Testing best practices

# Optional: install the `speedscope` package to easily upload the JSON report to https://www.speedscope.app
npm install -g speedscope
Generate the JSON report
bin/rspec-stackprof --speedscope=true <your_slow_spec>
# There will be the name of the report displayed when the script ends.
# Upload the JSON report to speedscope.app
speedscope tmp/<your-json-report>.json
How to interpret the flamegraph
Below are some useful tips to interpret and navigate the flamegraph:
There are several views available for the flamegraph. Left Heavy is particularly useful when there are a lot of function calls (for example, feature specs).
You can zoom in or out! See the navigation documentation
If you are working on a slow feature test, search for Capybara::DSL# in the search to see the capybara actions that are made, and how long they take!
See #414929 or #375004 for some analysis examples.
Optimize factory usage
A common cause of slow tests is excessive creation of objects, and thus
computation and DB time. Factories are essential to development, but they can
make inserting data into the DB so easy that we may be able to optimize.
The two basic techniques to bear in mind here are:
Reduce: avoid creating objects, and avoid persisting them.
Reuse: shared objects, especially nested ones we do not examine, can generally be shared.
To avoid creation, it is worth bearing in mind that:
instance_double and spy are faster than FactoryBot.build(...).
FactoryBot.build(...) and .build_stubbed are faster than .create.
Don’t create an object when you can use build, build_stubbed, attributes_for,
spy, or instance_double. Database persistence is slow!
Use Factory Doctor to find cases where database persistence is not needed in a given test.
Examples of factories optimization 1, 2.
# run test for path
FDOC=1 bin/rspec spec/[path]/[to]/[spec].rb
A common change is to use build or build_stubbed instead of create:
# Old
let(:project) { create(:project) }
# New
let(:project) { build(:project) }
Factory Profiler can help to identify repetitive database persistence via factories.
# run test for path
FPROF=1 bin/rspec spec/[path]/[to]/[spec].rb
# to visualize with a flamegraph
FPROF=flamegraph bin/rspec spec/[path]/[to]/[spec].rb
A common cause of a large number of created factories is factory cascades, which result when factories create and recreate associations.
They can be identified by a noticeable difference between total time and top-level time numbers:
   total   top-level     total time      time per call      top-level time               name
     208           0        9.5812s            0.0461s             0.0000s          namespace
     208          76       37.4214s            0.1799s            13.8749s            project
The table above shows us that we never create any namespace objects explicitly
(top-level == 0) - they are all created implicitly for us. But we still end up
with 208 of them (one for each project) and this takes 9.5 seconds.
In order to reuse a single object for all calls to a named factory in implicit parent associations,
FactoryDefault
can be used:
RSpec.describe API::Search, factory_default: :keep do
  let_it_be(:namespace) { create_default(:namespace) }
Then every project we create uses this namespace, without us having to pass
it as namespace: namespace. In order to make it work along with let_it_be, factory_default: :keep
must be explicitly specified. That keeps the default factory for every example in a suite instead of
recreating it for each example.
To prevent accidental reliance between test examples, objects created
with create_default are
frozen.
Maybe we don’t need to create 208 different projects - we
can create one and reuse it. In addition, we can see that only about 1/3 of the
projects we create are ones we ask for (76/208). There is benefit in setting
a default value for projects as well:
  let_it_be(:project) { create_default(:project) }
In this case, the total time and top-level time numbers match more closely:
   total   top-level     total time      time per call      top-level time               name
      31          30        4.6378s            0.1496s             4.5366s            project
       8           8        0.0477s            0.0477s             0.0477s          namespace
Let’s talk about let
There are various ways to create objects and store them in variables in your tests. They are, from least efficient to most efficient:
let! creates the object before each example runs. It also creates a new object for every example. You should only use this option if you need to create a clean object before each example without explicitly referring to it.
let lazily creates the object. It isn’t created until the object is called. let is generally inefficient as it creates a new object for every example. let is fine for simple values. However, more efficient variants of let are best when dealing with database models such as factories.
let_it_be_with_refind works similar to let_it_be_with_reload, but the former calls ActiveRecord::Base#find instead of ActiveRecord::Base#reload. reload is usually faster than refind.
let_it_be_with_reload creates an object one time for all examples in the same context, but after each example, the database changes are rolled back, and object.reload will be called to restore the object to its original state. This means you can make changes to the object before or during an example. However, there are cases where state leaks across other models can occur. In these cases, let may be an easier option, especially if only a few examples exist.
let_it_be creates an object one time for all of the examples in the same context. This is a great alternative to let and let! for objects that do not need to change from one example to another. Using let_it_be can dramatically speed up tests that create database models. See https://github.com/test-prof/test-prof/blob/master/docs/recipes/let_it_be.md#let-it-be for more details and examples.
Pro-tip: When writing tests, it is best to consider the objects inside a let_it_be as immutable, as there are some important caveats when modifying objects inside a let_it_be declaration (1, 2). To make your let_it_be objects immutable, consider using freeze: true:





    
# Before
let_it_be(:namespace) { create_default(:namespace) }
# After
let_it_be(:namespace, freeze: true) { create_default(:namespace) }
See https://github.com/test-prof/test-prof/blob/master/docs/recipes/let_it_be.md#state-leakage-detection for more information on let_it_be freezing.
let_it_be is the most optimized option since it instantiates an object once and shares its instance across examples. If you find yourself needing let instead of let_it_be, try let_it_be_with_reload.
