Introduction to Unit Testing Part 4: Design New Code For Testability

In the last post in this series, I covered what I think of as an important yet seldom discussed subject: how not to overwhelm yourself and get discouraged when you’re starting to unit test. In the post before that, I showed the basics of writing a unit test. But this leaves something of a gap. You can now write a unit test in a vacuum for an extremely simple class, and, when looking at your legacy code base, you know what to avoid. But you don’t necessarily how to write a non-trivial application with unit tests.

You might understand now how to write tests, assuming that you have some disconnected new class in your code base without dependencies and barriers to testing, but you wonder how to get to that point in the first place. I mean, your application doesn’t seem to need a prime number finder or a bowling score calculator. It needs you to add lines of code to existing methods or new methods to existing classes. It doesn’t really seem to need new classes, so the initial momentum and resolution you’ve built reading these first three posts to go and be a unit tester sort of fizzles anticlimactically when you stare at your code base.

What’s going on here?

Recognizing Inhospitable Terrain

The first thing to understand is that your code base probably wasn’t written with testability in mind. There’s nothing wrong with you for not being able to see where unit testing fits in, because it doesn’t. Last time I talked about things you’ll see that will torpedo your efforts to test a particular method or piece of code, but let me speak to some entire architectures and patterns that don’t lend themselves to testability. It’s not that code written with these technologies and patterns can’t be tested–it’s just that it won’t be easy for you. As you read through this, you might feel like I’ve read your code base’s mind.

  1. Active Record as an architectural pattern
    This is a pattern in which you create classes that are in-memory representations of database tables, views, or stored procedures. If you see in your code base a class called “Customer” that has methods like “GetById()”, “Update()” and “MoveNext()” you’ve got yourself an Active Record architecture. This architecture tightly couples your database to your domain logic and your domain logic to the rules for navigating through domain objects. You can’t test any of these objects since any operation you perform on them sends them scurrying off to create database connections and parameterized queries and all manner of other untestable stuff. And since decomposition and decoupling is the path toward unit testing, this sort of tight coupling of everything in your code is the path away from it.
  2. Winforms
    Winforms in the .NET world are tried and true when it comes to rapidly cranking out functional little applications, but you have to work really, really hard to make code that uses them testable. Q&A sites are littered with people trying to understand how to make Winforms testable, which should tell you that making them testable is not trivial. If you have Winforms and Active Record both in the same code base, at least the architecture is split into two concerns. But it’s split into two thoroughly untestable ones.
  3. Webforms
    See Winforms. Webforms is very similar in terms of framework testability, and for pretty similar reasons. Webforms is arguably even harder to test, however, because it’s predicated on spewing out reams and reams of HTML, CSS, and Javascript while allowing you to pretend you’re writing a desktop app. I’ve talked about my opinion of this technology before.
  4. Wizard/markup-reliant code
    Do you use the Webforms grid wizard thing to generate your grids? Do you define object data sources, such as DB connections or files, in the markup? Practices like these are the epitome of quick and dirty, rapid-prototyping implementations that hopelessly cross couple your applications beyond all testability. If this is something that’s done in your group/code base, testing is basically a non-starter until you go in a different architectural direction.
  5. Everything in your application is in a user control/form
    I’ve seen this called “Smart UI” and it basically means that there’s absolutely no separation of concerns in your code. The UI elements create database connections, write to files, implement business rules–they do everything. Code like this is impossible to unit test.

If any of this is sounding familiar, your task might be daunting. I have my own preferences, but I’m trying not to offer a value judgment here as much as I’m letting you know what you’re up against. I’m like a mortgage broker that’s saying to you, “if you want to own a home, that’s a great goal. But if you are eight months behind on your rent and have no personal savings, you’re going to have some work to do first.” If I’ve described your code base in the list above, you face different challenges than a green field developer. And since you’ve presumably been contributing to these code bases, you’re probably very used to implementation techniques that don’t result in testable code. You’re going to need to change your thinking and your coding practices in order to start writing testable code.

Once we’ve discussed how to get you writing testable code, I’ll come back to these macroscopic concerns and give some pointers for how to improve the situation. But, for now, on to a revised approach to coding. The following holds true whether you’re banging away at some legacy Winforms/Active Record application or starting a brand new MVC 4 site.

Add New Methods and Classes First, Ask Questions Later

First thing to abide by is to favor adding new things to the code over modifying existing things in the code. You may have heard this before in the context of the “Open/Closed Principle”, but that’s a guide for how to write your classes. (Basically, it admonishes you to write classes that others can extend and override rather than change.) I’m talking about how to deal with existing code bases. To put it simply and bluntly, it’s a lot easier to both code and test brand spankin’ new classes than to write and test changes to existing ones. We all know this. It’s at the heart of why we as developers always lean toward rewriting others’ code instead of understanding and working with it.

