In Defense of 100% Unit Test Coverage

Common reactions to proposed test coverage requirements.

original

Preamble

In this article, I’ll bring up three arguments in favor of 100% test coverage, three common arguments against it, and a few testing protips.

Arguments in favor:

1. Encouraging testable components
2. Encouraging collaboration
3. Fringe benefit: 100% unit test coverage

Arguments against:

1. Marginal benefit increases
2. Barrier to entry
3. Misleading priorities

I want to make two things very clear first:

All code, given time, is bad code

The world of code is constantly changing. It would be egotistical to believe our code perfect, as that would imply we’re done learning or changing. I’m incredibly grateful to the people who sat me down and told me exactly (and politely) why what I was writing was rubbish.

Many articles — even this one — with opinions about the Correct Way To Write Good Code™ make it sound like some previous works are bad code. Maybe they were then; maybe they are now. So what? What’s important is to learn from your mistakes, successes, and peers.

Code coverage is not a silver bullet

There is no single Correct Way To Write Good Code™. Good techniques help, but hyperfocusing on a single methodology will only help you adhere to that singular narrow line of thought. As the agile manifesto recommends:

Individuals and interactions over processes and tools.

That is to say, tools and processes are important, but it is more important to have competent people working together effectively.

Still, we shouldn’t discount a process or tool purely for not doing enough on its own. I don’t skip a thick scarf in the winter for not keeping me warm without a coat (it completes my outfit!). Code coverage is just one potential way to help keep your code tested.

Non-Goals

I don’t want to discuss whether we should unit test instead of higher-level tests or whether we should write tests in the first place. If you’re on a project that you truly believe doesn’t need to be unit tested, fine: this post isn’t for that project. If you don’t believe in unit tests in particular, this post definitely isn’t for you. If you don’t believe in any tests, either you have never written a large, persistent project or you enjoy living in constant danger.

What it's like to work in a large project without tests.

original

For a full discussion of how to properly write unit tests, read Roy Osherove’s The Art of Unit Testing. It’ll make you a better developer and a better person.

Arguments For

Encouraging Testable Components

If your code is well written, it likely follows good coding principles such as SOLID. Your classes and functions should have single responsibilities; the implementation details of dependencies should be abstracted away by their interfaces; changes to one component’s behavior shouldn’t require obscure changes to the internals of others.

Code that violates good coding principles likely violates good testability too: god classes are inherently difficult to capture all cases of; depending on implementation details of dependencies necessitates awkwardly implementation-specific tests… But does the reverse hold true? Does writing testable code imply adherence to SOLID? Surely there must be some overlap!

If high coverage encourages well-separated components, and well-separated components are much easier with SOLID principles, then it follows that requiring high coverage is a way of reinforcing SOLID.

However you define code testability, your core principles for unit tests in particular should at least require the tests be small and single-minded. Otherwise they’re not unit tests! Likewise for limiting external dependencies: tests that require external dependencies or many components likely aren’t unit tests.

We should note that complete coverage doesn’t enforce testable designs. There will always be ways for developers to write yucky god classes (see “Misleading Priorities” below). Rather, complete coverage makes it more difficult to write and maintain those classes by increasing the roadblocks for any change that doesn’t respect SOLID. It’s natural selection, except instead of starving or being hunted to death, your code gets more and more frustrating to write. If you’re having difficulties implementing 100% unit test coverage, you might just be exposing structural problems in your code.

Conclusion 1: If your code isn’t testable, it’s probably not great to begin with.

Encouraging Collaboration

Every junior developer I’ve ever seen join a team — me included — has started off their first feature with a giant 500+ line monstrosity of a class. Following the previous section, we know this to be bad. We must have a way of stopping these people from doing those terrible things, right?

The answer is you and me. Great code is inherently difficult: folks new to code or your frameworks will likely need help Doing the Right Thing. Requiring all code be in this ideal testable form forces developers to reach out when they face issues. It’s your job as someone who has read this post (ha) to help guide them towards the light. By helping them, you’re increasing both their ability to work in your environment and the quality of your code.

