This is a topic I'm not capable to cover succinctly yet. I'll try my best at the moment and revisit later.
- Confidence in implementation
- Freedom of change
- Act as documentation
- Reduce costly debugging
All practices come with applicable contexts. I've seen some well put articles around the necessities of unit testing, whatever the definition of unit is. One main context we are under is that it's going to be a moderately big real project, which means at least months of effort from a few Scrum teams of developers.
On the contrary to common image, tests actually take more skills and effort to write. If a test is not well written, it will soon lose maintenance thus become useless. A bad written code at least has functionality.
Different test has different cost in writing and running. So they are pictured as a pyramid, with expensive one at top, relatively cheap one at bottom.
We are living in a real world, everything comes with a cost. Test is no less the same. If the probability of a test goes wrong is low, and other tests have covered the same test object, there would be no reason to keep it. Tests also require maintenance in the whole life cycle of the belonging product. Anyone should not be afraid to change tests to reflect the change of business.
What's worthwhile to note is that it is reasonable to have a test for integration. Its objective is to test the links between all units, while the caveat is it's not necessary to cover multiple use cases which should be the responsibility of individual units. It seems unnerving for new TDD adopters, a second thought of this approach will reveal that you've lost not much in confidence.
However, I've seen an exception in recent project. The workflow covers quite a few units, and it has various branches which expects different behaviours from units. Then it makes sense to have a feature test for each branch, in case the target is lost during the development.
When we look at the pyramid and the real cases together, confusion arises. Where is the boundary for each type of test? Shall we include that component in this test of not?
- The more object a test is covering, the less correctness it should be aiming for. For example, the higher level test should cover less scenario.
- Be pragmatic, add tests if only extra confidence is necessary.
- Avoid overlaps.
- Cost versus benefits.
- Don't be afraid to introduce manual tests.
- Actively clean up tests
- Change design if testability demands
- Solitary and sociable: two styles of unit testing
- Including UI interaction, data source, all collaborating services, possibly third party services.
- Should be minimal, just cover the most valuable happy path.
- Costly to maintain.
In my working environment we use integration test for those tests involving infrastructure which is out of our control, like database, http requests, third party services. The test will call a real instance, albeit in a controlled environment like docker or wiremock, to make sure the interaction between units are as expected. This is distinguished from unit tests by not using mocked collaborator, and from e2e tests by mocking the connected instance.
Within the boundary of service, while with data source which is part of the infrastructure. It may only verify the correctness of data fetched or persisted, leave a full chain of business logic intact. It requires schema populating, data populating and data cleaning up. The schema populating happens only once for a whole run of test. The latter two are required for each test cases.
Sometimes this refers to the specific technics to test contracts between services by Pact.
It requires a continuous integration pipeline to ensure the value of this kind of test.
The definition of unit is always not consistent between teams. Martin Fowler's take on this topic is worth a reading so attached as reference.
In our job most of the time we adopt solitary unit tests. In some cases we are open to test two classes together if their interaction is the whole point of one of the unit. Like builder class, simple factory method, testing them alone would miss the context of their usages, and their contract might change but not detectable on the other side if tested solitary.
It's not suggested starting a feature with a blackbox acceptance test or e2e test because it won't provide necessary feedback for next steps in development. Instead start with a xUnit feature test, covering one business scenario, and start adding tests and implementation at unit level.
The seam between connecting parts in most cases are tested by feature tests. If some collaboration only happens for edge cases which is not covered by feature test, the choice to add one or not is up to the trade-off between cost and benefits. There is no golden rule and sometimes the lesson has to be learnt with found bugs in later stages.
Other hidden infrastructure as API gateway, load balancer, networks, CDN etc. can only be covered by E2E test, even costly. Question is whether to test features in independent module/service or the interaction among them.