Measurement of code coverage is one of those things that always catches my attention. On the one hand, I often find that organizations don’t necessarily know how much code they are covering during testing — which is really surprising! At the other end of the coverage spectrum, there are organizations for whom the number is so important that the quality and efficacy of the tests has become mostly irrelevant. They mindlessly chase the 100% dragon and believe that if they have that number the software is good, or even the best that it can be. This is at least as dangerous as not knowing what you’ve tested, in fact perhaps more so since it can give you a false sense of security.
Security. Code coverage can be a good and interesting number to assess your software quality, but it’s important to remember that it is a means, rather than an end. We don’t want coverage for coverage’s sake, we want coverage because it’s supposed to indicate that we’ve done a good job testing the software. If the tests themselves aren’t meaningful, then having more of them certainly doesn’t indicate better software. The important goal is to make sure every piece of code is tested, not just executed. Failing enough time and money to fully test everything, at least make sure that everything important is tested.
What this means is that while low coverage means we’re probably not testing enough, high coverage by itself doesn’t necessarily correlate to high quality — the picture is more complicated than that.
Obviously, having a happy medium where you have “enough” coverage to be comfortable about releasing the software with a good, stable, maintainable test suite that has “just enough tests” would be perfect. But still, these coverage traps are common.
The first trap is the “we don’t know our coverage” trap. This seems unreasonable to me — coverage tools are cheap and plentiful. A friend of mine suggests that organizations know their coverage number isn’t good, so developers and testers are loath to expose the poor coverage to management. I would hope this isn’t the usual case.
One real issue that teams encounter when trying to measure coverage is that the system is too complicated. When you build an application out of pieces on top of pieces on top of pieces, just knowing where to put the coverage counters can be a daunting task. I would suggest that if it’s actually difficult to measure the coverage in your application, you should think twice about the architecture.
A second way to fall into this trap happens with organizations that may have a lot of testing, but no real coverage number because they haven’t found a proper way to aggregate the numbers from different test runs. If you’re doing manual testing, functional testing, unit testing, and end-to-end testing, you can’t simply add the numbers up. Even if they are each achieving 25% coverage it is unlikely that it’s 100% when combined. In fact, it’s more likely to be closer to the 25% than to the 100% when you look into it.
The next trap is the “coverage is everything” perspective. Once teams are able to measure coverage, it’s not uncommon for managers to say “let’s increase that number.” Eventually the number itself becomes more important than the testing. Perhaps the best analogy is one I heard from Parasoft’s founder, Adam Kolawa:
“It’s like asking a pianist to cover 100% of the piano keys rather than hit just the keys that make sense in the context of a given piece of music. When he plays the piece, he gets whatever amount of key coverage makes sense.”
Therein lies the problem — mindless coverage is the same as mindless music. The coverage needs to reflect real, meaningful use of the code, otherwise it’s just noise.
And speaking of noise… the cost of coverage goes up as coverage increases. Remember that you not only need to create tests, but you have to maintain them going forward. If you’re not planning on re-using and maintaining a test, you should probably not waste the time creating it in the first place.
As the test suite gets larger, the amount of noise increases in unexpected ways. Twice as many tests may mean two or even three times as much noise. The meaningless tests end up creating more noise than good tests because they have no real context, but have to be dealt with each time the tests are executed. Talk about technical debt! Useless tests are a real danger.
Now, in certain industries, safetycritical industries for example, the 100% coverage metric is a requirement. But even in that case, it’s all too easy to treat any execution of a line of code as a meaningful test, which is simply not true. I have two basic questions I ask to determine if a test is a good test:
- What does it mean when the test fails? 2
- What does it mean when the test passes?
Ideally, when a test fails, we know something about what went wrong, and if the test is really good, it will point us in the right direction to fix it. All too often when a test fails, no one knows why, no one can reproduce it, and the test is ignored.
Conversely, when a test passes we should be able to know what was tested —– it should mean that a particular feature or piece of functionality is working properly. If you can’t answer one of those questions, you probably have a problem with your test.
If you can’t answer either of them, the test is probably more trouble than it’s worth.
The way out of this trap is first to understand that the coverage percentage itself isn’t the goal. The real goal is to create useful meaningful tests. This of course takes time.
In simple code, writing unit tests is simple, but in complex, real-world applications, it means writing stubs and mocks and using frameworks. This can take quite a bit of time and if you’re not doing it all the time, it’s easy to forget the nuances of the APIs involved. Even if you are serious about testing, the time it takes to create a really good test can be more than you expect.
Hopefully you’ve learned that coverage is important, and improving coverage is a worthy goal. But keep in mind that simply chasing the percentage isn’t nearly as valuable as writing stable, maintainable, meaningful tests.