AuthenticationCoding standardsConfiguration for City of Helsinki Django projectsDocumentation practicesLicensingOnboarding checklist for development projectsPractices for Accepting Third Party ContributionsPractices for Version ControlProject management practices when using ScrumTechnology choicesTesting requirementsTests that should be writtenTests that would be niceHow tests should be writtenHow to write easily testable codeWeb accessibility
... or, what kind of regression tests should City of Helsinki services have?
well-tested program - adapting to every change - no surprises here! -- Panu Kalliokoski, programming haikus
- integration tests that cover the happy paths of each feature and user
- this means a program that does what the user would do (against a real service) and checks that what happens is what should happen
- integration or unit tests that cover forbidden behavior
- a program that does what a hacker would do and checks that it fails
- unit tests for every method/function whose logic is "complicated"
- "complicated" means there are at least three execution paths in the method/function
- integration or unit tests that provide full code coverage at least for
all code that implements the running service
- this is meant to exclude build-time helper scripts and similar stuff. But consider testing them, too!
- at least rudimentary performance tests: how long does it take to load
each view 1000 times?
- these produce an important time series where you can find when you made a performance blunder
- randomised fuzz testing to see that the service doesn't break (or reports errors correctly)
- state invariants if your data has interdependencies
- these can be made on database level
- Prefer integration (and e2e) tests over unit tests.
- Use unit tests for stuff that is hard to understand even on a local (unit) level.
- Integration tests are good for covering a lot of stuff quickly. Unit tests are good for finding exactly where a problem is.
- Don't use mocks unless it simplifies tests considerably. That is,
don't redo software components in a simplified way.
- instead, use the real components for tests, too.
- if the real component is really hard to use, then a mock (or fake version) might be feasible. Consider fixing the real component, though.
- Don't overtest or undertest. (This is hard.)
- Undertesting means that broken things go unnoticed. Write enough tests to cover all execution paths and test the results are right from the unit/service user's point of view.
- Overtesting means that change is hard because most of the time goes into fixing tests. In the test, don't check for details that are irrelevant and might change.
- Use fixtures (such as prepopulated data in the database) if your tests
- However, don't make your tests depend very much on the detail of testing data. Testing data will inevitably have to be updated someday.
- Build stateless services if you can. Put state into specialised
stateful services (buckets, databases).
- This way, the state can always be initialised (by fixtures) so that tests work predictably.
- Use stateless handlers (pure functions/components) wherever you can.
- Model your program as functions that process input data into output data. Hoist all I/O and other interaction onto as high level in the program as possible.
- In UI programs, separate state handling from the rest of the program.
- You can test your reducers (state update functions) separately if they are "complicated".