# Old
let(:project) { create(:project) }
# New
let_it_be(:project) { create(:project) }
# If you need to expect changes to the object in the test
let_it_be_with_reload(:project) { create(:project) }
Here is an example of when let_it_be cannot be used, but let_it_be_with_reload allows for more efficiency than let:
let_it_be(:user) { create(:user) }
let_it_be_with_reload(:project) { create(:project) } # The test will fail if `let_it_be` is used
context 'with a developer' do
  before do
    project.add_developer(user)
  it 'project has an owner and a developer' do
    expect(project.members.map(&:access_level)).to match_array([Gitlab::Access::OWNER, Gitlab::Access::DEVELOPER])
context 'with a maintainer' do
  before do
    project.add_maintainer(user)
  it 'project has an owner and a maintainer' do
    expect(project.members.map(&:access_level)).to match_array([Gitlab::Access::OWNER, Gitlab::Access::MAINTAINER])
Stubbing methods within factories
You should avoid using allow(object).to receive(:method) in factories, as this makes the factory unable to be used with let_it_be, as described in common test setup.
Instead, you can use stub_method to stub the method:
  before(:create) do |user, evaluator|
    # Stub a method.
    stub_method(user, :some_method) { 'stubbed!' }
    # Or with arguments, including named ones
    stub_method(user, :some_method) { |var1| "Returning #{var1}!" }
    stub_method(user, :some_method) { |var1: 'default'| "Returning #{var1}!" }
  # Un-stub the method.
  # This may be useful where the stubbed object is created with `let_it_be`
  # and you want to reset the method between tests.
  after(:create) do  |user, evaluator|
    restore_original_method(user, :some_method)
    restore_original_methods(user)
stub_method does not work when used in conjunction with let_it_be_with_refind. This is because stub_method will stub a method on an instance and let_it_be_with_refind will create a new instance of the object for each run.

stub_method does not support method existence and method arity checks.

stub_method is supposed to be used in factories only. It’s strongly discouraged to be used elsewhere. Consider using RSpec mocks if available.
Stubbing member access level
To stub member access level for factory stubs like Project or Group use
stub_member_access_level:
let(:project) { build_stubbed(:project) }
let(:maintainer) { build_stubbed(:user) }
let(:policy) { ProjectPolicy.new(maintainer, project) }
it 'allows admin_project ability' do
  stub_member_access_level(project, maintainer: maintainer)
  expect(policy).to be_allowed(:admin_project)
Refrain from using this stub helper if the test code relies on persisting
project_authorizations or Member records. Use Project#add_member or Group#add_member instead.
Additional profiling metrics
We can use the rspec_profiling gem to diagnose, for instance, the number of SQL queries we’re making when running a test.
This could be caused by some application side SQL queries triggered by a test that could mock parts that are not under test (for example, !123810).
See the instructions in the performance docs.
Troubleshoot slow feature test
A slow feature test can generally be optimized the same way as any other test. However, there are some specific techniques that can make the troubleshooting session more fruitful.
See what the feature test is doing in the UI
# Before
bin/rspec ./spec/features/admin/admin_settings_spec.rb:992
# After
WEBDRIVER_HEADLESS=0 bin/rspec ./spec/features/admin/admin_settings_spec.rb:992
See Run :js spec in a visible browser for more info.
Search for Capybara::DSL# when using profiling
When using stackprof flamegraphs, search for Capybara::DSL# in the search to see the capybara actions that are made, and how long they take!
Identify slow tests
Running a spec with profiling is a good way to start optimizing a spec. This can
be done with:
bundle exec rspec --profile -- path/to/spec_file.rb
Which includes information like the following:
Top 10 slowest examples (10.69 seconds, 7.7% of total time):
  Issue behaves like an editable mentionable creates new cross-reference notes when the mentionable text is edited
    1.62 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:164
  Issue relative positioning behaves like a class that supports relative positioning .move_nulls_to_end manages to move nulls to the end, stacking if we cannot create enough space
    1.39 seconds ./spec/support/shared_examples/models/relative_positioning_shared_examples.rb:88
  Issue relative positioning behaves like a class that supports relative positioning .move_nulls_to_start manages to move nulls to the end, stacking if we cannot create enough space
    1.27 seconds ./spec/support/shared_examples/models/relative_positioning_shared_examples.rb:180
  Issue behaves like an editable mentionable behaves like a mentionable extracts references from its reference property
    0.99253 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:69
  Issue behaves like an editable mentionable behaves like a mentionable creates cross-reference notes
    0.94987 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:101
  Issue behaves like an editable mentionable behaves like a mentionable when there are cached markdown fields sends in cached markdown fields when appropriate
    0.94148 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:86
  Issue behaves like an editable mentionable when there are cached markdown fields when the markdown cache is stale persists the refreshed cache so that it does not have to be refreshed every time
    0.92833 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:153
  Issue behaves like an editable mentionable when there are cached markdown fields refreshes markdown cache if necessary
    0.88153 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:130
  Issue behaves like an editable mentionable behaves like a mentionable generates a descriptive back-reference
    0.86914 seconds ./spec/support/shared_examples/models/mentionable_shared_examples.rb:65
  Issue#related_issues returns only authorized related issues for given user
    0.84242 seconds ./spec/models/issue_spec.rb:335
Finished in 2 minutes 19 seconds (files took 1 minute 4.42 seconds to load)
277 examples, 0 failures, 1 pending
From this result, we can see the most expensive examples in our spec, giving us
a place to start. The most expensive examples here are in
shared examples; any reductions generally have a larger impact as
they are called in multiple places.





    
Top slow tests
We collect information about tests duration in rspec_profiling_stats project. The data is showed using GitLab Pages in this
With issue we defined thresholds for tests duration that can act a guide.
For tests that are not meeting the thresholds, we create issues automatically in order to improve them.
For tests that are slow for a legitimate reason and to skip issue creation, add allowed_to_be_slow: true.
Feature tests
Controllers and Requests tests
Other
Method
2023-02-15
67.42 seconds
44.66 seconds
76.86 seconds
Top slow test eliminating the maximum
2023-06-15
50.13 seconds
19.20 seconds
27.12
45.40 seconds
Avg for top 100 slow tests
Avoid repeating expensive actions
While isolated examples are very clear, and help serve the purpose of specs as
specification, the following example shows how we can combine expensive
actions:
subject { described_class.new(arg_0, arg_1) }
it 'creates an event' do
  expect { subject.execute }.to change(Event, :count).by(1)
it 'sets the frobulance' do
  expect { subject.execute }.to change { arg_0.reset.frobulance }.to('wibble')
it 'schedules a background job' do
  expect(BackgroundJob).to receive(:perform_async)
  subject.execute
If the call to subject.execute is expensive, then we are repeating the same
action just to make different assertions. We can reduce this repetition by
combining the examples:
it 'performs the expected side-effects' do
  expect(BackgroundJob).to receive(:perform_async)
  expect { subject.execute }
    .to change(Event, :count).by(1)
    .and change { arg_0.frobulance }.to('wibble')
Be careful doing this, as this sacrifices clarity and test independence for
performance gains.
When combining tests, consider using :aggregate_failures, so that the full
results are available, and not just the first failure.
In case you’re stuck
We have a backend_testing_performance domain expertise to list people that could help refactor slow backend specs.
To find people that could help, search for backend testing performance on the Engineering Projects page, or look directly in the www-gitlab-org project.
Feature category metadata
You must set feature category metadata for each RSpec example.
Tests depending on EE license
You can use if: Gitlab.ee? or unless: Gitlab.ee? on context/spec blocks to execute tests depending on whether running with FOSS_ONLY=1.
Example: SchemaValidator reads a different path depending on the license
Tests depending on SaaS
You can use the :saas RSpec metadata tag helper on context/spec blocks to test code that only runs on GitLab.com. This helper sets Gitlab.config.gitlab['url'] to Gitlab::Saas.com_url.
Coverage
simplecov is used to generate code test coverage reports.
These are generated automatically on the CI, but not when running tests locally. To generate partial reports
when you run a spec file on your machine, set the SIMPLECOV environment variable:
SIMPLECOV=1 bundle exec rspec spec/models/repository_spec.rb
Coverage reports are generated into the coverage folder in the app root, and you can open these in your browser, for example:
firefox coverage/index.html
Use the coverage reports to ensure your tests cover 100% of your code.
System / Feature tests
Before writing a new system test,
consider this guide around their use
Feature specs should be named ROLE_ACTION_spec.rb, such as
user_changes_password_spec.rb.
Use scenario titles that describe the success and failure paths.
Avoid scenario titles that add no information, such as “successfully”.
Avoid scenario titles that repeat the feature title.
Create only the necessary records in the database
Test a happy path and a less happy path but that’s it
Every other possible path should be tested with Unit or Integration tests
Test what’s displayed on the page, not the internals of ActiveRecord models.
For instance, if you want to verify that a record was created, add
expectations that its attributes are displayed on the page, not that
Model.count increased by one.
It’s ok to look for DOM elements, but don’t abuse it, because it makes the tests
more brittle
UI testing
When testing the UI, write tests that simulate what a user sees and how they interact with the UI.
This means preferring Capybara’s semantic methods and avoiding querying by IDs, classes, or attributes.
The benefits of testing in this way are that:
It ensures all interactive elements have an accessible name.
It is more readable, as it uses more natural language.
It is less brittle, as it avoids querying by IDs, classes, and attributes, which are not visible to the user.
We strongly recommend that you query by the element’s text label instead of by ID, class name, or data-testid.
If needed, you can scope interactions within a specific area of the page by using within.
As you will likely be scoping to an element such as a div, which typically does not have a label,
you may use a data-testid selector in this case.
You can use the be_axe_clean matcher to run axe automated accessibility testing in feature tests.
Externalized contents
For RSpec tests, expectations against externalized contents should call the same
externalizing method to match the translation. For example, you should use the _
method in Ruby.
See Internationalization for GitLab - Test files (RSpec) for details.
Actions
Where possible, use more specific actions, such as the ones below.
# good
click_button _('Submit review')
click_link _('UI testing docs')
fill_in _('Search projects'), with: 'gitlab' # fill in text input with text
select _('Updated date'), from: 'Sort by' # select an option from a select input
check _('Checkbox label')
uncheck _('Checkbox label')
choose _('Radio input label')
attach_file(_('Attach a file'), '/path/to/file.png')
# bad - interactive elements must have accessible names, so
# we should be able to use one of the specific actions above
find('.group-name', text: group.name).click
find('.js-show-diff-settings').click
find('[data-testid="submit-review"]').click
find('input[type="checkbox"]').click
find('.search').native.send_keys('gitlab')
Finders
Where possible, use more specific finders, such as the ones below.
# good
find_button _('Submit review')
find_button _('Submit review'), disabled: true
find_link _('UI testing docs')
find_link _('UI testing docs'), href: docs_url
find_field _('Search projects')
find_field _('Search projects'), with: 'gitlab' # find the input field with text
find_field _('Search projects'), disabled: true
find_field _('Checkbox label'), checked: true
find_field _('Checkbox label'), unchecked: true




    

# acceptable when finding a element that is not a button, link, or field
find('[data-testid="element"]')
Matchers
Where possible, use more specific matchers, such as the ones below.
# good
expect(page).to have_button _('Submit review')
expect(page).to have_button _('Submit review'), disabled: true
expect(page).to have_button _('Notifications'), class: 'is-checked' # assert the "Notifications" GlToggle is checked
expect(page).to have_link _('UI testing docs')
expect(page).to have_link _('UI testing docs'), href: docs_url # assert the link has an href
expect(page).to have_field _('Search projects')
expect(page).to have_field _('Search projects'), disabled: true
expect(page).to have_field _('Search projects'), with: 'gitlab' # assert the input field has text
expect(page).to have_checked_field _('Checkbox label')
expect(page).to have_unchecked_field _('Radio input label')
expect(page).to have_select _('Sort by')
expect(page).to have_select _('Sort by'), selected: 'Updated date' # assert the option is selected
expect(page).to have_select _('Sort by'), options: ['Updated date', 'Created date', 'Due date'] # assert an exact list of options
expect(page).to have_select _('Sort by'), with_options: ['Created date', 'Due date'] # assert a partial list of options
expect(page).to have_text _('Some paragraph text.')
expect(page).to have_text _('Some paragraph text.'), exact: true # assert exact match
expect(page).to have_current_path 'gitlab/gitlab-test/-/issues'
expect(page).to have_title _('Not Found')
# acceptable when a more specific matcher above is not possible
expect(page).to have_css 'h2', text: 'Issue title'
expect(page).to have_css 'p', text: 'Issue description', exact: true
expect(page).to have_css '[data-testid="weight"]', text: 2
expect(page).to have_css '.atwho-view ul', visible: true
Interacting with modals
Use the within_modal helper to interact with GitLab UI modals.
include Spec::Support::Helpers::ModalHelpers
within_modal do
  expect(page).to have_link _('UI testing docs')
  fill_in _('Search projects'), with: 'gitlab'
  click_button 'Continue'
Furthermore, you can use accept_gl_confirm for confirmation modals that only need to be accepted.
This is helpful when migrating window.confirm() to confirmAction.
include Spec::Support::Helpers::ModalHelpers
accept_gl_confirm do
  click_button 'Delete user'
You can also pass the expected confirmation message and button text to accept_gl_confirm.
include Spec::Support::Helpers::ModalHelpers
accept_gl_confirm('Are you sure you want to delete this user?', button_text: 'Delete') do
  click_button 'Delete user'
Other useful methods
After you retrieve an element using a finder method, you can invoke a number of
element methods
on it, such as hover.
Capybara tests also have a number of session methods available, such as accept_confirm.
Some other useful methods are shown below:
refresh # refresh the page
send_keys([:shift, 'i']) # press Shift+I keys to go to the Issues dashboard page
current_window.resize_to(1000, 1000) # resize the window
scroll_to(find_field('Comment')) # scroll to an element
You can also find a number of GitLab custom helpers in the spec/support/helpers/ directory.
Live debug
Sometimes you may need to debug Capybara tests by observing browser behavior.
You can pause Capybara and view the website on the browser by using the
live_debug method in your spec. The current page is automatically opened
in your default browser.
You may need to sign in first (the current user’s credentials are displayed in
the terminal).
To resume the test run, press any key.
For example:
$ bin/rspec spec/features/auto_deploy_spec.rb:34
Running via Spring preloader in process 8999
Run options: include {:locations=>{"./spec/features/auto_deploy_spec.rb"=>[34]}}
Current example is paused for live debugging
The current user credentials are: user2 / 12345678
Press any key to resume the execution of the example!
Back to the example!
Finished in 34.51 seconds (files took 0.76702 seconds to load)
1 example, 0 failures
live_debug only works on JavaScript enabled specs.
Run :js spec in a visible browser
Run the spec with WEBDRIVER_HEADLESS=0, like this:
WEBDRIVER_HEADLESS=0 bin/rspec some_spec.rb
The test completes quickly, but this gives you an idea of what’s happening.
Using live_debug with WEBDRIVER_HEADLESS=0 pauses the open browser, and does not
open the page again. This can be used to debug and inspect elements.
You can also add byebug or binding.pry to pause execution and step through
the test.
Screenshots
We use the capybara-screenshot gem to automatically take a screenshot on
failure. In CI you can download these files as job artifacts.
Also, you can manually take screenshots at any point in a test by adding the
methods below. Be sure to remove them when they are no longer needed! See
https://github.com/mattheworiordan/capybara-screenshot#manual-screenshots for
more.
Add screenshot_and_save_page in a :js spec to screenshot what Capybara
“sees”, and save the page source.
Add screenshot_and_open_image in a :js spec to screenshot what Capybara
“sees”, and automatically open the image.
The HTML dumps created by this are missing CSS.
This results in them looking very different from the actual application.
There is a small hack to add CSS which makes debugging easier.
Fast unit tests
Some classes are well-isolated from Rails. You should be able to test them
without the overhead added by the Rails environment and Bundler’s :default
group’s gem loading. In these cases, you can require 'fast_spec_helper'
instead of require 'spec_helper' in your test file, and your test should run
really fast because:
Gem loading is skipped
Rails app boot is skipped
GitLab Shell and Gitaly setup are skipped
Test repositories setup are skipped
fast_spec_helper also support autoloading classes that are located inside the
lib/ directory. If your class or module is using only
code from the lib/ directory, you don’t need to explicitly load any
dependencies. fast_spec_helper also loads all ActiveSupport extensions,
including core extensions that are commonly used in the Rails environment.
Note that in some cases, you might still have to load some dependencies using
require_dependency when a code is using gems or a dependency is not located
in lib/.





    
For example, if you want to test your code that is calling the
Gitlab::UntrustedRegexp class, which under the hood uses re2 library, you
should either:
Add require_dependency 're2' to files in your library that need re2 gem,
to make this requirement explicit. This approach is preferred.
Add it to the spec itself.
Use rubocop_spec_helper for RuboCop related specs.
It takes around one second to load tests that are using fast_spec_helper
instead of 30+ seconds in case of a regular spec_helper.
To verify that code and its specs are well-isolated from Rails, run the spec
individually via bin/rspec. Don’t use bin/spring rspec as it loads
spec_helper automatically.
subject and let variables
The GitLab RSpec suite has made extensive use of let(along with its strict, non-lazy
version let!) variables to reduce duplication. However, this sometimes comes at the cost of clarity,
so we need to set some guidelines for their use going forward:
let! variables are preferable to instance variables. let variables
are preferable to let! variables. Local variables are preferable to
let variables.
Use let to reduce duplication throughout an entire spec file.
Don’t use let to define variables used by a single test; define them as
local variables inside the test’s it block.
Don’t define a let variable inside the top-level describe block that’s
only used in a more deeply-nested context or describe block. Keep the
definition as close as possible to where it’s used.
Try to avoid overriding the definition of one let variable with another.
Don’t define a let variable that’s only used by the definition of another.
Use a helper method instead.
let! variables should be used only in case if strict evaluation with defined
order is required, otherwise let suffices. Remember that let is lazy and won’t
be evaluated until it is referenced.
Avoid referencing subject in examples. Use a named subject subject(:name), or a let variable instead, so
the variable has a contextual name.
If the subject is never referenced inside examples, then it’s acceptable to define the subject without a name.
Common test setup
let_it_be and before_all do not work with DatabaseCleaner’s deletion strategy. This includes migration specs, Rake task specs, and specs that have the :delete RSpec metadata tag.
For more information, see issue 420379.

In some cases, there is no need to recreate the same object for tests
again for each example. For example, a project and a guest of that project
are needed to test issues on the same project, so one project and user are enough for the entire file.
As much as possible, do not implement this using before(:all) or before(:context). If you do,
you would need to manually clean up the data as those hooks run outside a database transaction.
Instead, this can be achieved by using
let_it_be variables and the
before_all hook
from the test-prof gem.
let_it_be(:project) { create(:project) }
let_it_be(:user) { create(:user) }
before_all do
  project.add_guest(user)
This results in only one Project, User, and ProjectMember created for this context.
let_it_be and before_all are also available in nested contexts. Cleanup after the context
is handled automatically using a transaction rollback.
Note that if you modify an object defined inside a let_it_be block,
then you must do one of the following:
Reload the object as needed.
Use the let_it_be_with_reload alias.
Specify the reload option to reload for every example.
let_it_be_with_reload(:project) { create(:project) }
let_it_be(:project, reload: true) { create(:project) }
You can also use the let_it_be_with_refind alias, or specify the refind
option as well to completely load a new object.
let_it_be_with_refind(:project) { create(:project) }
let_it_be(:project, refind: true) { create(:project) }
Note that let_it_be cannot be used with factories that has stubs, such as allow.
The reason is that let_it_be happens in a before(:all) block, and RSpec does not
allow stubs in before(:all).
See this issue for more details.
To resolve, use let, or change the factory to not use stubs.
Time-sensitive tests
ActiveSupport::Testing::TimeHelpers
can be used to verify things that are time-sensitive. Any test that exercises or verifies something time-sensitive
should make use of these helpers to prevent transient test failures.
Example:
it 'is overdue' do
  issue = build(:issue, due_date: Date.tomorrow)
  travel_to(3.days.from_now) do
    expect(issue).to be_overdue
RSpec helpers
You can use the :freeze_time and :time_travel_to RSpec metadata tag helpers to help reduce the amount of
boilerplate code needed to wrap entire specs with the ActiveSupport::Testing::TimeHelpers
methods.
describe 'specs which require time to be frozen', :freeze_time do
  it 'freezes time' do
    right_now = Time.now
    expect(Time.now).to eq(right_now)
describe 'specs which require time to be frozen to a specific date and/or time', time_travel_to: '2020-02-02 10:30:45 -0700' do
  it 'freezes time to the specified date and time' do
    expect(Time.now).to eq(Time.new(2020, 2, 2, 17, 30, 45, '+00:00'))
Under the hood, these helpers use the around(:each) hook and the block syntax of the
ActiveSupport::Testing::TimeHelpers
methods:
around(:each) do |example|
  freeze_time { example.run }
around(:each) do |example|
  travel_to(date_or_time) { example.run }
Remember that any objects created before the examples run (such as objects created via let_it_be) will be outside spec scope.
If the time for everything needs to be frozen, before :all can be used to encapsulate the setup as well.
before :all do
  freeze_time
after :all do
  unfreeze_time
Timestamp truncation
Active Record timestamps are set by the Rails’ ActiveRecord::Timestamp
module using Time.now.
Time precision is OS-dependent,
and as the docs state, may include fractional seconds.
When Rails models are saved to the database,
any timestamps they have are stored using a type in PostgreSQL called timestamp without time zone,
which has microsecond resolution—i.e., six digits after the decimal.
So if 1577987974.6472975 is sent to PostgreSQL,
it truncates the last digit of the fractional part and instead saves 1577987974.647297.
The results of this can be a simple test like:
let_it_be(:contact) { create(:contact) }
data = Gitlab::HookData::IssueBuilder.new(issue).build
expect(data).to include('customer_relations_contacts' => [contact.hook_attrs])
Failing with an error along the lines of:
expected {
"assignee_id" => nil, "...1 +0000 } to include {"customer_relations_contacts" => [{:created_at => "2023-08-04T13:30:20Z", :first_name => "Sidney Jones3" }]}
Diff:
       @@ -1,35 +1,69 @@
       -"customer_relations_contacts" => [{:created_at=>"2023-08-04T13:30:20Z", :first_name=>"Sidney Jones3" }],
       +"customer_relations_contacts" => [{"




    
created_at"=>2023-08-04 13:30:20.245964000 +0000, "first_name"=>"Sidney Jones3" }],
The fix is to ensure we .reload the object from the database to get the timestamp with correct precision:
let_it_be(:contact) { create(:contact) }
data = Gitlab::HookData::IssueBuilder.new(issue).build
expect(data).to include('customer_relations_contacts' => [contact.reload.hook_attrs])
This explanation was taken from a blog post
by Maciek Rząsa.
You can see a merge request
where this problem arose and the backend pairing session
where it was discussed.
Feature flags in tests
This section was moved to developing with feature flags.
Pristine test environments
The code exercised by a single GitLab test may access and modify many items of
data. Without careful preparation before a test runs, and cleanup afterward,
a test can change data in a way that affects the behavior of
following tests. This should be avoided at all costs! Fortunately, the existing
test framework handles most cases already.
When the test environment does get polluted, a common outcome is
flaky tests. Pollution often manifests as an order
dependency: running spec A followed by spec B reliably fails, but running
spec B followed by spec A reliably succeeds. In these cases, you can use
rspec --bisect (or a manual pairwise bisect of spec files) to determine which
spec is at fault. Fixing the problem requires some understanding of how the test
suite ensures the environment is pristine. Read on to discover more about each
data store!
SQL database
This is managed for us by the database_cleaner gem. Each spec is surrounded in
a transaction, which is rolled back after the test completes. Certain specs
instead issue DELETE FROM queries against every table after completion. This
allows the created rows to be viewed from multiple database connections, which
is important for specs that run in a browser, or migration specs, among others.
One consequence of using these strategies, instead of the well-known
TRUNCATE TABLES approach, is that primary keys and other sequences are not
reset across specs. So if you create a project in spec A, then create a project
in spec B, the first has id=1, while the second has id=2.
This means that specs should never rely on the value of an ID, or any other
sequence-generated column. To avoid accidental conflicts, specs should also
avoid manually specifying any values in these kinds of columns. Instead, leave
them unspecified, and look up the value after the row is created.
TestProf in migration specs
Because of what is described above, migration specs can’t be run inside
a database transaction. Our test suite uses
TestProf to improve the runtime of the
test suite, but TestProf uses database transactions to perform these optimizations.
For this reason, we can’t use TestProf methods in our migration specs.
These are the methods that should not be used and should be replaced with
default RSpec methods instead:
let_it_be: use let or let! instead.
let_it_be_with_reload: use let or let! instead.
let_it_be_with_refind: use let or let! instead.
before_all: use before or before(:all) instead.
Redis
GitLab stores two main categories of data in Redis: cached items, and Sidekiq
jobs. View the full list of Gitlab::Redis::Wrapper descendants that are backed by
a separate Redis instance.
In most specs, the Rails cache is actually an in-memory store. This is replaced
between specs, so calls to Rails.cache.read and Rails.cache.write are safe.
However, if a spec makes direct Redis calls, it should mark itself with the
:clean_gitlab_redis_cache, :clean_gitlab_redis_shared_state or
:clean_gitlab_redis_queues traits as appropriate.
