- Unit tests
- Integration tests
- White-box tests at the system level (formerly known as System / Feature tests)
- Black-box tests at the system level, aka end-to-end tests
- EE-specific tests
- How to test at the correct level?
This diagram demonstrates the relative priority of each test type we use.
e2e stands for end-to-end.
As of 2019-05-01, we have the following distribution of tests per level:
|Test level||Community Edition||Enterprise Edition||Community + Enterprise Edition|
|Black-box tests at the system level (aka end-to-end or QA tests)||68 (0.14%)||31 (0.2%)||99 (0.17%)|
|White-box tests at the system level (aka system or feature tests)||5,471 (11.9%)||969 (7.4%)||6440 (10.9%)|
|Integration tests||8,333 (18.2%)||2,244 (17.2%)||10,577 (17.9%)|
|Unit tests||32,031 (69.7%)||9,778 (75.1%)||41,809 (71%)|
Formal definition: https://en.wikipedia.org/wiki/Unit_testing
These kind of tests ensure that a single unit of code (a method) works as expected (given an input, it has a predictable output). These tests should be isolated as much as possible. For example, model methods that don’t do anything with the database shouldn’t need a DB record. Classes that don’t need database records should use stubs/doubles as much as possible.
|Code path||Tests path||Testing engine||Notes|
|Karma & Jest||More details in the Frontend Testing guide section.|
|RSpec||More details in the Testing Rails migrations guide.|
Unit tests are on the lowest abstraction level and typically test functionality that is not directly perceivable by a user.
- Exported functions and classes: Anything exported can be reused at various places in ways you have no control over. You should document the expected behavior of the public interface with tests.
- Vuex actions: Any Vuex action must work in a consistent way, independent of the component it is triggered from.
- Vuex mutations: For complex Vuex mutations, you should separate the tests from other parts of the Vuex store to simplify problem-solving.
- Non-exported functions or classes: Anything not exported from a module can be considered private or an implementation detail, and doesn’t need to be tested.
- Constants: Testing the value of a constant means copying it, resulting in extra effort without additional confidence that the value is correct.
- Vue components: Computed properties, methods, and lifecycle hooks can be considered an implementation detail of components, are implicitly covered by component tests, and don’t need to be tested. For more information, see the official Vue guidelines.
- State of the class under test: Modifying the state of the class under test directly rather than using methods of the class avoids side effects in test setup.
- Other exported classes: Every class must be tested in isolation to prevent test scenarios from growing exponentially.
- Single DOM elements if passed as parameters: For tests only operating on single DOM elements, rather than a whole page, creating these elements is cheaper than loading an entire HTML fixture.
- All server requests: When running frontend unit tests, the backend may not be reachable, so all outgoing requests need to be mocked.
- Asynchronous background operations: Background operations cannot be stopped or waited on, so they will continue running in the following tests and cause side effects.
- Non-exported functions or classes: Everything that is not exported can be considered private to the module, and will be implicitly tested through the exported classes and functions.
- Methods of the class under test: By mocking methods of the class under test, the mocks will be tested and not the real methods.
- Utility functions (pure functions, or those that only modify parameters): If a function has no side effects because it has no state, it is safe to not mock it in tests.
- Full HTML pages: Avoid loading the HTML of a full page in unit tests, as it slows down tests.
Component tests cover the state of a single component that is perceivable by a user depending on external signals such as user input, events fired from other components, or application state.
- Vue components
- Vue applications: Vue applications may contain many components. Testing them on a component level requires too much effort. Therefore they are tested on frontend integration level.
- HAML templates: HAML templates contain only Markup and no frontend-side logic. Therefore they are not complete components.
- DOM: Operating on the real DOM is significantly slower than on the virtual DOM.
- Properties and state of the component under test: Similar to testing classes, modifying the properties directly (rather than relying on methods of the component) avoids side effects.
- Vuex store: To avoid side effects and keep component tests simple, Vuex stores are replaced with mocks.
- All server requests: Similar to unit tests, when running component tests, the backend may not be reachable, so all outgoing requests need to be mocked.
- Asynchronous background operations: Similar to unit tests, background operations cannot be stopped or waited on. This means they will continue running in the following tests and cause side effects.
- Child components:
Every component is tested individually, so child components are mocked.
- Methods or computed properties of the component under test: By mocking part of the component under test, the mocks will be tested and not the real component.
Formal definition: https://en.wikipedia.org/wiki/Integration_testing
These kind of tests ensure that individual parts of the application work well together, without the overhead of the actual app environment (i.e. the browser). These tests should assert at the request/response level: status code, headers, body. They’re useful to test permissions, redirections, what view is rendered etc.
|Code path||Tests path||Testing engine||Notes|
|RSpec||For N+1 tests, use request specs|
|Karma & Jest||More details below|
Integration tests cover the interaction between all components on a single page. Their abstraction level is comparable to how a user would interact with the UI.
