11 September 2012

Applying the test automation pyramid to the real world

You probaply heard of the test automation pyramid, which is a great idea to make automated testing work in an agile development environment. Martin Fowler explains the pyramid pretty well on his blog.

The pyramid has three layers UI, Service and Unit tests. Theses layers map well to standard web applications. But in more complex systems this simple model does not fit exactly at first view. But that is not true. It fits perfectly, if you understand it right.

What is wrong with the real world?

The test automation pyramid is created with web applications in mind. Think of a simple monolithic LAMP or Rails application. All the code is on the same server. In this case the idea of the pyramid fits perfectly, since we have an UI and a monolithic application.

In the real world, the situation differs a lot. Many applications do not have an UI at all. They might be so called backend services, which provide some logic, that works somewhere in the backround. The UI part of the software system is is probably written in a language like Ruby, while the backend services - maybe written in Java, Scala or Clojure are accessed through HTTP. But how do you apply the pyramid?

What is wrong with the pyramid?

The problem comes from the miss leading names of the layers in the pyramid. E. g. the top layer is called UI tests and our backend application has no UI. Sure there is hopefully an UI application, that access our backend service and has UI tests, but is that enough?

The solution is a simple renaming of the layers. Let us call the top layer function tests instead of UI tests. This name also applies well to non GUI applications. The tests in this layer should be very similar to UI tests.

Testing an application works

The functional test layer on top should prove the application works. This means, you have at least to start your application and access it through some mechanism similar to your production environment. In the best case your application will be deployed to a real server with a real database. Everything like authentication or access over the network should happen in a function test the same way it does in a production environment.

While the infrastructural requirements for theses tests are high and the tests itself might run very slow, there should be only few test cases in this category. You should not test failures here or some edge cases. Simply make sure your application is deployed, can access the database and answers as expected to some requests.

This kind of definition works for backend applications as well as for UI applications. UI tests are just a special form of functional tests. If you have both an UI application with UI tests and a backend service, your UI tests should run end to end, which also includes your backend services. In addition your backend service can also have its own functional tests, doing some basic HTTP requests. So you end up with functional tests on different layers of your system.

An important point is to know, that developers might not run theses test very frequently on their own local machines. But this is ok as long as some continous integration server will run them regularly.

Testing your application behaves correctly

When functional tests do not test the correct behaviour of your application, you need another layer to this. In the test pyramid the service test layer does this job.

Different to functional tests developers should run service tests very frequently. Thus the tests must execute fast enough to do this. Starting a Java EE application server might be much to slow for this. So service tests should skip this. Instead they should be written similar to unit tests, using a tool like Junit and instantiating objects from your code.

But what should be tested exactly? A good service test runs a typical scenario, that might appear similar in production. So take a typical conversation with your application and write it down in your test case. You should make sure, you focus on positive tests and skip error handling and edge cases at this level.

Service tests should make sure, the parts of your code work together as expected and behave like they should. This includes integration with third party libraries and even persistence. But always keep it simple. Use a simple database like SQLite or H2 instead of a full blown database server. Use simple HTTP mocks with static JSON files instead of real services in the backend. Keep your infrastructure as lean as possible to make running the tests easy.

Testing your code

A unit test should be executable as fast as possible. If some framework needs some seconds to initialize, skip it in the unit tests. Make your tests as fast as possible. There is a simple reason for this performance requirement. Developers should not run unit tests just frequently, they should run them as often as possible. You changed a line of code? Run the tests! This is how it should work and thus performance matters a lot.

Since unit tests should be very fast, it is cheap to run them, but it is also cheap to test edge cases and failure handling in unit tests. If your application should be able to deal with null values, unit tests are the place to prove it. If your application should validate some data, unit tests are the place to prove it. If your application should apply an algorithm, unit tests are the place to prove it.

Unit tests can do much more. They can help other developers to understand your code and its behaviour better. They can help you to change your code, without fearing to break it. And they can help you to find errors in the code, when they fail.

When should your tests run?

Unit tests can be found near the production code. In multi module projects, each module will have its unit tests. Most tools like Maven or Gradle in the Java world or Rake in the Ruby world will run the unit test by default, so there should be no probem for your continous integration server. Java IDEs have also integration for common test frameworks.

Little harder is to run service tests. They might be much slower than unit tests, so there should not be so much test cases to run. Important is, that a standard build like gradle build or mvn clean package will run theses tests. After all modules have been compiled, all unit tests should be executed. Now it is time to run the integration tests. They prove, the standard interaction with your software works well. Integration tests can be placed in a separate module to make this work best. But at least keep them separated from your unit tests to reduce problems.

The hardest part are the functional tests. Since their primary goal is to prove, the software runs and works as expected, there is no reason to execute them all the time as part of the standard build. Instead you should make theses tests easily executable without building the whole software. It is great, when a continous integration server deploys the software on a test machine with a real database, real HTTP and real authentication and than runs the functional tests against this system.

Other ways

Sure this is one solution to the problem and you may have found another way to work with different test levels. What do you think? How did you solve this? What worked great? What was bad? Leave me a comment.