Gotchas

The purpose of this guide is to document potential “gotchas” that contributors might encounter or should avoid during development of GitLab CE and EE.

Do not read files from app/assets directory

Omnibus GitLab has dropped the app/assets directory, after asset compilation. The ee/app/assets, vendor/assets directories are dropped as well.

This means that reading files from that directory fails in Omnibus-installed GitLab instances: 

file = Rails.root.join('app/assets/images/logo.svg')

# This file does not exist, read will fail with:
# Errno::ENOENT: No such file or directory @ rb_sysopen
File.read(file)

Do not assert against the absolute value of a sequence-generated attribute

Consider the following factory: 

FactoryBot.define do
  factory :label do
    sequence(:title) { |n| "label#{n}" }
  end
end

Consider the following API spec: 

require 'spec_helper'

RSpec.describe API::Labels do
  it 'creates a first label' do
    create(:label)

    get api("/projects/#{project.id}/labels", user)

    expect(response).to have_gitlab_http_status(:ok)
    expect(json_response.first['name']).to eq('label1')
  end

  it 'creates a second label' do
    create(:label)

    get api("/projects/#{project.id}/labels", user)

    expect(response).to have_gitlab_http_status(:ok)
    expect(json_response.first['name']).to eq('label1')
  end
end

When run, this spec doesn’t do what we might expect: 

1) API::API reproduce sequence issue creates a second label
   Failure/Error: expect(json_response.first['name']).to eq('label1')

     expected: "label1"
          got: "label2"

     (compared using ==)

This is because FactoryBot sequences are not reset for each example.

Remember that sequence-generated values exist only to avoid having to explicitly set attributes that have a uniqueness constraint when using a factory.

Solution

If you assert against a sequence-generated attribute’s value, you should set it explicitly. Also, the value you set shouldn’t match the sequence pattern.

For instance, using our :label factory, writing create(:label, title: 'foo') is ok, but create(:label, title: 'label1') is not.

Following is the fixed API spec: 

require 'spec_helper'

RSpec.describe API::Labels do
  it 'creates a first label' do
    create(:label, title: 'foo')

    get api("/projects/#{project.id}/labels", user)

    expect(response).to have_gitlab_http_status(:ok)
    expect(json_response.first['name']).to eq('foo')
  end

  it 'creates a second label' do
    create(:label, title: 'bar')

    get api("/projects/#{project.id}/labels", user)

    expect(response).to have_gitlab_http_status(:ok)
    expect(json_response.first['name']).to eq('bar')
  end
end

Avoid using expect_any_instance_of or allow_any_instance_of in RSpec

Why

  • Because it is not isolated therefore it might be broken at times.

  • Because it doesn’t work whenever the method we want to stub was defined in a prepended module, which is very likely the case in EE. We could see error like this: 

    1.1) Failure/Error: expect_any_instance_of(ApplicationSetting).to receive_messages(messages)
     Using `any_instance` to stub a method (elasticsearch_indexing) that has been defined on a prepended module (EE::ApplicationSetting) is not supported.

Alternatives

Instead, use any of these:

  • expect_next_instance_of
  • allow_next_instance_of
  • expect_next_found_instance_of
  • allow_next_found_instance_of

For example: 

# Don't do this:
expect_any_instance_of(Project).to receive(:add_import_job)

# Don't do this:
allow_any_instance_of(Project).to receive(:add_import_job)

We could write: 

# Do this:
expect_next_instance_of(Project) do |project|
  expect(project).to receive(:add_import_job)
end

# Do this:
allow_next_instance_of(Project) do |project|
  allow(project).to receive(:add_import_job)
end

# Do this:
expect_next_found_instance_of(Project) do |project|
  expect(project).to receive(:add_import_job)
end

# Do this:
allow_next_found_instance_of(Project) do |project|
  allow(project).to receive(:add_import_job)
end

Since Active Record is not calling the .new method on model classes to instantiate the objects, you should use expect_next_found_instance_of or allow_next_found_instance_of mock helpers to set up mock on objects returned by Active Record query & finder methods._

It is also possible to set mocks and expectations for multiple instances of the same Active Record model by using the expect_next_found_(number)_instances_of and allow_next_found_(number)_instances_of helpers, like so; 

expect_next_found_2_instances_of(Project) do |project|
  expect(project).to receive(:add_import_job)
end

allow_next_found_2_instances_of(Project) do |project|
  allow(project).to receive(:add_import_job)
end

If we also want to initialize the instance with some particular arguments, we could also pass it like: 

# Do this:
expect_next_instance_of(MergeRequests::RefreshService, project, user) do |refresh_service|
  expect(refresh_service).to receive(:execute).with(oldrev, newrev, ref)
end

This would expect the following: 

# Above expects:
refresh_service = MergeRequests::RefreshService.new(project, user)
refresh_service.execute(oldrev, newrev, ref)

Do not rescue Exception

See “Why is it bad style to rescue Exception => e in Ruby?”.

This rule is enforced automatically by RuboCop.

Do not use inline JavaScript in views

Using the inline :javascript Haml filters comes with a performance overhead. Using inline JavaScript is not a good way to structure your code and should be avoided.

We’ve removed these two filters in an initializer.

Further reading

Storing assets that do not require pre-compiling

Assets that need to be served to the user are stored under the app/assets directory, which is later pre-compiled and placed in the public/ directory.

However, you cannot access the content of any file from within app/assets from the application code, as we do not include that folder in production installations as a space saving measure. 

support_bot = Users::Internal.support_bot

# accessing a file from the `app/assets` folder
support_bot.avatar = Rails.root.join('app', 'assets', 'images', 'bot_avatars', 'support_bot.png').open

support_bot.save!

While the code above works in local environments, it errors out in production installations as the app/assets folder is not included.

Solution

The alternative is the lib/assets folder. Use it if you need to add assets (like images) to the repository that meet the following conditions:

  • The assets do not need to be directly served to the user (and hence need not be pre-compiled).
  • The assets do need to be accessed via application code.

In short:

Use app/assets for storing any asset that needs to be precompiled and served to the end user. Use lib/assets for storing any asset that does not need to be served to the end user directly, but is still required to be accessed by the application code.

MR for reference: !37671

Do not override has_many through: or has_one through: associations

Associations with the :through option should not be overridden as we could accidentally destroy the wrong object.

This is because the destroy() method behaves differently when acting on has_many through: and has_one through: associations. 

group.users.destroy(id)

The code example above reads as if we are destroying a User record, but behind the scenes, it is destroying a Member record. This is because the users association is defined on Group as a has_many through: association: 

class Group < Namespace
  has_many :group_members, -> { where(requested_at: nil).where.not(members: { access_level: Gitlab::Access::MINIMAL_ACCESS }) }, dependent: :destroy, as: :source

  has_many :users, through: :group_members
end

And Rails has the following behavior on using destroy() on such associations:

If the :through option is used, then the join records are destroyed instead, not the objects themselves.

This is why a Member record, which is the join record connecting a User and Group, is being destroyed.

Now, if we override the users association, so like: 

class Group < Namespace
  has_many :group_members, -> { where(requested_at: nil).where.not(members: { access_level: Gitlab::Access::MINIMAL_ACCESS }) }, dependent: :destroy, as: :source

  has_many :users, through: :group_members

  def users
    super.where(admin: false)
  end
end

The overridden method now changes the above behavior of destroy(), such that if we execute 

group.users.destroy(id)

a User record will be deleted, which can lead to data loss.

In short, overriding a has_many through: or has_one through: association can prove dangerous. To prevent this from happening, we are introducing an automated check in !131455.

For more information, see issue 424536.