A necessary caveat for this point is to remember that we’ve all been junior developers. Don’t go blazing into code reviews with hellfire and brimstone: be kind and nurture the new people. Nothing kills professional motivation or feeds the imposter syndrome quite like a senior developer or scrappy youngster telling you you’re terrible.

Conclusion 2: A rising tide lifts all boats.

Fringe benefit: 100% unit test coverage

Requiring unit test coverage for everything at the very least ensures that you try to unit test everything. Nothing can truly validate that code is doing what’s intended like tests can. Team members leave; designs change; documentation becomes outdated and discarded. Untested code is at least as likely to break as tested code.

From another angle: unit tests provide a combination safety net & sample documentation source for your code. When looking at a section of code you haven’t seen before, unit tests will always show at least one way of interacting with it. When you progress to changing that section, the unit tests will validate your changes haven’t broken anything. 100% coverage helps enforce you add tests to cover most or all of your sections.

From a third angle: if you discover a bug and want to add a unit test for it, the only way to guarantee this is possible is to guarantee all code paths are at least reachable by unit tests. 100% coverage is a good way to do that.

Conclusion 3: Test all your code.

Arguments Against

Marginal benefit increases

The Pareto Principle states that roughly 80% of effects come from 20% of causes. It’s often applied to coding in that 80% of work comes from 20% of tasks. We tend to spend a small amount of time writing what we want to and most of the rest banging our faces into the keyboard over bugs or edge cases.

Similarly, we can capture most of a system’s behavior with a small number of tests, but expanding to all edge cases dramatically increases the amount of test work. The amount of work to increase test coverage to 100% from 80% is much more than to 80% from 0% (and, likely, the same holds true for going to 100% from 95% compared to 95% from 80%). Some consider the extra work for complete coverage not worth the system improvements.

Barrier to entry

Every code requirement, from unit tests to strict compiler flags to formatting rules, makes it more difficult to contribute. Trying to stamp out imperfections inherently makes it more difficult to achieve the minimum code quality bar. Developer fatigue can easily be exacerbated by adhering to seemingly arbitrary and near-useless extra rules.

For example, I don’t enforce 100% code coverage in most of my open source projects. Some are rapid prototypes where significant tests would be cumbersome. In others, most of the contributors are college students who don’t know what a proper unit test is, let alone how to write one that follows SOLID principles.

Misleading priorities

I’ve seen a lot of junior developers jump into test coverage debates using only this argument in favor of 100% coverage: “It makes sure we test everything!“… I honestly don’t consider it reason enough alone to justify such a hard requirement. Just because you run a line of code doesn’t mean it’s tested. Tests could be only coincidentally running that line but really testing something else. 100% unit test coverage alone does not enforce good tests, nor testable designs, nor good code in general. That’s your responsibility as code authors and reviewers.

Rebuttal?

🤷‍ Up to you.

Every project is a unique blend of conflicting shareholder and developer desires. Use critical thinking and be willing to change. The worst thing we can do here is blindly impose one project’s requirements to another’s opinions.

That being said, you should definitely add 100% code coverage, enable an extremely strict lint configuration, and convert to TypeScript because I know what’s best for you and your code. 🙃

Protips

“Encouraging Collaboration” argues that developers less familiar in an area often face knowledge-based limitations that could be resolved by more experienced developers. It would be irresponsible of me to make that claim without providing a few examples of common confusing cases.

(Code snippets are in TypeScript, but you can get the drift even if you’ve never written JavaScript or TypeScript before.)

“Impossible” cases

TL;DR: obey the single responsibility principle.

Pretending a private getHandler is supposed to only be called with "apple" or "banana":

private getHandler(fruit: string): IHandler {
    switch (fruit) {
        case "apple":
            return new AppleHandler();
        case "banana":
            return new BananaHandler();
        // (imagine 24 other cases here)
        default:
            throw new Error("This should never be hit!");
    }
}

Some would say that we “need” that error code to enforce that the fruit parameter is correct. Since it’s bad practice to test internals but generally impossible to hit this situation, you could conclude it’s impossible to fully unit test this method…

But adding never-to-be-run code is not the best way of handling this kind of validation!

