According to the Oxford Dictionary, entropy can be defined as:
gradual decline into disorder
When I think about integration or end-to-end tests I've encountered in my career, entropy definitely comes to mind. I have written and encountered integration tests that were written when the project was under active development, but became more and more brittle and lost value as time went on. As a result, developers begin skipping them in CI builds and soon neglect them altogether. In this article, I want to talk about the two main reasons I believe this entropy occurs and what can be done to make sure integration tests thrive and provide value throughout the lifespan of a project.
Note: Integration tests could have many different meanings to different developers. For the purposes of this article I'm referring specifically to end to end tests that either test a UI (Selenium, Cypress, etc.) or test the data returned by a REST API.
The most common reason for brittle integration tests I've seen is a reliance on data. I do believe that end-to-end tests should manipulate data all the way to the data store in order to assure stakeholders that the application is working properly. However, these tests shouldn't assume that the data will always and forever be in the correct state for an accurate run of the test. Data change is inevitable and no one wants to waste hours trying to diagnose a failing test because of a change in data.
There are several ways to mitigate this risk. The easiest would be to write setup and tear down code for integration tests that ensure that data is in the right state before and after the run. Of course, if integration tests can be written in a way that they are not reliant on any specific data this is even better. A more complicated solution could be to utilize a Docker container that contains known data specifically for integration tests.
In my opinion, every integration test should be tied to a business requirement and this link should be immediately obvious to anyone who encounters the test. Unfortunately, this doesn't always happen. Tests that make an HTTP request to an API endpoint and assert against data or tests that look for a specific element to appear on a UI don't necessarily meet this need. When the test inevitably breaks for a developer unfamiliar with the original requirements, the test should make its value clear. Otherwise, it's all to easy to decide the test isn't valuable anymore and delete it.
My favorite method for linking tests to actual business requirements is to use Behavior Driven Development. You may be familiar with this concept as BDD, Cucumber, or Gherkin, but the idea is the same. If integration tests can be written in a behavior oriented fashion that business owners can easily understand and verify, the result is a huge win for the project.