Given a choice of adding unit tests or integration tests to an existing system, which is better to start with and why?

I'm currently consulting开发者_如何学运维 on an existing system, and I suspect the right next step is to add unit tests because of the types of exceptions that are occurring (null pointers, null lists, invalid return data). However, an employee who has a "personal investment" in the application insists on integration tests, even though the problems being reported are not related to specific use cases failing. In this case is it better to start with unit or integration tests?

Typically, it is very difficult to retrofit an untested codebase to have unit tests. There will be a high degree of coupling and getting unit tests to run will be a bigger time sink than the returns you'll get. I recommend the following:

• Get at least one copy of Working Effectively With Legacy Code by Michael Feathers and go through it together with people on the team. It deals with this exact issue.
• Enforce a rigorous unit testing (preferably TDD) policy on all new code that gets written. This will ensure new code doesn't become legacy code and getting new code to be tested will drive refactoring of the old code for testability.
• If you have the time (which you probably won't), write a few key focused integration tests over critical paths of your system. This is a good sanity check that the refactoring you're doing in step #2 isn't breaking core functionality.

Integration tests have an important role to play, but central to the testing of your code is unit-tests.

In the beginning, you will probably be forced to do integration tests only. The reason is that your code base is very heavily coupled (just a wild guess since there are no unit tests). Tight coupling means that you cannot create an instance of an object for test without creating a lot of related objects first. This makes any tests integration-tests per definition. It is crucial that you write these integration tests, as should be used as base lines for your bug-finding/refactoring efforts.

1. Write tests that document the bug.

2. Fix the bug so all created unit-tests are green.

3. It is time to be a good boyscout (leave the campsite/code in better order that it was when you entered) : Write tests that documents the functionality of the class that contained the bug.

4. As a part of your boyscout efforts, you start to decouple the class from others. Dependency Injection is THE tool here. Think that no other classes should be constructed inside other classses -- they should be injected as interfaces instead.

5. Finally, when you have decoupled the class, you can decouple the tests as well. Now, when you are injecting interfacing instead of creating concrete instances inside the tested class, you can make stubs/mocks instead. Suddenly your tests have become unit-tests!!

6. You can create integration tests as well, where you inject concrete classes instead of stubs and mocks. Just remember to keep them far away from the unit-tests; preferably in another assembly. Unit-tests should be able to run all the time, and run very fast don't let them be slowed down by slow integration tests.

The answer to the question depends on the context in which it is being asked. If you are looking to bring an existing codebase, and you are considering rewriting or replacing large portions of the code then it will be more valuable to design a comprehensive set of integration tests around the components you wish to rewrite or replace. On the other hand, if you are taking responsibility for an existing system that needs to be support and maintained, you might want to first start with unit tests to make sure that your more focused changes do not introduce errors.

I'll put it another way. If someone sends you an old car, take a look at it. If you are going to replace all of the components right away, then don't bother testing the minute performance characteristics of the fuel injector. If, on the other hand, you are going to be maintaining the car, as is, go ahead and write targeted unit tests around the components you are going to be fixing.

General rule, code without unit tests are brittle, systems without integrations are brittle. If you are going to be focused on low-level code changes, write Unit Tests first. If you are going to be focused on system-level changes, write integration tests.

And, also, make sure to ignore everything you read on sites like this. No one here knows the specifics of your project.

Choosing between integration tests and unit tests is highly subjective. It depends on various metrics of the codebase, most notably cohesion and coupling of the classes.

The generic advice that I would provide is that if classes that are loosely coupled, then test setup is going to consume lesser time, and hence, it would be much easier to start writing unit tests (especially against the more critical classes in the codebase).

On the other hand, in the event of high coupling, you might be better off writing integration tests against the more critical code paths, starting especially with a class that is loosely coupled (and resident much higher up in the execution stack). At the same time, attempts must be made to refactor the classes involved to reduce coupling (while using the integration tests as a safety net).