Background jobs / Sidekiq
By default, Sidekiq jobs are enqueued into a jobs array and aren’t processed.
If a test queues Sidekiq jobs and need them to be processed, the
:sidekiq_inline trait can be used.
The :sidekiq_might_not_need_inline trait was added when
Sidekiq inline mode was changed to fake mode
to all the tests that needed Sidekiq to actually process jobs. Tests with
this trait should be either fixed to not rely on Sidekiq processing jobs, or their
:sidekiq_might_not_need_inline trait should be updated to :sidekiq_inline if
the processing of background jobs is needed/expected.
The usage of perform_enqueued_jobs is useful only for testing delayed mail
deliveries, because our Sidekiq workers aren’t inheriting from ApplicationJob
/ ActiveJob::Base.
DNS requests are stubbed universally in the test suite
(as of !22368), as DNS can
cause issues depending on the developer’s local network. There are RSpec labels
available in spec/support/dns.rb which you can apply to tests if you need to
bypass the DNS stubbing, like this:
it "really connects to Prometheus", :permit_dns do
And if you need more specific control, the DNS blocking is implemented in
spec/support/helpers/dns_helpers.rb and these methods can be called elsewhere.
Rate Limiting
Rate limiting is enabled in the test suite. Rate limits
may be triggered in feature specs that use the :js trait. In most cases, triggering rate
limiting can be avoided by marking the spec with the :clean_gitlab_redis_rate_limiting
trait. This trait clears the rate limiting data stored in Redis cache between specs. If
a single test triggers the rate limit, the :disable_rate_limit can be used instead.
Stubbing File methods
In the situations where you need to
the contents of a file use the stub_file_read, and
expect_file_read helper methods which handle the stubbing for
File.read correctly. These methods stub File.read for the given
filename, and also stub File.exist? to return true.
If you need to manually stub File.read for any reason be sure to:
Stub and call the original implementation for other file paths.
Then stub File.read for only the file path you are interested in.
Otherwise File.read calls from other parts of the codebase get
stubbed incorrectly.
# bad, all Files will read and return nothing
allow(File).to receive(:read)
# good
stub_file_read(my_filepath, content: "fake file content")
# also OK
allow(File).to receive(:read).and_call_original
allow(File).to receive(:read).with(my_filepath).and_return("fake file_content")
File system
File system data can be roughly split into “repositories”, and “everything else”.
Repositories are stored in tmp/tests/repositories. This directory is emptied
before a test run starts, and after the test run ends. It is not emptied between
specs, so created repositories accumulate in this directory over the
lifetime of the process. Deleting them is expensive, but this could lead to
pollution unless carefully managed.
To avoid this, hashed storage
is enabled in the test suite. This means that repositories are given a unique
path that depends on their project’s ID. Because the project IDs are not reset
between specs, each spec gets its own repository on disk,
and prevents changes from being visible between specs.
If a spec manually specifies a project ID, or inspects the state of the
tmp/tests/repositories/ directory directly, then it should clean up the
directory both before and after it runs. In general, these patterns should be
completely avoided.
Other classes of file linked to database objects, such as uploads, are generally
managed in the same way. With hashed storage enabled in the specs, they are
written to disk in locations determined by ID, so conflicts should not occur.
Some specs disable hashed storage by passing the :legacy_storage trait to the
projects factory. Specs that do this must never override the path of the
project, or any of its groups. The default path includes the project ID, so it
does not conflict. If two specs create a :legacy_storage project with the same
path, they use the same repository on disk and lead to test environment
pollution.
Other files must be managed manually by the spec. If you run code that creates a
tmp/test-file.csv file, for instance, the spec must ensure that the file is
removed as part of cleanup.
Persistent in-memory application state
All the specs in a given rspec run share the same Ruby process, which means
they can affect each other by modifying Ruby objects that are accessible between
specs. In practice, this means global variables, and constants (which includes
Ruby classes, modules, etc).
Global variables should generally not be modified. If absolutely necessary, a
block like this can be used to ensure the change is rolled back afterwards:
around(:each) do |example|
  old_value = $0
  begin
    $0 = "new-value"
    example.run
  ensure
    $0 = old_value
If a spec needs to modify a constant, it should use the stub_const helper to
ensure the change is rolled back.
If you need to modify the contents of the ENV constant, you can use the
stub_env helper method instead.
While most Ruby instances are not shared between specs, classes
and modules generally are. Class and module instance variables, accessors,
class variables, and other stateful idioms, should be treated in the same way as
global variables. Don’t modify them unless you have to! In particular, prefer
using expectations, or dependency injection along with stubs, to avoid the need
for modifications. If you have no other choice, an around block like the global
variables example can be used, but avoid this if at all possible.
Elasticsearch specs
Specs that require Elasticsearch must be marked with the :elastic trait. This
creates and deletes indices before and after all examples.
The :elastic_delete_by_query trait was added to reduce runtime for pipelines by creating and deleting indices at the
start and end of each context only. The Elasticsearch delete by query API
is used to delete data in all indices (except the migrations index) between examples to ensure a clean index.