Let’s take a step back and evaluate whether this complex private method (often a red flag) violates the Single Responsibility Principle to do something its class shouldn’t be taking care of. Is creating a handler for a fruit part of that component’s responsibilities? Likely not; that should be a separate class or function.

const handlers = new Map<string, () => IHandler>([
    ["apple", () => new AppleHandler()],
    ["banana", () => new BananaHandler()],
    // ...
]);

export const getHandler = (fruit: string) => {
    const creator = handlers.get(fruit);
    if (creator === undefined) {
        throw new Error(`Unknown fruit: '${fruit}'.`);
    }
    return creator();
};

Great! Now we can still throw an error for an unknown fruit in a testable way… and our code has been split up to better adhere to SOLID principles.

That’s why collaborative authoring is essential for well tested code. This method and many others are things you may gradually pick up as you see them or have them verbally ingrained into your brain via peer review.

Externalities

TL;DR: know when sections of code shouldn’t be completely covered.

Have you ever interacted with a library that requires you inherit one of their classes in order to use it?

/**
 * Extend me and pass instances to other parts of this library.
 */
declare abstract class AbstractHost {
    protected act(action: IAction);
}

Forcing you to extend the constructor is troublesome because it makes it impossible to test the sub-class without coincidentally calling the parent constructor. If the class does very little, such as set up dependency injections, that might be fine… but what if the parent class makes network calls, connects to databases, or performs other operations unfriendly to tests? Not good! Class extensions make it impossible to stub out the dependency on the library because you’re forced to include the library in your code. How would you stop that network call from happening when testing behavior intended to call act? You can’t! Oh no!

It’s a sad fact of life that the universe contains bad code (often written by good developers) you’ll need to interact with, and you might need to unit test your sub-classes of that bad code. There are still things you can do to make it better!

Sometimes the things you have deal with are so bad you can’t do the right thing. Maybe they dealt with external difficult-to-test systems. Maybe you wrote them years ago. Take the blow, minimize the damage, and be good when you can. In some scenarios it makes sense to mark parts of code as “untestable” only for the case of dealing with external dependencies. Larger projects I’ve been on tend to have up to three sections excluded from code coverage reports:

• Adapters: Interfaces on top of external or native code.
• Externals: Compiled external code not distributed via package management.
• Mains: Entry points for the application that call into native APIs.

Separate Test Responsibilities

TL;DR: unit tests should only test the component they’re meant to.

You have a component, Parent, whose sole responsibility is to contain multiple components of type Child. Child contains some basic text — maybe with a few edge cases.

You write rudimentary “this doesn’t crash” tests for both and some more advanced tests for Parent that exercise edge cases in Child. You’ve now achieved 100% unit test coverage for these new components. Hooray!…

Unfortunately, you’re no longer writing unit tests: these are more like integration tests. You’ve lost some of the benefits of focused unit tests:

• Child changes may fail Parent tests, which will be confusing to debug.
• Child logic tests go through logic in Parent, which adds extra work to writing test changes.
• Most importantly, if we change Child to be allowed as children of other component types (and delete Parent), we would then have to move the logic in Parent testing Child into Child’s unit tests. That’s extra work when we’re not changing Child at all!

It’s better to put the tests for each component in that component’s tests in the first place. It might be a little more work to write more test scaffolding in the beginning, but in the long term you’re buying more stability in your tests.

Takeaways

1. All code, given time, is bad code.
2. There is no silver bullet for good code.
3. If your code isn’t testable, it’s probably not great to begin with.
4. A rising tide lifts all boats.
5. Test all your code.
6. Exercise critical thinking.

Many thanks to the kind humans who helped proofread this article!

Chris Bevan, Derek Mehlhorn, Nayomi Mitchell, Ian Craig, Helen Anderson, Adam Reineke, Shawn Lee, Jesse Freitas: you’re all superb teammates. 🙌

Also my parents. Solid grammar checking there. 👌

Liked this post? Thanks! Let the world know: