Guidelines for reusing abstractions

As GitLab has grown, different patterns emerged across the codebase. Service classes, serializers, and presenters are just a few. These patterns made it easy to reuse code, but at the same time make it easy to accidentally reuse the wrong abstraction in a particular place.

Why these guidelines are necessary

Code reuse is good, but sometimes this can lead to shoehorning the wrong abstraction into a particular use case. This in turn can have a negative impact on maintainability, the ability to easily debug problems, or even performance.

An example would be to use ProjectsFinder in IssuesFinder to limit issues to those belonging to a set of projects. While initially this may seem like a good idea, both classes provide a very high level interface with very little control. This means that IssuesFinder may not be able to produce a better optimized database query, as a large portion of the query is controlled by the internals of ProjectsFinder.

To work around this problem, you would use the same code used by ProjectsFinder, instead of using ProjectsFinder itself directly. This allows you to compose your behavior better, giving you more control over the behavior of the code.

To illustrate, consider the following code from IssuableFinder#projects: 

return @projects = project if project?

projects =
  if current_user && params[:authorized_only].presence && !current_user_related?
    current_user.authorized_projects
  elsif group
    finder_options = { include_subgroups: params[:include_subgroups], exclude_shared: true }
    GroupProjectsFinder.new(group: group, current_user: current_user, options: finder_options).execute
  else
    ProjectsFinder.new(current_user: current_user).execute
  end

@projects = projects.with_feature_available_for_user(klass, current_user).reorder(nil)

Here we determine what projects to scope our data to, using three different approaches. When a group is specified, we use GroupProjectsFinder to retrieve all the projects of that group. On the surface this seems harmless: it is easy to use, and we only need two lines of code.

In reality, things can get hairy very quickly. For example, the query produced by GroupProjectsFinder may start out simple. Over time more and more functionality is added to this (high level) interface. Instead of only affecting the cases where this is necessary, it may also start affecting IssuableFinder in a negative way. For example, the query produced by GroupProjectsFinder may include unnecessary conditions. Since we’re using a finder here, we can’t easily opt-out of that behavior. We could add options to do so, but then we’d need as many options as we have features. Every option adds two code paths, which means that for four features we have to cover 8 different code paths.

A much more reliable (and pleasant) way of dealing with this, is to use the underlying bits that make up GroupProjectsFinder directly. This means we may need a little bit more code in IssuableFinder, but it also gives us much more control and certainty. This means we might end up with something like this: 

return @projects = project if project?

projects =
  if current_user && params[:authorized_only].presence && !current_user_related?
    current_user.authorized_projects
  elsif group
    current_user
      .owned_groups(subgroups: params[:include_subgroups])
      .projects
      .any_additional_method_calls
      .that_might_be_necessary
  else
    current_user
      .projects_visible_to_user
      .any_additional_method_calls
      .that_might_be_necessary
  end

@projects = projects.with_feature_available_for_user(klass, current_user).reorder(nil)

This is just a sketch, but it shows the general idea: we would use whatever the GroupProjectsFinder and ProjectsFinder finders use under the hoods.

End goal

The guidelines in this document are meant to foster better code reuse, by clearly defining what can be reused where, and what to do when you cannot reuse something. Clearly separating abstractions makes it harder to use the wrong one, makes it easier to debug the code, and (hopefully) results in fewer performance problems.

Abstractions

Now let’s take a look at the various abstraction levels available, and what they can (or cannot) reuse. For this we can use the following table, which defines the various abstractions and what they can (not) reuse:

Abstraction Service classes Finders Presenters Serializers Model instance method Model class methods Active Record Worker
Controller/API endpoint Yes Yes Yes Yes Yes No No No
Service class Yes Yes No No Yes No No Yes
Finder No No No No Yes Yes No No
Presenter No Yes No No Yes Yes No No
Serializer No Yes No No Yes Yes No No
Model class method No No No No Yes Yes Yes No
Model instance method No Yes No No Yes Yes Yes Yes
Worker Yes Yes No No Yes No No Yes

Controllers

Everything in app/controllers.

Controllers should not do much work on their own, instead they pass input to other classes and present the results.

API endpoints

Everything in lib/api (the REST API) and app/graphql (the GraphQL API).

API endpoints have the same abstraction level as controllers.

Service classes

Everything that resides in app/services.

Service classes represent operations that coordinates changes between models (such as entities and value objects). Changes impact the state of the application.

  1. When an object makes no changes to the state of the application, then it’s not a service. It may be a finder or a value object.
  2. When there is no operation, there is no need to execute a service. The class would probably be better designed as an entity, a value object, or a policy.

When implementing a service class, consider using the following patterns:

  1. A service class initializer should contain in its arguments:
    1. A model instance that is being acted upon. Should be first positional argument of the initializer. The argument name of the argument is left to the developer’s discretion, such as: issue, project, merge_request.
    2. When service represents an action initiated by a user or executed in the context of a user, the initializer must have the current_user: keyword argument. Services with the current_user: argument run high-level business logic and must validate user authorization to perform their operations.
    3. When service does not have a user context and it’s not directly initiated by a user (like background service or side-effects), the current_user: argument is not needed. This describes low-level domain logic or instance-wide logic.
    4. For all additional data required by a service, the explicit keyword arguments are recommended. When a service requires too long of a list of arguments, consider splitting them into:
      • params: A hash with model properties that will be assigned directly.
      • options: A hash with extra parameters (which need to be processed, and are not model properties). The options hash should be stored in an instance variable. 
      # merge_request: A model instance that is being acted upon.
# assignee: new MR assignee that will be assigned to the MR
#   after the service is executed.
def initialize(merge_request, assignee:)
  @merge_request = merge_request
  @assignee = assignee
end

      # issue: A model instance that is being acted upon.
# current_user: Current user.
# params: Model properties.
# options: Configuration for this service. Can be any of the following:
#   - notify: Whether to send a notification to the current user.
#   - cc: Email address to copy when sending a notification.
def initialize(issue:, current_user:, params: {}, options: {})
  @issue = issue
  @current_user = current_user
  @params = params
  @options = options
end
  2. The service class should implements a single public instance method #execute, which invokes service class behavior:
    • The #execute method takes no arguments. All required data is passed into initializer.
  3. If a return value is needed, the #execute method should returns its result via ServiceResponse object.

Several base classes implement the service classes convention. You may consider inheriting from:

  • BaseContainerService for services scoped by container (project or group).
  • BaseProjectService for services scoped to projects.
  • BaseGroupService for services scoped to groups.

For some domains or bounded contexts, it may make sense for service classes to use different patterns. For example, the Remote Development domain uses a layered architecture with domain logic isolated to a separate domain layer following a standard pattern, which allows for a very minimal service layer which consists of only a single reusable CommonService class. It also uses functional patterns with stateless singleton class methods. See the Remote Development service layer code example for more details. However, even though the invocation signature of services via this pattern is different, it still respects the standard Service Class contract of always returning all results via a ServiceResponse object.

Classes that are not service objects should be created elsewhere, such as in lib.

ServiceResponse

Service classes usually have an execute method, which can return a ServiceResponse. You can use ServiceResponse.success and ServiceResponse.error to return a response in execute method.

In a successful case: 

response = ServiceResponse.success(message: 'Branch was deleted')

response.success? # => true
response.error? # => false
response.status # => :success
response.message # => 'Branch was deleted'

In a failed case: 

response = ServiceResponse.error(message: 'Unsupported operation')

response.success? # => false
response.error? # => true
response.status # => :error
response.message # => 'Unsupported operation'

An additional payload can also be attached: 

response = ServiceResponse.success(payload: { issue: issue })

response.payload[:issue] # => issue

Error responses can also specify the failure reason which can be used by the caller to understand the nature of the failure. The caller, if an HTTP endpoint, could translate the reason symbol into an HTTP status code: 

response = ServiceResponse.error(
  message: 'Job is in a state that cannot be retried',
  reason: :job_not_retrieable)

if response.success?
  head :ok
elsif response.reason == :job_not_retriable
  head :unprocessable_entity
else
  head :bad_request
end

For common failures such as resource :not_found or operation :forbidden, we could leverage the Rails HTTP status symbols as long as they are sufficiently specific for the domain logic involved. For other failures use domain-specific reasons whenever possible.

For example: :job_not_retriable, :duplicate_package, :merge_request_not_mergeable.

Finders

Everything in app/finders, typically used for retrieving data from a database.

Finders cannot reuse other finders in an attempt to better control the SQL queries they produce.

Finders’ execute method should return ActiveRecord::Relation. Exceptions can be added to spec/support/finder_collection_allowlist.yml. See #298771 for more details.

Presenters

Everything in app/presenters, used for exposing complex data to a Rails view, without having to create many instance variables.

See the documentation for more information.

Serializers

Everything in app/serializers, used for presenting the response to a request, typically in JSON.

Models

Classes and modules in app/models represent domain concepts that encapsulate both data and behavior.

These classes can interact directly with a data store (like ActiveRecord models) or can be a thin wrapper (Plain Old Ruby Objects) on top of ActiveRecord models to express a richer domain concept.

Entities and Value Objects that represent domain concepts are considered domain models.

Some examples:

Model class methods

These are class methods defined by GitLab itself, including the following methods provided by Active Record:

  • find
  • find_by_id
  • delete_all
  • destroy
  • destroy_all

Any other methods such as find_by(some_column: X) are not included, and instead fall under the “Active Record” abstraction.

Model instance methods

Instance methods defined on Active Record models by GitLab itself. Methods provided by Active Record are not included, except for the following methods:

  • save
  • update
  • destroy
  • delete

Active Record

The API provided by Active Record itself, such as the where method, save, delete_all, and so on.

Worker

Everything in app/workers.

Use SomeWorker.perform_async or SomeWorker.perform_in to schedule Sidekiq jobs. Never directly invoke a worker using SomeWorker.new.perform.