Now, this might not win you friends. In shops where people tend to write procedural code (you know, the kind you’re trying to get away from writing), they seem to have some weird fear of creating too many classes and masochistic attachment to monolithic structures. You might have to compromise or practice on your own if you run afoul of the project’s architect, but the exercise is invaluable. It’s going to propel you toward decoupling as a default rather than an exception. Doing new things? Time for a new class and some unit tests.

I know what you’re thinking: “But what if the thing that needs to be done has to be done in the middle of some method somewhere?” Well, instantiate your new class at that point and use it. “But what if it needs a bunch of fields from the class it’s in and variables from the method?” Pass them in through the constructor or method call. “But won’t that make my design bad?” It already is bad, but at least now you’re making part of it testable. When your code is testable and under test, everything is easier to fix later.

You’ll have to use some discretion, obviously, but shift your attitude here. Don’t look at a project that has nothing but .aspx files and their code behind and hide classes in there for fear of breaking with tradition. Boldly add pure .cs files to the code base. Add unit tests for those .cs files you’re creating. (Apologies to Java readers, but this has no real Java equivalent that I can think of having used. Plus, the Java stack in general seems not to have the same level of untestable cruft built into frameworks.) It’s a lot harder for someone, even the project architect, to give you a hard time if your new way of doing things is covered by unit tests. Even if they’re hostile Expert Beginners, they’re hard pressed not to sound silly if they say, “we don’t do that here.” You’ll at least have a better chance of pushing this change through with the unit tests than without them.

Ask Questions in the Right Order: What, How, When

Now that you’re creating a lot of new classes and instantiating them in the old untestable ones, it’s time to start working on what kind of code you write. If you’ve practiced and come back, I suspect that you’re starting to be able to write a few useful tests but are perhaps still struggling. And I bet it’s because the line between where the old class ends and your new one begins is a little hazy. Maybe they share some common fields. Maybe when you instantiate the class you’re testing, you hand it a “this” reference so that it can go picking through the properties on the untestable behemoth from which you’re escaping. This is the next thing we need to tighten up–stop doing that. A clear, concise division of labor between the classes is necessary, and it’s not possible if they share all of the same fields, properties, state, etc. That’s like a break up where you two continue to live together, share a car, and go to the movies on weekends.

The best way to achieve this clean split is with good abstraction, and the best way to do that is to remember “what, how, when.” When it’s time to change the code base, remember that you want to favor creating a new class. But before you do that, ask yourself “what?” but not the other two questions. What should I name this class? What should it do? What should it expose as its public methods? Don’t start thinking about how those methods or classes should work and don’t you dare start thinking about when anything at all should happen–just think about what. Give it a good name that defines a clear purpose, and then give it a good set of methods and properties that draw attention to why it’s a different concept than the class that will be using it. Ask yourself what the boundaries between the two classes will be so that you can minimize the amount of shared information. And now, start stubbing out those methods with no implementation and start stubbing out some test methods with names that say what the methods will do.

At this point, you’re ready for “how.” Start actually implementing the “what” and testing that your implementation works in the unit tests. Believe me, it’s much, much easier to implement methods this way. If you think about “what” and “how” at the same time, you start writing confusing code that not even you have faith in when you’re done. Implementation alone is much easier when you have a clear picture of “what,” and unit testing is a breeze.

Once everything is implemented, you can start thinking about “when.” When should you instantiate the class and when should you call its methods? But don’t spend too much time with “when” in your head because it’s dangerous. Does that sound weird? Let me explain.

“When Code” is Monolithic Code

Picture two methods. One is a 700 line juggernaut with more control flow statements than you can keep track of without a spreadsheet. The other is five lines long, consisting only of an initialize statement, a foreach, and a return statement. With which would you rather work? I imagine the response is unanimous here. Even if you tend to crank out these kinds of large methods, when you step through the debugger, looking for the cause of a bug and finding yourself in some huge method, your heart sinks and you settle in with snacks and caffeine because it’s going to be a long day.

Now with these two methods in mind, imagine if we were pair programming together and I simply asked “when?” With the tiny method, you’d probably say, “What do you mean by ‘when’–I mean, you initialize before the loop and you return when you find the record you’re looking for. What a weird question!” With the other method, you’d probably affect a thousand-yard stare and say, “Man, I don’t even know where to begin.” But the answer to that question would fill pages. Books. Because you set the first loop counter j equal to the third loop counter k about twenty lines before the fourth try-catch and fifteen lines before you set the middle loop counter j equal to four. Unless, of course, you threw that exception up on line 2090, in which case j might never have been initialized. Er, wait, I think that happened somewhere near the fifth while loop in that else condition up there. Oh, there’s so much “when,” but it’s all slammed together in a method where you can’t possibly test any of it. Lots of thinking about “when” breeds huge methods like a Petri dish for bacteria.

“When” code is procedural at its core, and procedural, “when” code is an anathema to object-oriented unit testing, which is all about “what” and “how.” Remember earlier in the series when I said that multi-threaded code was really, really hard to test? Well, that’s just a subset of an idea called “temporal coupling,” and what we’re talking about here also falls under that umbrella. Temporal coupling is what happens when things have to be executed in a specific order or else they do not work.

Imagine that you’re coding up a model for someone’s day. When you think of how to do this, do you think, “first he gets up, then he brushes his teeth, then he showers, then he puts on his clothes, then… then he comes home, then he eats dinner, then he watches TV, then he goes to bed?” Do you code this up with constructs like:

public void GoAboutMyDay()
{
    bool wokeUp = WakeUp();
    if (wokeUp == true)
    {
        bool showered = Shower();
        if (showered == true)
        {
            bool putOnClothes = PutOnClothes();
            if(putOnClothes)
                //You get the idea
        }
    }
}

This method is all about “when.” It’s entirely procedural, and it’s going to be horrifying when it’s complete. When you get into the fortieth nested if condition, maybe someone will come along and flatten it out with a bunch of inverted early returns. Or maybe not, because maybe some of if clauses start sprouting else conditions with loops in them. And maybe the methods being called start communicating with one another via boolean flag fields in the class. Who knows–this thing is on the precipice of becoming unstoppable. It might just achieve sentience at some point, so that when you try to start deleting conditionals, it says, “I can’t let you do that, Dave,” and puts them back.

The root problem behind it all is the “when” and the procedural thinking because you’re orienting the implementation around the order of the activities rather than the nature of the activities. Unit testing is all about deconstructing things into their smallest possible chunks and asserting things about those chunks. Temporal coupling and “when” logic is all about chaining and fusing things together.

If you were thinking about “what” first here, you would form a much different mental model of a person’s day. You’d say things to yourself like, “well, during the course of a person’s day, he probably wakes up, gets dressed, eats breakfast–well, actually eats one or more meals–maybe works if it’s a weekday, goes to bed at some point,” etc. Whereas in the procedural “when” modeling you were necessarily building a juggernaut method, here you’re dreaming up the names of methods and/or classes that can be unit tested separately and in isolation. It’s no reach to say, “okay, let’s have a Meal class that will have the following methods…”

Only at the end will you decide “when.” You’ll decide it after you’ve stubbed things out with the “what” and implemented/unit tested them with the “how.” “When” is a detail that you should allow yourself to figure out at any point down the line. If you nail down what and how, you will have testable, modular, and manageable code as you create your classes.

Other Design Considerations for Your New Classes

I’ll wrap up here with a few additional tips for creating testable designs when adding code-to-code bases:

  1. Avoid using fields to communicate between your methods by setting flags and tracking state. Favor having methods that can be executed at any time and in any order.
  2. Don’t instantiate things in your constructor. Favor passing them in (we’ll talk about this in detail in a future post in the series).
  3. Similarly, don’t have a lot of code or do a lot of work in your constructor. This will make your class painful to setup for test.
  4. In your methods, accept parameters that are as decomposed as possible. For instance, don’t accept a Customer object if all you do with it is read its SSN property. In that case, just ask for the SSN.
  5. Avoid writing public static methods. These are easy enough to test (often), but they start introducing testability problems when you write code that uses them. (This might be hard to swallow at first, but mull over the idea of simply not using static methods anymore.)
  6. The earlier you start writing your unit tests, the better. If you find that you’re having a hard time testing your new code, it’s more likely a problem with the code than with unit testing it, and if you write tests early, you’ll discover these problems before you get too far and fix them.

This post has covered ways to write unit tests “from here forward” and ways to stop adding untested code to code bases. In the next post, I’ll talk about how to start getting the legacy code under test.

Addendum: Mitigating the Hostile Test Environments

Finally, as promised, here are ways to accommodate testing in the less-than-ideal architectures mentioned above, if you’re curious or want to do some more research:

  1. Instead of Active Record, look at some kind of ORM solution like NHibernate or Entity Framework. These are tools that generate all of the code for you to access the database so that you don’t have to worry about testing that code and you can focus on writing only your (testable) domain code. Barring that, try to separate the three concerns of Active Record objects: modeling the database, connecting to the database, and modeling a domain object. The first concern adds no value, and the second two can be broken out into separate objects where the only thing hard to test is the actual database access.
  2. Instead of Winforms, favor WPF when possible. If that isn’t possible, see if you can use the Model-View-Presenter (MVP) pattern to move as much logic out of the untestable code-behind as possible.
  3. To be blunt, from a testing/decoupling perspective, Webforms is a disaster. You can have some limited success by adopting a more passive binding model and moving as much code out of the code-behind as possible, but it’s all pretty awkward. Webforms really seems more about rapid-prototyping and Microsoft-Accessing web development than producing scalable, sophisticated architectures.
  4. If you’re using wizards to generate your application’s architecture, cut it out. If you’re defining implementation details in markup, cut it out. Markup is for layout, not unit-testable business logic or state logic. If you depend on definitions in markup to drive your application’s behavior, you’re relying exorbitantly on a third-party framework, which is always extremely brittle from a testability perspective.
  5. To fix Smart UI, you just have to factor toward a more decoupled architecture. Start pulling different concerns out of the user controls and forms and finding a home for them.