The :elastic_clean trait creates and deletes indices between examples to ensure a clean index. This way, tests are not
polluted with non-essential data. If using the :elastic or :elastic_delete_by_query trait
is causing issues, use :elastic_clean instead. :elastic_clean is significantly slower than the other traits
and should be used sparingly.
Most tests for Elasticsearch logic relate to:
Creating data in PostgreSQL and waiting for it to be indexed in Elasticsearch.
Searching for that data.
Ensuring that the test gives the expected result.
There are some exceptions, such as checking for structural changes rather than individual records in an index.
Elasticsearch indexing uses Gitlab::Redis::SharedState.
Therefore, the Elasticsearch traits dynamically use the :clean_gitlab_redis_shared_state trait.
You do not need to add :clean_gitlab_redis_shared_state manually.
Specs using Elasticsearch require that you:
Create data in PostgreSQL and then index it into Elasticsearch.
Enable Application Settings for Elasticsearch (which is disabled by default).
To do so, use:
before do
  stub_ee_application_setting(elasticsearch_search: true, elasticsearch_indexing: true)
Additionally, you can use the ensure_elasticsearch_index! method to overcome the asynchronous nature of Elasticsearch.
It uses the Elasticsearch Refresh API
to make sure all operations performed on an index since the last refresh are available for search. This method is typically
called after loading data into PostgreSQL to ensure the data is indexed and searchable.
You can use the SEARCH_SPEC_BENCHMARK environment variable to benchmark test setup steps:
SEARCH_SPEC_BENCHMARK=1 bundle exec rspec ee/spec/lib/elastic/latest/merge_request_class_proxy_spec.rb
Test Snowplow events
Snowplow performs runtime type checks by using the contracts gem.
Because Snowplow is by default disabled in tests and development, it can be hard to
catch exceptions when mocking Gitlab::Tracking.
To catch runtime errors due to type checks you can use expect_snowplow_event, which checks for
calls to Gitlab::Tracking#event.
describe '#show' do
  it 'tracks snowplow events' do
    get :show
    expect_snowplow_event(
      category: 'Experiment',
      action: 'start',
      namespace: group,
      project: project
    expect_snowplow_event(
      category: 'Experiment',
      action: 'sent',
      property: 'property',
      label: 'label',
      namespace: group,
      project: project
When you want to ensure that no event got called, you can use expect_no_snowplow_event.
  describe '#show' do
    it 'does not track any snowplow events' do
      get :show
      expect_no_snowplow_event(category: described_class.name, action: 'some_action')
Even though category and action can be omitted, you should at least
specify a category to avoid flaky tests. For example,
Users::ActivityService may track a Snowplow event after an API
request, and expect_no_snowplow_event will fail if that happens to run
when no arguments are specified.
Test Snowplow context against the schema
The Snowplow schema matcher
helps to reduce validation errors by testing Snowplow context against the JSON schema.
The schema matcher accepts the following parameters:
schema path
context
To add a schema matcher spec:
Add a new schema to the Iglu repository,
then copy the same schema to the spec/fixtures/product_intelligence/ directory.
In the copied schema, remove the "$schema" key and value. We do not need it for specs
and the spec fails if we keep the key, as it tries to look for the schema in the URL.
Use the following snippet to call the schema matcher:
match_snowplow_context_schema(schema_path: '<filename from step 1>', context: <Context Hash> )
Table-based / Parameterized tests
This style of testing is used to exercise one piece of code with a comprehensive
range of inputs. By specifying the test case once, alongside a table of inputs
and the expected output for each, your tests can be made easier to read and more
compact.
We use the RSpec::Parameterized
gem. A short example, using the table syntax and checking Ruby equality for a
range of inputs, might look like this:
describe "#==" do
  using RSpec::Parameterized::TableSyntax
  let(:one) { 1 }
  let(:two) { 2 }
  where(:a, :b, :result) do
    1         | 1         | true
    1         | 2         | false
    true      | true      | true
    true      | false     | false
    ref(:one) | ref(:one) | true  # let variables must be referenced using `ref`
    ref(:one) | ref(:two) | false
  with_them do
    it { expect(a == b).to eq(result) }
    it 'is isomorphic' do
      expect(b == a).to eq(result)
If, after creating a table-based test, you see an error that looks like this:
NoMethodError:
  undefined method `to_params'
  param_sets = extracted.is_a?(Array) ? extracted : extracted.to_params
                                                                       ^^^^^^^^^^
  Did you mean?  to_param
That indicates that you need to include the line using RSpec::Parameterized::TableSyntax in the spec file.
Only use simple values as input in the where block. Using procs, stateful
objects, FactoryBot-created objects, and similar items can lead to
unexpected results.
Prometheus tests
Prometheus metrics may be preserved from one test run to another. To ensure that metrics are
reset before each example, add the :prometheus tag to the RSpec test.
Matchers
Custom matchers should be created to clarify the intent and/or hide the
complexity of RSpec expectations. They should be placed under
spec/support/matchers/. Matchers can be placed in subfolder if they apply to
a certain type of specs only (such as features or requests) but shouldn’t be if
they apply to multiple type of specs.
be_like_time
Time returned from a database can differ in precision from time objects
in Ruby, so we need flexible tolerances when comparing in specs.
The PostgreSQL time and timestamp types
have the resolution of 1 microsecond.
However, the precision of Ruby Time can vary depending on the OS.
Consider the following snippet: