gitlab-org--gitlab-foss/doc/development/ee_features.md
Tim Zallmann be1523c1ba Merge branch 'add-documentation-about-backporting-changes' into 'master'
Add documentation about backporting changes from EE to CE

See merge request gitlab-org/gitlab-ce!18003
2018-04-03 13:42:24 +00:00

20 KiB

Guidelines for implementing Enterprise Edition features

  • Write the code and the tests.: As with any code, EE features should have good test coverage to prevent regressions.
  • Write documentation.: Add documentation to the doc/ directory. Describe the feature and include screenshots, if applicable.
  • Submit a MR to the www-gitlab-com project.: Add the new feature to the [EE features list][ee-features-list].

Act as CE when unlicensed

Since the implementation of GitLab CE features to work with unlicensed EE instance GitLab Enterprise Edition should work like GitLab Community Edition when no license is active. So EE features always should be guarded by project.feature_available? or group.feature_available? (or License.feature_available? if it is a system-wide feature).

CE specs should remain untouched as much as possible and extra specs should be added for EE. Licensed features can be stubbed using the spec helper stub_licensed_features in EE::LicenseHelpers.

Separation of EE code

We want a single code base eventually, but before we reach the goal, we still need to merge changes from GitLab CE to EE. To help us get there, we should make sure that we no longer edit CE files in place in order to implement EE features.

Instead, all EE code should be put inside the ee/ top-level directory. The rest of the code should be as close to the CE files as possible.

EE-specific comments

When complete separation can't be achieved with the ee/ directory, you can wrap code in EE specific comments to designate the difference from CE/EE and add some context for someone resolving a conflict.

# EE-specific start
stub_licensed_features(variable_environment_scope: true)
# EE specific end
-# EE-specific start
= render 'ci/variables/environment_scope', form_field: form_field, variable: variable
-# EE-specific end

EE-specific comments should not be backported to CE.

Detection of EE-only files

For each commit (except on master), the ee-files-location-check CI job tries to detect if there are any new files that are EE-only. If any file is detected, the job fails with an explanation of why and what to do to make it pass.

Basically, the fix is simple: git mv <file> ee/<file>.

How to name your branches?

For any EE branch, the job will try to detect its CE counterpart by removing any ee- prefix or -ee suffix from the EE branch name, and matching the last branch that contains it.

For instance, from the EE branch new-shiny-feature-ee (or ee-new-shiny-feature), the job would find the corresponding CE branches:

  • new-shiny-feature
  • ce-new-shiny-feature
  • new-shiny-feature-ce
  • my-super-new-shiny-feature-in-ce

Whitelist some EE-only files that cannot be moved to ee/

The ee-files-location-check CI job provides a whitelist of files or folders that cannot or should not be moved to ee/. Feel free to open an issue to discuss adding a new file/folder to this whitelist.

For instance, it was decided that moving EE-only files from qa/ to ee/qa/ would make it difficult to build the gitLab-{ce,ee}-qa Docker images and it was not worth the complexity.

EE-only features

If the feature being developed is not present in any form in CE, we don't need to put the codes under EE namespace. For example, an EE model could go into: ee/app/models/awesome.rb using Awesome as the class name. This is applied not only to models. Here's a list of other examples:

  • ee/app/controllers/foos_controller.rb
  • ee/app/finders/foos_finder.rb
  • ee/app/helpers/foos_helper.rb
  • ee/app/mailers/foos_mailer.rb
  • ee/app/models/foo.rb
  • ee/app/policies/foo_policy.rb
  • ee/app/serializers/foo_entity.rb
  • ee/app/serializers/foo_serializer.rb
  • ee/app/services/foo/create_service.rb
  • ee/app/validators/foo_attr_validator.rb
  • ee/app/workers/foo_worker.rb

This works because for every path that are present in CE's eager-load/auto-load paths, we add the same ee/-prepended path in config/application.rb.

EE features based on CE features

For features that build on existing CE features, write a module in the EE namespace and prepend it in the CE class. This makes conflicts less likely to happen during CE to EE merges because only one line is added to the CE class - the prepend line.

Since the module would require an EE namespace, the file should also be put in an ee/ sub-directory. For example, we want to extend the user model in EE, so we have a module called ::EE::User put inside ee/app/models/ee/user.rb.

This is also not just applied to models. Here's a list of other examples:

  • ee/app/controllers/ee/foos_controller.rb
  • ee/app/finders/ee/foos_finder.rb
  • ee/app/helpers/ee/foos_helper.rb
  • ee/app/mailers/ee/foos_mailer.rb
  • ee/app/models/ee/foo.rb
  • ee/app/policies/ee/foo_policy.rb
  • ee/app/serializers/ee/foo_entity.rb
  • ee/app/serializers/ee/foo_serializer.rb
  • ee/app/services/ee/foo/create_service.rb
  • ee/app/validators/ee/foo_attr_validator.rb
  • ee/app/workers/ee/foo_worker.rb

Overriding CE methods

To override a method present in the CE codebase, use prepend. It lets you override a method in a class with a method from a module, while still having access the class's implementation with super.

There are a few gotchas with it:

  • you should always extend ::Gitlab::Utils::Override and use override to guard the "overrider" method to ensure that if the method gets renamed in CE, the EE override won't be silently forgotten.
  • when the "overrider" would add a line in the middle of the CE implementation, you should refactor the CE method and split it in smaller methods. Or create a "hook" method that is empty in CE, and with the EE-specific implementation in EE.
  • when the original implementation contains a guard clause (e.g. return unless condition), we cannot easily extend the behaviour by overriding the method, because we can't know when the overridden method (i.e. calling super in the overriding method) would want to stop early. In this case, we shouldn't just override it, but update the original method to make it call the other method we want to extend, like a template method pattern. For example, given this base:
      class Base
        def execute
          return unless enabled?
    
          # ...
          # ...
        end
      end
    
    Instead of just overriding Base#execute, we should update it and extract the behaviour into another method:
      class Base
        def execute
          return unless enabled?
    
          do_something
        end
    
        private
    
        def do_something
          # ...
          # ...
        end
      end
    
    Then we're free to override that do_something without worrying about the guards:
      module EE::Base
        extend ::Gitlab::Utils::Override
    
        override :do_something
        def do_something
          # Follow the above pattern to call super and extend it
        end
      end
    
    This would require updating CE first, or make sure this is back ported to CE.

When prepending, place them in the ee/ specific sub-directory, and wrap class or module in module EE to avoid naming conflicts.

For example to override the CE implementation of ApplicationController#after_sign_out_path_for:

def after_sign_out_path_for(resource)
  current_application_settings.after_sign_out_path.presence || new_user_session_path
end

Instead of modifying the method in place, you should add prepend to the existing file:

class ApplicationController < ActionController::Base
  prepend EE::ApplicationController
  # ...

  def after_sign_out_path_for(resource)
    current_application_settings.after_sign_out_path.presence || new_user_session_path
  end

  # ...
end

And create a new file in the ee/ sub-directory with the altered implementation:

module EE
  module ApplicationController
    extend ::Gitlab::Utils::Override

    override :after_sign_out_path_for
    def after_sign_out_path_for(resource)
      if Gitlab::Geo.secondary?
        Gitlab::Geo.primary_node.oauth_logout_url(@geo_logout_state)
      else
        super
      end
    end
  end
end

Use self-descriptive wrapper methods

When it's not possible/logical to modify the implementation of a method. Wrap it in a self-descriptive method and use that method.

For example, in CE only an admin is allowed to access all private projects/groups, but in EE also an auditor has full private access. It would be incorrect to override the implementation of User#admin?, so instead add a method full_private_access? to app/models/users.rb. The implementation in CE will be:

def full_private_access?
  admin?
end

In EE, the implementation ee/app/models/ee/users.rb would be:

override :full_private_access?
def full_private_access?
  super || auditor?
end

In lib/gitlab/visibility_level.rb this method is used to return the allowed visibilty levels:

def levels_for_user(user = nil)
  if user.full_private_access?
    [PRIVATE, INTERNAL, PUBLIC]
  elsif # ...
end

See CE MR and EE MR for full implementation details.

Code in app/controllers/

In controllers, the most common type of conflict is with before_action that has a list of actions in CE but EE adds some actions to that list.

The same problem often occurs for params.require / params.permit calls.

Mitigations

Separate CE and EE actions/keywords. For instance for params.require in ProjectsController:

def project_params
  params.require(:project).permit(project_params_attributes)
end

# Always returns an array of symbols, created however best fits the use case.
# It _should_ be sorted alphabetically.
def project_params_attributes
  %i[
    description
    name
    path
  ]
end

In the EE::ProjectsController module:

def project_params_attributes
  super + project_params_attributes_ee
end

def project_params_attributes_ee
  %i[
    approvals_before_merge
    approver_group_ids
    approver_ids
    ...
  ]
end

Code in app/models/

EE-specific models should extend EE::Model.

For example, if EE has a specific Tanuki model, you would place it in ee/app/models/ee/tanuki.rb.

Code in app/views/

It's a very frequent problem that EE is adding some specific view code in a CE view. For instance the approval code in the project's settings page.

Mitigations

Blocks of code that are EE-specific should be moved to partials. This avoids conflicts with big chunks of HAML code that that are not fun to resolve when you add the indentation to the equation.

EE-specific views should be placed in ee/app/views/ee/, using extra sub-directories if appropriate.

Code in lib/

Place EE-specific logic in the top-level EE module namespace. Namespace the class beneath the EE module just as you would normally.

For example, if CE has LDAP classes in lib/gitlab/ldap/ then you would place EE-specific LDAP classes in ee/lib/ee/gitlab/ldap.

Code in lib/api/

It can be very tricky to extend EE features by a single line of prepend, and for each different Grape feature, we might need different strategies to extend it. To apply different strategies easily, we would use extend ActiveSupport::Concern in the EE module.

Put the EE module files following EE features based on CE features.

EE API routes

For EE API routes, we put them in a prepended block:

module EE
  module API
    module MergeRequests
      extend ActiveSupport::Concern

      prepended do
        params do
          requires :id, type: String, desc: 'The ID of a project'
        end
        resource :projects, requirements: ::API::API::PROJECT_ENDPOINT_REQUIREMENTS do
          # ...
        end
      end
    end
  end
end

Note that due to namespace differences, we need to use the full qualifier for some constants.

EE params

We can define params and utilize use in another params definition to include params defined in EE. However, we need to define the "interface" first in CE in order for EE to override it. We don't have to do this in other places due to prepend, but Grape is complex internally and we couldn't easily do that, so we'll follow regular object-oriented practices that we define the interface first here.

For example, suppose we have a few more optional params for EE, given this CE API code:

module API
  class MergeRequests < Grape::API
    # EE::API::MergeRequests would override the following helpers
    helpers do
      params :optional_params_ee do
      end
    end

    prepend EE::API::MergeRequests

    params :optional_params do
      # CE specific params go here...

      use :optional_params_ee
    end
  end
end

And then we could override it in EE module:

module EE
  module API
    module MergeRequests
      extend ActiveSupport::Concern

      prepended do
        helpers do
          params :optional_params_ee do
            # EE specific params go here...
          end
        end
      end
    end
  end
end

This way, the only difference between CE and EE for that API file would be prepend EE::API::MergeRequests.

EE helpers

To make it easy for an EE module to override the CE helpers, we need to define those helpers we want to extend first. Try to do that immediately after the class definition to make it easy and clear:

module API
  class JobArtifacts < Grape::API
    # EE::API::JobArtifacts would override the following helpers
    helpers do
      def authorize_download_artifacts!
        authorize_read_builds!
      end
    end

    prepend EE::API::JobArtifacts
  end
end

And then we can follow regular object-oriented practices to override it:

module EE
  module API
    module JobArtifacts
      extend ActiveSupport::Concern

      prepended do
        helpers do
          def authorize_download_artifacts!
            super
            check_cross_project_pipelines_feature!
          end
        end
      end
    end
  end
end

EE-specific behaviour

Sometimes we need EE-specific behaviour in some of the APIs. Normally we could use EE methods to override CE methods, however API routes are not methods and therefore can't be simply overridden. We need to extract them into a standalone method, or introduce some "hooks" where we could inject behavior in the CE route. Something like this:

module API
  class MergeRequests < Grape::API
    helpers do
      # EE::API::MergeRequests would override the following helpers
      def update_merge_request_ee(merge_request)
      end
    end

    prepend EE::API::MergeRequests

    put ':id/merge_requests/:merge_request_iid/merge' do
      merge_request = find_project_merge_request(params[:merge_request_iid])

      # ...

      update_merge_request_ee(merge_request)

      # ...
    end
  end
end

Note that update_merge_request_ee doesn't do anything in CE, but then we could override it in EE:

module EE
  module API
    module MergeRequests
      extend ActiveSupport::Concern

      prepended do
        helpers do
          def update_merge_request_ee(merge_request)
            # ...
          end
        end
      end
    end
  end
end

EE route_setting

It's very hard to extend this in an EE module, and this is simply storing some meta-data for a particular route. Given that, we could simply leave the EE route_setting in CE as it won't hurt and we are just not going to use those meta-data in CE.

We could revisit this policy when we're using route_setting more and whether or not we really need to extend it from EE. For now we're not using it much.

Utilizing class methods for setting up EE-specific data

Sometimes we need to use different arguments for a particular API route, and we can't easily extend it with an EE module because Grape has different context in different blocks. In order to overcome this, we could use class methods from the API class.

For example, in one place we need to pass an extra argument to at_least_one_of so that the API could consider an EE-only argument as the least argument. This is not quite beautiful but it's working:

module API
  class MergeRequests < Grape::API
    def self.update_params_at_least_one_of
      %i[
        assignee_id
        description
      ]
    end

    prepend EE::API::MergeRequests

    params do
      at_least_one_of(*::API::MergeRequests.update_params_at_least_one_of)
    end
  end
end

And then we could easily extend that argument in the EE class method:

module EE
  module API
    module MergeRequests
      extend ActiveSupport::Concern

      class_methods do
        def update_params_at_least_one_of
          super.push(*%i[
            squash
          ])
        end
      end
    end
  end
end

It could be annoying if we need this for a lot of routes, but it might be the simplest solution right now.

Code in spec/

When you're testing EE-only features, avoid adding examples to the existing CE specs. Also do no change existing CE examples, since they should remain working as-is when EE is running without a license.

Instead place EE specs in the ee/spec folder.

JavaScript code in assets/javascripts/

To separate EE-specific JS-files we should also move the files into an ee folder.

For example there can be an app/assets/javascripts/protected_branches/protected_branches_bundle.js and an EE counterpart ee/app/assets/javascripts/protected_branches/protected_branches_bundle.js.

See the frontend guide performance section for information on managing page-specific javascript within EE.

SCSS code in assets/stylesheets

To separate EE-specific styles in SCSS files, if a component you're adding styles for is limited to only EE, it is better to have a separate SCSS file in appropriate directory within app/assets/stylesheets. See backporting changes for instructions on how to merge changes safely.

In some cases, this is not entirely possible or creating dedicated SCSS file is an overkill, e.g. a text style of some component is different for EE. In such cases, styles are usually kept in stylesheet that is common for both CE and EE, and it is wise to isolate such ruleset from rest of CE rules (along with adding comment describing the same) to avoid conflicts during CE to EE merge.

Bad

.section-body {
  .section-title {
    background: $gl-header-color;
  }

  &.ee-section-body {
    .section-title {
      background: $gl-header-color-cyan;
    }
  }
}

Good

.section-body {
  .section-title {
    background: $gl-header-color;
  }
}

// EE-specific start
.section-body.ee-section-body {
  .section-title {
    background: $gl-header-color-cyan;
  }
}
// EE-specific end

Backporting changes from EE to CE

When working in EE-specific features, you might have to tweak a few files that are not EE-specific. Here is a workflow to make sure those changes end up backported safely into CE too. (This approach does not refer to changes introduced via csslab.)

  1. Make your changes in the EE branch. If possible, keep a separated commit (to be squashed) to help backporting and review.
  2. Open merge request to EE project.
  3. Apply the changes you made to CE files in a branch of the CE project. (Tip: Use patch with the diff from your commit in EE branch)
  4. Open merge request to CE project, referring it's a backport of EE changes and link to MR open in EE.
  5. Once EE MR is merged, the MR towards CE can be merged. But not before.

Note: regarding SCSS, make sure the files living outside /ee/ don't diverge between CE and EE projects.

gitlab-svgs

Conflicts in app/assets/images/icons.json or app/assets/images/icons.svg can be resolved simply by regenerating those assets with yarn run svg.