Writing unit tests for code is surprisingly a hot button topic within software development. Many developers have very strong opinions in how tests should be approached, written, and implemented.
Over the past few months, several of our engineering departments have been pushing for more adoption of frontend tests for their code. Thanks to this push, our Frontend Platforms team has found that we need to put a lot more effort into the documentation around writing tests. Several engineers, from both our team and other feature teams, have been writing docs, pairing with others to write tests, and writing guides and presentations around unit testing.
Through all this work, we have slowly shifted our mindset from approaching testing as something that is a consistent, unit-based, solid foundation for an application to something that is a bit more fluid, a bit farther away from testing each unit individually, and something that changes as the application changes.
The key takeaway I have learned from a coworker is that the type and value of the tests you write depends on if you are writing the code for the application at the same time as the test or if you write the test after the feature is code-complete.
Writing tests before a code-complete feature is completely different from writing them when the feature is done
We have found that a lot of the pain points around writing unit tests are a result of attempting to write unit tests after the feature is code complete and deployed to production.
A lot of our other learnings have been so simply summarized by this tweet by Guillermo Rauch:
Write tests. Not too many. Mostly integration.— Guillermo Rauch (@rauchg) December 10, 2016
This short tweet is packed with so much wisdom, the core takeaways we have had as a team have been:
Code coverage is one of those feel-good stats that we seek as developers, giving us a rush when we see the coverage report come back with a higher percentage than we had previously. Unfortunately it has absolutely no value to the end user of your application.
Users do not care what your code coverage percentage is
You never see consumer products showcase these stats to their users in their footers, nor have I ever seen another developer choose an npm dependency on their code coverage percentage.
Your code will change, and your application will change as well, you shouldn't worry about how well covered the code is, and instead focus on how well the user flow is covered. Do you have a test for that checkout flow on site? What about for the user login flow? These are the things worth capturing coverage for, real parts of the application and not the number of lines of code.
In my experience, preferring to write integration tests rather than unit tests is even more contentious than the point above about code coverage, many developers seem to extrapolate from TDD that tests must be focused to units of code rather than writen for the larger picture of the feature or application as a whole.
The tests that I have found that are worth keeping around over time are these integration tests, ones that aren't at all worried about the implementation of the feature but rather the user flow through the feature.
The most important goal in software development (keep in mind writing tests is part of this work) is to deliver a working, enjoyable experience for the customer. Your customer won't care about the implementation of the checkout button, or if you are using some middleware for authentication. The key, when writing code, is to think about the user. Sometimes your user may be another developer using your service to implement another feature, or it may be a customer looking to buy their favorite bed frame.
I think the key for getting value out of software testing is not that you must write integration tests through all stages of development, but rather that over time the tests that should remain in the codebase should be scoped as integration tests.
Many of the developers I have talked with consider tests as these solid, never changing pillars of a codebase; however in my experience the type of tests, and the value of these tests, changes dramatically depending on what I am doing. If I am working on a refactor of a component for example, I may want to setup some visual regression tests, then I can refactor with confidence, and when the refactor work is done I can tear down the tests.
Unfortunately many of our modern testing tools don't really communicate this ephemerality of testing software, either through their implementation or through their documentation. This leads new developers to approach writing tests as something to do once, ship it and then forget it.
Often unit tests, those written to accomplish a code coverage goal or test a particular implementation detail, add friction to development that so many developers worry about when they try to get into testing their code. A good mental model for tests is that often pure unit tests should have a short half-life within the codebase, meaning they only exist for a short amount of time (frequently only during the very early stages of a new feature or application), and the half-life of integration tests is much longer and should remain in the codebase for the lifetime of the feature.
Unit tests should have a short half-life within your codebase, focus on code that has a long half-life instead