maandag 9 juli 2012

1 + 1 = 3. The flaws of strict (j)unit testing.

Sometimes extreme is good, sometimes it isn't. I like eXtreme Programming a lot: applying best programming practices to the extreme to make exceptional software. However, I believe sometimes you need to be pragmatic. When you apply unit testing to the extreme, I think you are missing important feedback.

The concept that 1 + 1 = 3 is especially relevant when writing object-oriented code. Two or more classes working together often reinforce each other's behavior and the result of their cooperation is more than just the sum of their individual parts.

What I am trying to dissuade is blindly applying strict unit testing to the extreme: some people believe every object different from the object tested should be mocked or stubbed.
The basic idea is good: you don't want to test things twice. However, I think people who only write tests this way to test their entire systems are ignoring important issues with this way of testing. I am pleading for what I call Behavioral Unit Testing (BUT), meaning that the unit tests should test the actual result of the code and stop focusing on implementation, in the end expressing as much business value as possible.

Changing behavior of one class can have impact on another class using this class
The worst nightmare of each committed developer is breaking code without noticing it. Suppose you have a Person class with a getMainAddress method, returning null if no main address is set. Changing its behavior so it throws a NoAddressFoundException impacts other classes using this method. When those classes are tested using mocks, you won't see you broke something by changing the behavior of the Person class. However, if you use real Person instances in the other classes' tests and you wrote tests for persons without a main address, you will notice the changed behavior of the Person class has impact on the other classes as well and you can fix your tests accordingly.

Behavioral Unit Tests better express business value than mock tests
I assume everyone agrees that testing behavior is more desirable and valuable than testing your implementaton. Besides, when writing a BUT, you have far more chance you can talk about the expected behaviour of your test with business people, reinforcing the ubiquitous language throughout your team. Another advantage of testing behavior is that you think in two ways about the problem you're solving: what should be the solution (this is your test) and how to get to the solution (the implementation). That way you reduce the chance of false positives: working tests while the implementation is wrong.

Mock tests provide less feedback
When writing mock tests to the extreme, you are having the risk of expecting an algorithm to work without really testing it thoroughly. Never thought your algorithm would work and it turned out otherwise? Then come over to our team and stop writing tests, it's such a waste of time ;-)

Strict mock tests are not very refactor-proof
This one is obvious to me: If you have a mock test and you refactor your domain code, chances are you need to change your mock test too (unless you just renamed some methods). When you use real domain objects, you don't need to change any test (assuming you only refactored without changing behavior). Since I see tests as a safety net, I feel a lot safer having behavior tests then I would when I'd only have mock tests. A special case of this: when mock tests fail, developers tend to fix it by just adding or modifying a verify (I've added/changed it in the code, so now I'll add/change it in my test). Consecutively, people will think less about why tests are failing, they just fix tests by doing in the test what they already did in the code.

Cucumber and Fitnesse
When I am telling people I prefer behavior tests over mock tests, people always start worrying about the immense overhead which would be introduced by fancy tools such as Cucumber and Fitnesse. Actually, I'm not at all talking about these tools. They surely add extra value by offering visual tooling for business people, but that's not my plan. I am just telling you can write the same test in junit as you would in Cucumber or fitnesse with the same business value, without the fancy gui, but also without all of the overhead.

There will always be people saying testing things twice is bad because of the DRY principle. I can tell those people consciously, I'd rather have 10 tests failing for the same reason than zero failing tests while I broke my implementation.

Other people are telling me they use mocks to hide complexity occuring behind the scenes of a cooperating object. I think this is a valid remark. In that case, make sure you have a solid idea about the api of your cooperating object and moreover be aware when this api is changing and adapt your tests accordingly. With api I mean not only the method signature, but also exception handling, how it deals with unexpected parameter values, ...

There are also some other cases where I think the use of mocks in unit tests are justified:
  • checking delegation
  • checking corner cases that are really hard to simulate (for example system errors, transaction timeouts, ...)
  • stubbing repositories, factories, very complex calculations, ...
  • checking notifications sent to loosely coupled components
  • interactions with other systems

Always watch your concrete situation, cause there's no silver bullet. However, when you have a choice, try to reduce the number of mocks since you don't want to (re)write your tests with each refactoring. Furthermore, with mocks, there's always a chance of introducing non-obvious blind spots in your unit tests.

1 opmerking:

  1. Great Post Kristof!
    I could not agree more. Mocking Domain Objects is Evil. Period.
    Often the reason people mock Domain Objects is because they are difficult to build and have nastly dependencies and use all kinds of things to get them properly initialized. So yes we mock them. That is not fixing the root cause, but soothing the consequence.