- Page bundles (
- Vue applications outside of page bundles: Testing Vue applications as a whole ensures the corresponding frontend components integrate well.
- HAML views (use fixtures instead): Rendering HAML views requires a Rails environment including a running database, which you cannot rely on in frontend tests.
- All server requests: Similar to unit and component tests, when running component tests, the backend may not be reachable, so all outgoing requests must be mocked.
- Asynchronous background operations that are not perceivable on the page: Background operations that affect the page must be tested on this level. All other background operations cannot be stopped or waited on, so they will continue running in the following tests and cause side effects.
- DOM: Testing on the real DOM ensures your components work in the intended environment. Part of DOM testing is delegated to cross-browser testing.
- Properties or state of components: On this level, all tests can only perform actions a user would do. For example: to change the state of a component, a click event would be fired.
- Vuex stores: When testing the frontend code of a page as a whole, the interaction between Vue components and Vuex stores is covered as well.
and use methods which are deprecated in Rails 5 (#23768).
These kind of tests ensure the GitLab Rails application (for example,
gitlab) works as expected from a browser point of view.
- knowledge of the internals of the application are still required
- data needed for the tests are usually created directly using RSpec factories
- expectations are often set on the database or objects state
These tests should only be used when:
- the functionality/component being tested is small
- the internal state of the objects/database needs to be tested
- it cannot be tested at a lower level
For instance, to test the breadcrumbs on a given page, writing a system test makes sense since it’s a small component, which cannot be tested at the unit or controller level.
Only test the happy path, but make sure to add a test case for any regression that couldn’t have been caught at lower levels with better tests (for example, if a regression is found, regression tests should be added at the lowest level possible).
|Tests path||Testing engine||Notes|
|Capybara + RSpec||If your test has the |
In contrast to frontend integration tests, feature tests make requests against the real backend instead of using fixtures. This also implies that database queries are executed which makes this category significantly slower.
- The RSpec testing guidelines.
- System / Feature tests in the Testing Best Practices.
- Issue #26159 which aims at combining those guidelines with this page.
- Use cases that require a backend, and cannot be tested using fixtures.
- Behavior that is not part of a page bundle, but defined globally.
:js flag is added to the test to make sure the full environment is loaded:
scenario 'successfully', :js do sign_in(create(:admin)) end
The steps of each test are written using (capybara methods).
XHR (XMLHttpRequest) calls might require you to use
wait_for_requests in between steps, such as:
find('.form-control').native.send_keys(:enter) wait_for_requests expect(page).not_to have_selector('.card')
If we’re confident that the low-level components work well (and we should be if we have enough Unit & Integration tests), we shouldn’t need to duplicate their thorough testing at the System test level.
It’s very easy to add tests, but a lot harder to remove or improve tests, so one should take care of not introducing too many (slow and duplicated) tests.
The reasons why we should follow these best practices are as follows:
- System tests are slow to run since they spin up the entire application stack in a headless browser, and even slower when they integrate a JS driver
The QA framework and instance-level scenarios are part of GitLab Rails so that they’re always in-sync with the codebase (especially the views).
- knowledge of the internals of the application are not required
- data needed for the tests can only be created using the GUI or the API
- expectations can only be made against the browser page and API responses
Every new feature should come with a test plan.
|Tests path||Testing engine||Notes|
|Capybara + RSpec + Custom QA framework||Tests should be placed under their corresponding Product category|
See end-to-end tests for more information.
Smoke tests are quick tests that may be run at any time (especially after the pre-deployment migrations).
These tests run against the UI and ensure that basic functionality is working.
See Smoke Tests for more information.
GitLab QA orchestrator is a tool that allows to test that all these pieces integrate well together by building a Docker image for a given version of GitLab Rails and running end-to-end tests (i.e. using Capybara) against it.
Learn more in the GitLab QA orchestrator README.
EE-specific tests follows the same organization, but under the
As many things in life, deciding what to test at each level of testing is a trade-off:
- Unit tests are usually cheap, and you should consider them like the basement of your house: you need them to be confident that your code is behaving correctly. However if you run only unit tests without integration / system tests, you might miss the big / picture !
- Integration tests are a bit more expensive, but don’t abuse them. A system test is often better than an integration test that is stubbing a lot of internals.
Another way to see it is to think about the “cost of tests”, this is well explained in this article and the basic idea is that the cost of a test includes:
- The time it takes to write the test
- The time it takes to run the test every time the suite runs
- The time it takes to understand the test
- The time it takes to fix the test if it breaks and the underlying code is OK
- Maybe, the time it takes to change the code to make the code testable.
There are cases where the behaviour you are testing is not worth the time spent running the full application, for example, if you are testing styling, animation, edge cases or small actions that don’t involve the backend, you should write an integration test using Jasmine.