PHPUnit best practices to organize tests

Question

I am currently going to start from scratch with the phpunit tests for a project. So I was looking into some projects (like Zend) to see how they are doing things and how they organizing their tests.

Most things are pretty clear, only thing I have some problems with is how to organize the test suites properly. Zend has an AllTests.php from which loads others test suites.
Tough looking at the class it is useing PHPUnit_Framework_TestSuite to create a suite object and then add the other suites to it, but if I look in the PHPUnit docs for organizing tests in PHPUnit versions after 3.4 there is only a description for XML or FileHierarchy. The one using classes to organize the tests was removed.
I haven't found anything that this method is deprecated and projects like Zend are still using it.

But if it is deprecated, how would I be able to organize tests in the same structure with the xml configuration? Executing all tests is no problem, but how would I organize the tests (in the xml) if I only wanted to execute a few tests. Maybe creating several xmls where I only specify a few tests/test suites to be run?

So if I would want to only test module1 and module2 of the application, would I have an extra xml for each and defining test suites only for those modules (classes used by the module) in it. And also one that defines a test suite for all tests?

Or would it be better to use the @group annotation on the specific tests to mark them to be for module1 or module2?

Thanks in advance for pointing me to some best practices.

Answer 1

Basic Directory Structure:

I have been experimenting with keeping the test code right next to the code being tested, literally in the same directory with a slightly different file name from the file with the code it is testing. So far I am liking this approach. The idea is you don't have to spend time and energy keeping the directory structure in sync between your code and your test code. So if you change the name of the directory the code is in, you don't then also need to go and find and change the directory name for the test code. This also causes you to spend less time looking for the test code that goes with some code as it is right there next to it. This even makes it less of a hassle to create the file with the test code to begin with because you don't have to first find the directory with the tests, possibly create a new directory to match the one you are creating tests for, and then create the test file. You just create the test file right there.

One huge advantage of this is it means the other employees (not you because you would never do this) will be less likely to avoid writing test code to begin with because it is just too much work. Even as they add methods to existing classes they will be less likely to not feel like adding tests to the existing test code, because of the low friction of finding the test code.

One disadvantage is this makes it harder to release your production code without the tests accompanying it. Although if you use strict naming conventions it still might be possible. For example, I have been using ClassName.php, ClassNameUnitTest.php, and ClassNameIntegrationTest.php. When I want to run all the unit tests, there is a suite that looks for files ending in UnitTest.php. The integration test suite works similarly. If I wanted to, I could use a similar technique to prevent the tests from getting released to production.

Another disadvantage of this approach is when you are just looking for actual code, not test code, it takes a little more effort to differentiate between the two. But I feel this is actually a good thing as it forces us to feel the pain of the reality that test code is code too, it adds its' own maintenance costs, and is just as vitally a part of the code as anything else, not just something off to the side somewhere.

One test class per class:

This is far from experimental for most programmers, but it is for me. I am experimenting with only having one test class per class being tested. In the past I had an entire directory for each class being tested and then I had several classes inside that directory. Each test class setup the class being tested in a certain way, and then had a bunch of methods each one with a different assertion made. But then I started noticing certain conditions I would get these objects into had stuff in common with other conditions it got into from other test classes. The duplication become too much to handle, so I started creating abstractions to remove it. The test code became very difficult to understand and maintain. I realized this, but I couldn't see an alternative that made sense to me. Just having one test class per class seemed like it would not be able to test nearly enough situations without becoming overwhelming to have all that test code inside one test class. Now I have a different perspective on it. Even if I was right, this is a huge dampener on other programmers, and myself, wanting to write and maintain the tests. Now I am experimenting with forcing myself to have one test class per class being tested. If I run into too many things to test in that one test class, I am experimenting with seeing this as an indication that the class being tested is doing too much, and should be broken up into multiple classes. For removing duplication I am trying to stick to simpler abstractions as much as possible that allows everything to exist in one readable test class.

UPDATE I am still using and liking this approach, but I have found a very good technique for reducing the amount of test code and the amount of duplication. It is important to write reusable assertion methods inside the test class itself that gets heavily used by the test methods in that class. It helps me to come up with the right types of assertion methods if I think of them as internal DSLs (something Uncle Bob promotes, well actually he promotes actually making internal DSLs). Sometimes you can take this DSL concept even further (actually make a DSL) by accepting a string parameter that has a simple value that refers to what kind of test you are trying to perform. For example, one time I made a reusable asssertion method that accepted a $left, $comparesAs, and a $right parameter. This made the tests very short and readable as the code read something like $this->assertCmp('a', '<', 'b').

Honestly, I can't emphasize that point enough, it is the entire foundation of making writing tests something that is sustainable (that you and the other programmers want to keep doing). It makes it possible for the value that tests add to outweigh what they take away. The point is not that you need to use that exact technique, the point is you need to use some kind of reusable abstractions that allow you to write short and readable tests. It might seem like I'm getting off topic from the question, but I'm really not. If you don't do this, you will eventually fall into the trap of needing to create multiple test classes per class being tested, and things really break down from there.

