terrible piece of code. Your task is to understand it and clean it up.

Cleaning, of course, starts with testing.

Session one: understand the code. Just read it. Don't change it. Don't start refactoring. Please don't start fixing it. Just find out what it does.

Know what it does? Alright, prove it. Start writing tests. Write tests until a method is completely covered. If any of them go red, your assumptions were wrong. But that's OK, just correct them. When you're covered and green, sit back and smile. Session one finished five minutes ago. Everyone's staring at you.

What was hard? Did you feel the urge to start changing the code? Yes? Good. You're normal. We all do. The trick is to not give in to instinct.

Right. Onto session two. Session two is a bit different. Testing's too easy. Let's make something that makes the test. We're going to create a golden master.

The trivia project is a nice one with which to do this, as all the output is dependent on a single random number generator. And while the numbers are pseudorandom, we can control which branch of the randomness comes out simply by providing a seed. By a small change to the application, you can pass the random number seed in through the command line rather than using the default seed (which is usually the current time). Something like this:

public static void main(String[] args) {
    Random random = args.length >= 1 ? new Random(Integer.parseInt(args[0])) : new Random();
    ...
}

Now we can generate a ton of test cases simply by varying the input and capturing the output. By storing the input and output in files, we can write a simple test that iterates through all of them, runs the application again and asserts that the output has not changed. We have a massive suite of regression tests, which we can use to defend against unwanted changes.

This one was a tricky one to grasp for a lot of people, including me the first time I tried it. We had a bit of trouble explaining random number seeds, and some people were scratching their heads over the large number of tests. This just requires a bit of one-on-one chatting when people get stuck.

It's important to note that the golden master is not a replacement for unit tests. Rather, it gives you a safety net. The tests will tell you that something changed, but not what. They also won't tell you whether it should have changed—you need to decide that for yourself. What you really have here are a lot of very slow, crap tests. This is better than no tests at all, but not by much. The next thing to do is to go back to what we were doing in session one: writing unit tests to cover code. Only then can we refactor, safe in the knowledge that if we screw up, our unit tests will probably catch the problem, and if not, we have the golden master tests sitting there, ready.

Our third session brought in another technique that has been dubbed "subclass to test". The idea is to identify what Michael Feathers calls a seam: an area of code you can modify without changing the actual code itself. In this case, we can modify a method for testing by subclassing and overriding the specific method. Combined with extracting methods, this is a very powerful tool for getting code under test. For example, if I have some code that uses a static factory to get a reference to the database:

public User createUser() {
    User user = new User();
    Database.getDatabase().save(user);
    return user;
}

We can yank that out into a method, using automated refactoring tools to do so if we have them available.

public class UserMaker {
    public User createUser() {
        User user = new User();
        database().save(user);
        return user;
    }

    protected Database database() {
        return Database.getDatabase();
    }
}

Now we have a seam. We can override that in tests to return a mock so we don't have to touch the database.

public class UserMakerTest {
    private final Database database = mock(Database.class);

    ...

    private static class UserMakerForTesting extends UserMaker {
        @Override
        protected Database database() {
            return database;
        }
    }
}

Look at that. Previously untestable code is now testable. Easy, right?

This was the first half of the day. I think it went pretty well, and the feedback was great. In my next post, I'll be tackling the second half: fixing the code base.

Part two of the Legacy Code Retreat double act is now available. I hope you enjoy it.