Taking this a step further:

You can take these ideas even further if you are willing to make your own testing framework. Instead of having the tests be in a separate file named after the class it is testing, place the tests directly in the code. The code that tests a class should be inside a class. This makes it even easier to find the test code that goes with the code because it is not in a different file. This means that if you rename the class, you don't need to then find and rename the test class as well. This makes you and others even less likely to not begin, maintain, and improve tests because of the friction of switching over to the test class. Each class will have more code now and at times it will be annoying to have the test code mixed in when you just want to see the code, but this pain is actually a good thing. Just as I explained above, this pain is forcing you to deal with reality. The test code is just as much a real and important part of your code as everything else is. If the combination of the code and the tests is getting too much, it means your class is trying to do too much and needs to be broken up, or you need some better abstractions or DSLs to keep it shorter and/or more readable.

This blurs the lines between tests and code, and this is a good thing. Some of assert methods that you initially create to be used by the tests will then prove to be useful by the regular code too and will already be there for it to call. This approach also makes it easier to see when it is better for some assertions to run all the time as part of the code instead of only with the tests. That can alleviate some of the burden of the test code, where certain conditions will always be tested instead having to test it over and over again. You can place keep some of these assertions behind a config flag so they only run when the tests are running. That can be helpful when certain assertions may not hold true in a production environment when many requests are hitting the system all the time and changing data, but are expected to be true in an automated testing environment. For example, if you just inserted a new row into a table, in a test environment you would expect the count to be one larger than before, but it may not be practical to guarantee this a production environment, especially if there is already a lot of legacy code or technical debt.

One way you can design such a testing framework is by by relying heavily on PHP attributes. A testing framework can look for certain attributes above a method that tells it what values to pass in for the parameters and what assertions to run. This works especially well for pure code where all you need to do is pass in parameters and test the return value, but you can definitely do other types of testing and assertions as well. In cases where this does not offer enough flexibility, it still does because of a trick. You can have a method that only has test code in it, and you can get this method to be run with the other tests by placing the appropriate attribute above it, perhaps even passing different values into it. The main disadvantage of this approach is it seems impossible for the tests you specify in the attributes to be very well type checked by something like Psalm or PhpStan. Perhaps you could get around this by generating separate, more traditional tests from these attributes, and then that code can be type checked, but I have not tried this yet.

Answer 2

I'll start of by linking to the manual and then going into what I've seen and heard in the field.

Organizing phpunit test suites

Module / Test folder organization in the file system

My recommended approach is combining the file system with an xml config.

tests/
 \ unit/
 | - module1
 | - module2
 - integration/
 - functional/

with a phpunit.xml with a simple:

<testsuites>
  <testsuite name="My whole project">
    <directory>tests</directory>
  </testsuite>
</testsuites>

you can split the testsuites if you want to but thats a project to project choice.

Running phpunit will then execute ALL tests and running phpunit tests/unit/module1 will run all tests of module1.

Organization of the "unit" folder

The most common approach here is to mirror your source/ directory structure in your tests/unit/ folder structure.

You have one TestClass per ProductionClass anyways so it's a good approach in my book.

In file organization

• One class per file.

It's not going to work anyways if you have more than one test class in one file so avoid that pitfall.

• Don't have a test namespace

It just makes writing the test more verbose as you need an additional use statement so I'd say the testClass should go in the same namespace as the production class but that is nothing PHPUnit forces you to do. I've just found it to be easier with no drawbacks.

Executing only a few tests

For example phpunit --filter Factory executes all FactoryTests while phpunit tests/unit/logger/ executes everything logging related.

You can use @group tags for something like issue numbers, stories or something but for "modules" I'd use the folder layout.

Multiple xml files

It can be useful to create multiple xml files if you want to have:

• one without code coverage
• one just for the unit tests (but not for the functional or integration or long running tests)
• other common "filter" cases
• PHPBB3 for example does that for their phpunit.xmls

Code coverage for your tests

As it is related to starting a new project with tests:

• My suggestion is to use @covers tags like described in my blog (Only for unit tests, always cover all non public functions, always use covers tags.
• Don't generate coverage for your integration tests. It gives you a false sense of security.
• Always use whitelisting to include all of your production code so the numbers don't lie to you!

Autoloading and bootstrapping your tests

You don't need any sort of auto loading for your tests. PHPUnit will take care of that.

Use the <phpunit bootstrap="file"> attribute to specify your test bootstrap. tests/bootstrap.php is a nice place to put it. There you can set up your applications autoloader and so on (or call your applications bootstrap for that matter).

Summary

• Use the xml configuration for pretty much everything
• Separate unit and integration tests
• Your unit test folders should mirror your applications folder structure
• To only execute specif tests use phpunit --filter or phpunit tests/unit/module1
• Use the strict mode from the get go and never turn it off.

Sample projects to look at

• Sebastian Bergmanns "Bank Account" example project
• phpBB3 Even so they have to fight some with their legacy ;)
• Symfony
• Doctrine2