GraphQL API style guide

This document outlines the style guide for the GitLab GraphQL API.

How GitLab implements GraphQL

We use the GraphQL Ruby gem written by Robert Mosolgo.

All GraphQL queries are directed to a single endpoint (app/controllers/graphql_controller.rb#execute), which is exposed as an API endpoint at /api/graphql.

Deep Dive

In March 2019, Nick Thomas hosted a Deep Dive (GitLab team members only: https://gitlab.com/gitlab-org/create-stage/issues/1) on the GitLab GraphQL API to share his domain specific knowledge with anyone who may work in this part of the codebase in the future. You can find the recording on YouTube, and the slides on Google Slides and in PDF. Everything covered in this deep dive was accurate as of GitLab 11.9, and while specific details may have changed since then, it should still serve as a good introduction.

GraphiQL

GraphiQL is an interactive GraphQL API explorer where you can play around with existing queries. You can access it in any GitLab environment on https://<your-gitlab-site.com>/-/graphql-explorer. For example, the one for GitLab.com.

Authentication

Authentication happens through the GraphqlController, right now this uses the same authentication as the Rails application. So the session can be shared.

It's also possible to add a private_token to the query string, or add a HTTP_PRIVATE_TOKEN header.

Global IDs

The GitLab GraphQL API uses Global IDs (i.e: "gid://gitlab/MyObject/123") and never database primary key IDs.

Global ID is a convention used for caching and fetching in client-side libraries.

See also:

We have a custom scalar type (Types::GlobalIDType) which should be used as the type of input and output arguments when the value is a GlobalID. The benefits of using this type instead of ID are:

  • it validates that the value is a GlobalID
  • it parses it into a GlobalID before passing it to user code
  • it can be parameterized on the type of the object (e.g. GlobalIDType[Project]) which offers even better validation and security.

Consider using this type for all new arguments and result types. Remember that it is perfectly possible to parameterize this type with a concern or a supertype, if you want to accept a wider range of objects (e.g. GlobalIDType[Issuable] vs GlobalIDType[Issue]).

Types

We use a code-first schema, and we declare what type everything is in Ruby.

For example, app/graphql/types/issue_type.rb:

graphql_name 'Issue'

field :iid, GraphQL::ID_TYPE, null: true
field :title, GraphQL::STRING_TYPE, null: true

# we also have a method here that we've defined, that extends `field`
markdown_field :title_html, null: true
field :description, GraphQL::STRING_TYPE, null: true
markdown_field :description_html, null: true

We give each type a name (in this case Issue).

The iid, title and description are scalar GraphQL types. iid is a GraphQL::ID_TYPE, a special string type that signifies a unique ID. title and description are regular GraphQL::STRING_TYPE types.

When exposing a model through the GraphQL API, we do so by creating a new type in app/graphql/types. You can also declare custom GraphQL data types for scalar data types (for example TimeType).

When exposing properties in a type, make sure to keep the logic inside the definition as minimal as possible. Instead, consider moving any logic into a presenter:

class Types::MergeRequestType < BaseObject
  present_using MergeRequestPresenter

  name 'MergeRequest'
end

An existing presenter could be used, but it is also possible to create a new presenter specifically for GraphQL.

The presenter is initialized using the object resolved by a field, and the context.

Nullable fields

GraphQL allows fields to be "nullable" or "non-nullable". The former means that null may be returned instead of a value of the specified type. In general, you should prefer using nullable fields to non-nullable ones, for the following reasons:

  • It's common for data to switch from required to not-required, and back again
  • Even when there is no prospect of a field becoming optional, it may not be available at query time
    • For instance, the content of a blob may need to be looked up from Gitaly
    • If the content is nullable, we can return a partial response, instead of failing the whole query
  • Changing from a non-nullable field to a nullable field is difficult with a versionless schema

Non-nullable fields should only be used when a field is required, very unlikely to become optional in the future, and very easy to calculate. An example would be id fields.

A non-nullable GraphQL schema field is an object type followed by the exclamation point (bang) !. Here's an example from the gitlab_schema.graphql file:

  id: ProjectID!

Here's an example of a non-nullable GraphQL array:


  errors: [String!]!

Further reading:

Exposing Global IDs

In keeping with the GitLab use of Global IDs, always convert database primary key IDs into Global IDs when you expose them.

All fields named id are converted automatically into the object's Global ID.

Fields that are not named id need to be manually converted. We can do this using Gitlab::GlobalID.build, or by calling #to_global_id on an object that has mixed in the GlobalID::Identification module.

Using an example from Types::Notes::DiscussionType:

field :reply_id, GraphQL::ID_TYPE

def reply_id
  ::Gitlab::GlobalId.build(object, id: object.reply_id)
end

Connection types

NOTE: For specifics on implementation, see Pagination implementation.

GraphQL uses cursor based pagination to expose collections of items. This provides the clients with a lot of flexibility while also allowing the backend to use different pagination models.

To expose a collection of resources we can use a connection type. This wraps the array with default pagination fields. For example a query for project-pipelines could look like this:

query($project_path: ID!) {
  project(fullPath: $project_path) {
    pipelines(first: 2) {
      pageInfo {
        hasNextPage
        hasPreviousPage
      }
      edges {
        cursor
        node {
          id
          status
        }
      }
    }
  }
}

This would return the first 2 pipelines of a project and related pagination information, ordered by descending ID. The returned data would look like this:

{
  "data": {
    "project": {
      "pipelines": {
        "pageInfo": {
          "hasNextPage": true,
          "hasPreviousPage": false
        },
        "edges": [
          {
            "cursor": "Nzc=",
            "node": {
              "id": "gid://gitlab/Pipeline/77",
              "status": "FAILED"
            }
          },
          {
            "cursor": "Njc=",
            "node": {
              "id": "gid://gitlab/Pipeline/67",
              "status": "FAILED"
            }
          }
        ]
      }
    }
  }
}

To get the next page, the cursor of the last known element could be passed:

query($project_path: ID!) {
  project(fullPath: $project_path) {
    pipelines(first: 2, after: "Njc=") {
      pageInfo {
        hasNextPage
        hasPreviousPage
      }
      edges {
        cursor
        node {
          id
          status
        }
      }
    }
  }
}

To ensure that we get consistent ordering, we append an ordering on the primary key, in descending order. This is usually id, so we add order(id: :desc) to the end of the relation. A primary key must be available on the underlying table.

Shortcut fields

Sometimes it can seem easy to implement a "shortcut field", having the resolver return the first of a collection if no parameters are passed. These "shortcut fields" are discouraged because they create maintenance overhead. They need to be kept in sync with their canonical field, and deprecated or modified if their canonical field changes. Use the functionality the framework provides unless there is a compelling reason to do otherwise.

For example, instead of latest_pipeline, use pipelines(last: 1).

Exposing permissions for a type

To expose permissions the current user has on a resource, you can call the expose_permissions passing in a separate type representing the permissions for the resource.

For example:

module Types
  class MergeRequestType < BaseObject
    expose_permissions Types::MergeRequestPermissionsType
  end
end

The permission type inherits from BasePermissionType which includes some helper methods, that allow exposing permissions as non-nullable booleans:

class MergeRequestPermissionsType < BasePermissionType
  present_using MergeRequestPresenter

  graphql_name 'MergeRequestPermissions'

  abilities :admin_merge_request, :update_merge_request, :create_note

  ability_field :resolve_note,
                description: 'Indicates the user can resolve discussions on the merge request.'
  permission_field :push_to_source_branch, method: :can_push_to_source_branch?
end
  • permission_field: Acts the same as graphql-ruby's field method but setting a default description and type and making them non-nullable. These options can still be overridden by adding them as arguments.
  • ability_field: Expose an ability defined in our policies. This behaves the same way as permission_field and the same arguments can be overridden.
  • abilities: Allows exposing several abilities defined in our policies at once. The fields for these must all be non-nullable booleans with a default description.

Feature flags

Developers can add feature flags to GraphQL fields in the following ways:

  • Add the feature_flag property to a field. This allows the field to be hidden from the GraphQL schema when the flag is disabled.
  • Toggle the return value when resolving the field.

You can refer to these guidelines to decide which approach to use:

  • If your field is experimental, and its name or type is subject to change, use the feature_flag property.
  • If your field is stable and its definition doesn't change, even after the flag is removed, toggle the return value of the field instead. Note that all fields should be nullable anyway.

feature_flag property

The feature_flag property allows you to toggle the field's visibility within the GraphQL schema. This removes the field from the schema when the flag is disabled.

A description is appended to the field indicating that it is behind a feature flag.

WARNING: If a client queries for the field when the feature flag is disabled, the query fails. Consider this when toggling the visibility of the feature on or off on production.

The feature_flag property does not allow the use of feature gates based on actors. This means that the feature flag cannot be toggled only for particular projects, groups, or users, but instead can only be toggled globally for everyone.

Example:

field :test_field, type: GraphQL::STRING_TYPE,
      null: true,
      description: 'Some test field.',
      feature_flag: :my_feature_flag

Toggle the value of a field

This method of using feature flags for fields is to toggle the return value of the field. This can be done in the resolver, in the type, or even in a model method, depending on your preference and situation.

When applying a feature flag to toggle the value of a field, the description of the field must:

  • State that the value of the field can be toggled by a feature flag.
  • Name the feature flag.
  • State what the field returns when the feature flag is disabled (or enabled, if more appropriate).

Example:

field :foo, GraphQL::STRING_TYPE,
      null: true,
      description: 'Some test field. Will always return `null`' \
                   'if `my_feature_flag` feature flag is disabled.'

def foo
  object.foo if Feature.enabled?(:my_feature_flag, object)
end

Deprecating fields and enum values

The GitLab GraphQL API is versionless, which means we maintain backwards compatibility with older versions of the API with every change. Rather than removing a field or enum value, we need to deprecate it instead. The deprecated parts of the schema can then be removed in a future release in accordance with the GitLab deprecation process.

Fields and enum values are deprecated using the deprecated property. The value of the property is a Hash of:

  • reason - Reason for the deprecation.
  • milestone - Milestone that the field was deprecated.

Example:

field :token, GraphQL::STRING_TYPE, null: true,
      deprecated: { reason: 'Login via token has been removed', milestone: '10.0' },
      description: 'Token for login.'

The original description of the things being deprecated should be maintained, and should not be updated to mention the deprecation. Instead, the reason is appended to the description.

Deprecation reason style guide

Where the reason for deprecation is due to the field or enum value being replaced, the reason must be:

Use `otherFieldName`

Example:

field :designs, ::Types::DesignManagement::DesignCollectionType, null: true,
      deprecated: { reason: 'Use `designCollection`', milestone: '10.0' },
      description: 'The designs associated with this issue.',
module Types
  class TodoStateEnum < BaseEnum
    value 'pending', deprecated: { reason: 'Use PENDING', milestone: '10.0' }
    value 'done', deprecated: { reason: 'Use DONE', milestone: '10.0' }
    value 'PENDING', value: 'pending'
    value 'DONE', value: 'done'
  end
end

If the field is not being replaced by another field, a descriptive deprecation reason should be given.

See also Aliasing and deprecating mutations.

Enums

GitLab GraphQL enums are defined in app/graphql/types. When defining new enums, the following rules apply:

  • Values must be uppercase.
  • Class names must end with the string Enum.
  • The graphql_name must not contain the string Enum.

For example:

module Types
  class TrafficLightStateEnum < BaseEnum
    graphql_name 'TrafficLightState'
    description 'State of a traffic light'

    value 'RED', description: 'Drivers must stop.'
    value 'YELLOW', description: 'Drivers must stop when it is safe to.'
    value 'GREEN', description: 'Drivers can start or keep driving.'
  end
end

If the enum is used for a class property in Ruby that is not an uppercase string, you can provide a value: option that adapts the uppercase value.

In the following example:

  • GraphQL inputs of OPENED are converted to 'opened'.
  • Ruby values of 'opened' are converted to "OPENED" in GraphQL responses.
module Types
  class EpicStateEnum < BaseEnum
    graphql_name 'EpicState'
    description 'State of a GitLab epic'

    value 'OPENED', value: 'opened', description: 'An open Epic.'
    value 'CLOSED', value: 'closed', description: 'A closed Epic.'
  end
end

Enum values can be deprecated using the deprecated keyword.

Defining GraphQL enums dynamically from Rails enums

If your GraphQL enum is backed by a Rails enum, then consider using the Rails enum to dynamically define the GraphQL enum values. Doing so binds the GraphQL enum values to the Rails enum definition, so if values are ever added to the Rails enum then the GraphQL enum automatically reflects the change.

Example:

module Types
  class IssuableSeverityEnum < BaseEnum
    graphql_name 'IssuableSeverity'
    description 'Incident severity'

    ::IssuableSeverity.severities.keys.each do |severity|
      value severity.upcase, value: severity, description: "#{severity.titleize} severity."
    end
  end
end

JSON

When data to be returned by GraphQL is stored as JSON, we should continue to use GraphQL types whenever possible. Avoid using the GraphQL::Types::JSON type unless the JSON data returned is truly unstructured.

If the structure of the JSON data varies, but is one of a set of known possible structures, use a union. An example of the use of a union for this purpose is !30129.

Field names can be mapped to hash data keys using the hash_key: keyword if needed.

For example, given the following simple JSON data:

{
  "title": "My chart",
  "data": [
    { "x": 0, "y": 1 },
    { "x": 1, "y": 1 },
    { "x": 2, "y": 2 }
  ]
}

We can use GraphQL types like this:

module Types
  class ChartType < BaseObject
    field :title, GraphQL::STRING_TYPE, null: true, description: 'Title of the chart.'
    field :data, [Types::ChartDatumType], null: true, description: 'Data of the chart.'
  end
end

module Types
  class ChartDatumType < BaseObject
    field :x, GraphQL::INT_TYPE, null: true, description: 'X-axis value of the chart datum.'
    field :y, GraphQL::INT_TYPE, null: true, description: 'Y-axis value of the chart datum.'
  end
end

Descriptions

All fields and arguments must have descriptions.

A description of a field or argument is given using the description: keyword. For example:

field :id, GraphQL::ID_TYPE, description: 'ID of the resource.'

Descriptions of fields and arguments are viewable to users through:

Description style guide

To ensure consistency, the following should be followed whenever adding or updating descriptions:

  • Mention the name of the resource in the description. Example: 'Labels of the issue' (issue being the resource).
  • Use "{x} of the {y}" where possible. Example: 'Title of the issue'. Do not start descriptions with The.
  • Descriptions of GraphQL::BOOLEAN_TYPE fields should answer the question: "What does this field do?". Example: 'Indicates project has a Git repository'.
  • Always include the word "timestamp" when describing an argument or field of type Types::TimeType. This lets the reader know that the format of the property is Time, rather than just Date.
  • Must end with a period (.).

Example:

field :id, GraphQL::ID_TYPE, description: 'ID of the issue.'
field :confidential, GraphQL::BOOLEAN_TYPE, description: 'Indicates the issue is confidential.'
field :closed_at, Types::TimeType, description: 'Timestamp of when the issue was closed.'

copy_field_description helper

Sometimes we want to ensure that two descriptions are always identical. For example, to keep a type field description the same as a mutation argument when they both represent the same property.

Instead of supplying a description, we can use the copy_field_description helper, passing it the type, and field name to copy the description of.

Example:

argument :title, GraphQL::STRING_TYPE,
          required: false,
          description: copy_field_description(Types::MergeRequestType, :title)

Authorization

Authorizations can be applied to both types and fields using the same abilities as in the Rails app.

If the:

  • Currently authenticated user fails the authorization, the authorized resource is returned as null.
  • Resource is part of a collection, the collection is filtered to exclude the objects that the user's authorization checks failed against.

Also see authorizing resources in a mutation.

NOTE: Try to load only what the currently authenticated user is allowed to view with our existing finders first, without relying on authorization to filter the records. This minimizes database queries and unnecessary authorization checks of the loaded records.

Type authorization

Authorize a type by passing an ability to the authorize method. All fields with the same type is authorized by checking that the currently authenticated user has the required ability.

For example, the following authorization ensures that the currently authenticated user can only see projects that they have the read_project ability for (so long as the project is returned in a field that uses Types::ProjectType):

module Types
  class ProjectType < BaseObject
    authorize :read_project
  end
end

You can also authorize against multiple abilities, in which case all of the ability checks must pass.

For example, the following authorization ensures that the currently authenticated user must have read_project and another_ability abilities to see a project:

module Types
  class ProjectType < BaseObject
    authorize [:read_project, :another_ability]
  end
end

Field authorization

Fields can be authorized with the authorize option.

For example, the following authorization ensures that the currently authenticated user must have the owner_access ability to see the project:

module Types
  class MyType < BaseObject
    field :project, Types::ProjectType, null: true, resolver: Resolvers::ProjectResolver, authorize: :owner_access
  end
end

Fields can also be authorized against multiple abilities, in which case all of ability checks must pass. This requires explicitly passing a block to field:

module Types
  class MyType < BaseObject
    field :project, Types::ProjectType, null: true, resolver: Resolvers::ProjectResolver do
      authorize [:owner_access, :another_ability]
    end
  end
end

If the field's type already has a particular authorization then there is no need to add that same authorization to the field.

Type and Field authorizations together

Authorizations are cumulative, so where authorizations are defined on a field, and also on the field's type, then the currently authenticated user would need to pass all ability checks.

In the following simplified example the currently authenticated user would need both first_permission and second_permission abilities in order to see the author of the issue.

class UserType
  authorize :first_permission
end
class IssueType
  field :author, UserType, authorize: :second_permission
end

Resolvers

We define how the application serves the response using resolvers stored in the app/graphql/resolvers directory. The resolver provides the actual implementation logic for retrieving the objects in question.

To find objects to display in a field, we can add resolvers to app/graphql/resolvers.

Arguments can be defined within the resolver in the same way as in a mutation. See the Mutation arguments section.

To limit the amount of queries performed, we can use BatchLoader.

Writing resolvers

Our code should aim to be thin declarative wrappers around finders and services. You can repeat lists of arguments, or extract them to concerns. Composition is preferred over inheritance in most cases. Treat resolvers like controllers: resolvers should be a DSL that compose other application abstractions.

For example:

class PostResolver < BaseResolver
  type Post.connection_type, null: true
  authorize :read_blog
  description 'Blog posts, optionally filtered by name'

  argument :name, [::GraphQL::STRING_TYPE], required: false, as: :slug

  alias_method :blog, :object

  def resolve(**args)
    PostFinder.new(blog, current_user, args).execute
  end
end

You should never re-use resolvers directly. Resolvers have a complex life-cycle, with authorization, readiness and resolution orchestrated by the framework, and at each stage lazy values can be returned to take advantage of batching opportunities. Never instantiate a resolver or a mutation in application code.

Instead, the units of code reuse are much the same as in the rest of the application:

  • Finders in queries to look up data.
  • Services in mutations to apply operations.
  • Loaders (batch-aware finders) specific to queries.

Note that there is never any reason to use batching in a mutation. Mutations are executed in series, so there are no batching opportunities. All values are evaluated eagerly as soon as they are requested, so batching is unnecessary overhead. If you are writing:

  • A Mutation, feel free to lookup objects directly.
  • A Resolver or methods on a BaseObject, then you want to allow for batching.

Error handling

Resolvers may raise errors, which will be converted to top-level errors as appropriate. All anticipated errors should be caught and transformed to an appropriate GraphQL error (see Gitlab::Graphql::Errors). Any uncaught errors will be suppressed and the client will receive the message Internal service error.

The one special case is permission errors. In the REST API we return 404 Not Found for any resources that the user does not have permission to access. The equivalent behavior in GraphQL is for us to return null for all absent or unauthorized resources. Query resolvers should not raise errors for unauthorized resources.

The rationale for this is that clients must not be able to distinguish between the absence of a record and the presence of one they do not have access to. To do so is a security vulnerability, since it leaks information we want to keep hidden.

In most cases you don't need to worry about this - this is handled correctly by the resolver field authorization we declare with the authorize DSL calls. If you need to do something more custom however, remember, if you encounter an object the current_user does not have access to when resolving a field, then the entire field should resolve to null.

Deriving resolvers (BaseResolver.single and BaseResolver.last)

For some simple use cases, we can derive resolvers from others. The main use case for this is one resolver to find all items, and another to find one specific one. For this, we supply convenience methods:

  • BaseResolver.single, which constructs a new resolver that selects the first item.
  • BaseResolver.last, with constructs a resolver that selects the last item.

The correct singular type is inferred from the collection type, so we don't have to define the type here.

Before you make use of these methods, consider if it would be simpler to either:

  • Write another resolver that defines its own arguments.
  • Write a concern that abstracts out the query.

Using BaseResolver.single too freely is an anti-pattern. It can lead to non-sensical fields, such as a Project.mergeRequest field that just returns the first MR if no arguments are given. Whenever we derive a single resolver from a collection resolver, it must have more restrictive arguments.

To make this possible, use the when_single block to customize the single resolver. Every when_single block must:

  • Define (or re-define) at least one argument.
  • Make optional filters required.

For example, we can do this by redefining an existing optional argument, changing its type and making it required:

class JobsResolver < BaseResolver
  type JobType.connection_type, null: true
  authorize :read_pipeline

  argument :name, [::GraphQL::STRING_TYPE], required: false

  when_single do
    argument :name, ::GraphQL::STRING_TYPE, required: true
  end

  def resolve(**args)
    JobsFinder.new(pipeline, current_user, args.compact).execute
  end

Here we have a simple resolver for getting pipeline jobs. The name argument is optional when getting a list, but required when getting a single job.

If there are multiple arguments, and neither can be made required, we can use the block to add a ready condition:

class JobsResolver < BaseResolver
  alias_method :pipeline, :object

  type JobType.connection_type, null: true
  authorize :read_pipeline

  argument :name, [::GraphQL::STRING_TYPE], required: false
  argument :id, [::Types::GlobalIDType[::Job]],
           required: false,
           prepare: ->(ids, ctx) { ids.map(&:model_id) }

  when_single do
    argument :name, ::GraphQL::STRING_TYPE, required: false
    argument :id, ::Types::GlobalIDType[::Job],
             required: false
             prepare: ->(id, ctx) { id.model_id }

    def ready?(**args)
      raise ::Gitlab::Graphql::Errors::ArgumentError, 'Only one argument may be provided' unless args.size == 1
    end
  end

  def resolve(**args)
    JobsFinder.new(pipeline, current_user, args.compact).execute
  end

Then we can use these resolver on fields:

# In PipelineType

field :jobs, resolver: JobsResolver, description: 'All jobs.'
field :job, resolver: JobsResolver.single, description: 'A single job.'

Correct use of Resolver#ready?

Resolvers have two public API methods as part of the framework: #ready?(**args) and #resolve(**args). We can use #ready? to perform set-up, validation or early-return without invoking #resolve.

Good reasons to use #ready? include:

  • validating mutually exclusive arguments (see validating arguments)
  • Returning Relation.none if we know before-hand that no results are possible
  • Performing setup such as initializing instance variables (although consider lazily initialized methods for this)

Implementations of Resolver#ready?(**args) should return (Boolean, early_return_data) as follows:

def ready?(**args)
  [false, 'have this instead']
end

For this reason, whenever you call a resolver (mainly in tests - as framework abstractions Resolvers should not be considered re-usable, finders are to be preferred), remember to call the ready? method and check the boolean flag before calling resolve! An example can be seen in our GraphQLHelpers.

Look-Ahead

The full query is known in advance during execution, which means we can make use of lookahead to optimize our queries, and batch load associations we know we need. Consider adding lookahead support in your resolvers to avoid N+1 performance issues.

To enable support for common lookahead use-cases (pre-loading associations when child fields are requested), you can include LooksAhead. For example:

# Assuming a model `MyThing` with attributes `[child_attribute, other_attribute, nested]`,
# where nested has an attribute named `included_attribute`.
class MyThingResolver < BaseResolver
  include LooksAhead

  # Rather than defining `resolve(**args)`, we implement: `resolve_with_lookahead(**args)`
  def resolve_with_lookahead(**args)
    apply_lookahead(MyThingFinder.new(current_user).execute)
  end

  # We list things that should always be preloaded:
  # For example, if child_attribute is always needed (during authorization
  # perhaps), then we can include it here.
  def unconditional_includes
    [:child_attribute]
  end

  # We list things that should be included if a certain field is selected:
  def preloads
    {
        field_one: [:other_attribute],
        field_two: [{ nested: [:included_attribute] }]
    }
  end
end

The final thing that is needed is that every field that uses this resolver needs to advertise the need for lookahead:

  # in ParentType
  field :my_things, MyThingType.connection_type, null: true,
        extras: [:lookahead], # Necessary
        resolver: MyThingResolver,
        description: 'My things.'

For an example of real world use, please see ResolvesMergeRequests.

Negated arguments

Negated filters can filter some resources (for example, find all issues that have the bug label, but don't have the bug2 label assigned). The not argument is the preferred syntax to pass negated arguments:

issues(labelName: "bug", not: {labelName: "bug2"}) {
  nodes {
    id
    title
  }
}

To avoid duplicated argument definitions, you can place these arguments in a reusable module (or class, if the arguments are nested). Alternatively, you can consider to add a helper resolver method.

Metadata

When using resolvers, they can and should serve as the SSoT for field metadata. All field options (apart from the field name) can be declared on the resolver. These include:

  • type (this is particularly important, and is planned to be mandatory)
  • extras
  • description

Example:

module Resolvers
  MyResolver < BaseResolver
    type Types::MyType, null: true
    extras [:lookahead]
    description 'Retrieve a single MyType'
  end
end

Pass a parent object into a child Presenter

Sometimes you need to access the resolved query parent in a child context to compute fields. Usually the parent is only available in the Resolver class as parent.

To find the parent object in your Presenter class:

  1. Add the parent object to the GraphQL context from within your resolver's resolve method:

      def resolve(**args)
        context[:parent_object] = parent
      end
  2. Declare that your resolver or fields require the parent field context. For example:

      # in ChildType
      field :computed_field, SomeType, null: true,
            method: :my_computing_method,
            extras: [:parent], # Necessary
            description: 'My field description.'
    
      field :resolver_field, resolver: SomeTypeResolver
    
      # In SomeTypeResolver
    
      extras [:parent]
      type SomeType, null: true
      description 'My field description.'
  3. Declare your field's method in your Presenter class and have it accept the parent keyword argument. This argument contains the parent GraphQL context, so you have to access the parent object with parent[:parent_object] or whatever key you used in your Resolver:

      # in ChildPresenter
      def my_computing_method(parent:)
        # do something with `parent[:parent_object]` here
      end
    
      # In SomeTypeResolver
    
      def resolve(parent:)
        # ...
      end

For an example of real-world use, check this MR that added scopedPath and scopedUrl to IterationPresenter

Mutations

Mutations are used to change any stored values, or to trigger actions. In the same way a GET-request should not modify data, we cannot modify data in a regular GraphQL-query. We can however in a mutation.

Building Mutations

Mutations are stored in app/graphql/mutations, ideally grouped per resources they are mutating, similar to our services. They should inherit Mutations::BaseMutation. The fields defined on the mutation are returned as the result of the mutation.

Update mutation granularity

The service-oriented architecture in GitLab means that most mutations call a Create, Delete, or Update service, for example UpdateMergeRequestService. For Update mutations, a you might want to only update one aspect of an object, and thus only need a fine-grained mutation, for example MergeRequest::SetWip.

It's acceptable to have both fine-grained mutations and coarse-grained mutations, but be aware that too many fine-grained mutations can lead to organizational challenges in maintainability, code comprehensibility, and testing. Each mutation requires a new class, which can lead to technical debt. It also means the schema becomes very big, and we want users to easily navigate our schema. As each new mutation also needs tests (including slower request integration tests), adding mutations slows down the test suite.

To minimize changes:

  • Use existing mutations, such as MergeRequest::Update, when available.
  • Expose existing services as a coarse-grained mutation.

When a fine-grained mutation might be more appropriate:

  • Modifying a property that requires specific permissions or other specialized logic.
  • Exposing a state-machine-like transition (locking issues, merging MRs, closing epics, etc).
  • Accepting nested properties (where we accept properties for a child object).
  • The semantics of the mutation can be expressed clearly and concisely.

See issue #233063 for further context.

Naming conventions

Each mutation must define a graphql_name, which is the name of the mutation in the GraphQL schema.

Example:

class UserUpdateMutation < BaseMutation
  graphql_name 'UserUpdate'
end

Our GraphQL mutation names are historically inconsistent, but new mutation names should follow the convention '{Resource}{Action}' or '{Resource}{Action}{Attribute}'.

Mutations that create new resources should use the verb Create.

Example:

  • CommitCreate

Mutations that update data should use:

  • The verb Update.
  • A domain-specific verb like Set, Add, or Toggle if more appropriate.

Examples:

  • EpicTreeReorder
  • IssueSetWeight
  • IssueUpdate
  • TodoMarkDone

Mutations that remove data should use:

  • The verb Delete rather than Destroy.
  • A domain-specific verb like Remove if more appropriate.

Examples:

  • AwardEmojiRemove
  • NoteDelete

If you need advice for mutation naming, canvass the Slack #graphql channel for feedback.

Arguments

Arguments for a mutation are defined using argument.

Example:

argument :my_arg, GraphQL::STRING_TYPE,
         required: true,
         description: "A description of the argument."

Each GraphQL argument defined is passed to the #resolve method of a mutation as keyword arguments.

Example:

def resolve(my_arg:)
  # Perform mutation ...
end

graphql-ruby wraps up arguments into an input type.

For example, the mergeRequestSetWip mutation defines these arguments (some through inheritance):

argument :project_path, GraphQL::ID_TYPE,
         required: true,
         description: "The project the merge request to mutate is in."

argument :iid, GraphQL::STRING_TYPE,
         required: true,
         description: "The IID of the merge request to mutate."

argument :wip,
         GraphQL::BOOLEAN_TYPE,
         required: false,
         description: <<~DESC
                      Whether or not to set the merge request as a WIP.
                      If not passed, the value will be toggled.
                      DESC

These arguments automatically generate an input type called MergeRequestSetWipInput with the 3 arguments we specified and the clientMutationId.

Object identifier arguments

In keeping with the GitLab use of Global IDs, mutation arguments should use Global IDs to identify an object and never database primary key IDs.

Where an object has an iid, prefer to use the full_path or group_path of its parent in combination with its iid as arguments to identify an object rather than its id.

Fields

In the most common situations, a mutation would return 2 fields:

  • The resource being modified
  • A list of errors explaining why the action could not be performed. If the mutation succeeded, this list would be empty.

By inheriting any new mutations from Mutations::BaseMutation the errors field is automatically added. A clientMutationId field is also added, this can be used by the client to identify the result of a single mutation when multiple are performed within a single request.

The resolve method

Similar to writing resolvers, the resolve method of a mutation should aim to be a thin declarative wrapper around a service.

The resolve method receives the mutation's arguments as keyword arguments. From here, we can call the service that modifies the resource.

The resolve method should then return a hash with the same field names as defined on the mutation including an errors array. For example, the Mutations::MergeRequests::SetWip defines a merge_request field:

field :merge_request,
      Types::MergeRequestType,
      null: true,
      description: "The merge request after mutation."

This means that the hash returned from resolve in this mutation should look like this:

{
  # The merge request modified, this will be wrapped in the type
  # defined on the field
  merge_request: merge_request,
  # An array of strings if the mutation failed after authorization.
  # The `errors_on_object` helper collects `errors.full_messages`
  errors: errors_on_object(merge_request)
}

Mounting the mutation

To make the mutation available it must be defined on the mutation type that is stored in graphql/types/mutation_types. The mount_mutation helper method defines a field based on the GraphQL-name of the mutation:

module Types
  class MutationType < BaseObject
    include Gitlab::Graphql::MountMutation

    graphql_name "Mutation"

    mount_mutation Mutations::MergeRequests::SetWip
  end
end

Generates a field called mergeRequestSetWip that Mutations::MergeRequests::SetWip to be resolved.

Authorizing resources

To authorize resources inside a mutation, we first provide the required abilities on the mutation like this:

module Mutations
  module MergeRequests
    class SetWip < Base
      graphql_name 'MergeRequestSetWip'

      authorize :update_merge_request
    end
  end
end

We can then call authorize! in the resolve method, passing in the resource we want to validate the abilities for.

Alternatively, we can add a find_object method that loads the object on the mutation. This would allow you to use the authorized_find! helper method.

When a user is not allowed to perform the action, or an object is not found, we should raise a Gitlab::Graphql::Errors::ResourceNotAvailable error which is correctly rendered to the clients.

Errors in mutations

We encourage following the practice of errors as data for mutations, which distinguishes errors by who they are relevant to, defined by who can deal with them.

Key points:

  • All mutation responses have an errors field. This should be populated on failure, and may be populated on success.
  • Consider who needs to see the error: the user or the developer.
  • Clients should always request the errors field when performing mutations.
  • Errors may be reported to users either at $root.errors (top-level error) or at $root.data.mutationName.errors (mutation errors). The location depends on what kind of error this is, and what information it holds.
  • Mutation fields must have null: true

Consider an example mutation doTheThing that returns a response with two fields: errors: [String], and thing: ThingType. The specific nature of the thing itself is irrelevant to these examples, as we are considering the errors.

There are three states a mutation response can be in:

Success

In the happy path, errors may be returned, along with the anticipated payload, but if everything was successful, then errors should be an empty array, since there are no problems we need to inform the user of.

{
  data: {
    doTheThing: {
      errors: [] // if successful, this array will generally be empty.
      thing: { .. }
    }
  }
}

Failure (relevant to the user)

An error that affects the user occurred. We refer to these as mutation errors. In this case there is typically no thing to return:

{
  data: {
    doTheThing: {
      errors: ["you cannot touch the thing"],
      thing: null
    }
  }
}

Examples of this include:

  • Model validation errors: the user may need to change the inputs.
  • Permission errors: the user needs to know they cannot do this, they may need to request permission or sign in.
  • Problems with application state that prevent the user's action, for example: merge conflicts, the resource was locked, and so on.

Ideally, we should prevent the user from getting this far, but if they do, they need to be told what is wrong, so they understand the reason for the failure and what they can do to achieve their intent, even if that is as simple as retrying the request.

It is possible to return recoverable errors alongside mutation data. For example, if a user uploads 10 files and 3 of them fail and the rest succeed, the errors for the failures can be made available to the user, alongside the information about the successes.

Failure (irrelevant to the user)

One or more non-recoverable errors can be returned at the top level. These are things over which the user has little to no control, and should mainly be system or programming problems, that a developer needs to know about. In this case there is no data:

{
  errors: [
    {"message": "argument error: expected an integer, got null"},
  ]
}

This is the result of raising an error during the mutation. In our implementation, the messages of argument errors and validation errors are returned to the client, and all other StandardError instances are caught, logged and presented to the client with the message set to "Internal server error". See GraphqlController for details.

These represent programming errors, such as:

  • A GraphQL syntax error, where an Int was passed instead of a String, or a required argument was not present.
  • Errors in our schema, such as being unable to provide a value for a non-nullable field.
  • System errors: for example, a Git storage exception, or database unavailability.

The user should not be able to cause such errors in regular usage. This category of errors should be treated as internal, and not shown to the user in specific detail.

We need to inform the user when the mutation fails, but we do not need to tell them why, since they cannot have caused it, and nothing they can do fixes it, although we may offer to retry the mutation.

Categorizing errors

When we write mutations, we need to be conscious about which of these two categories an error state falls into (and communicate about this with frontend developers to verify our assumptions). This means distinguishing the needs of the user from the needs of the client.

Never catch an error unless the user needs to know about it.

If the user does need to know about it, communicate with frontend developers to make sure the error information we are passing back is useful.

See also the frontend GraphQL guide.

Aliasing and deprecating mutations

The #mount_aliased_mutation helper allows us to alias a mutation as another name within MutationType.

For example, to alias a mutation called FooMutation as BarMutation:

mount_aliased_mutation 'BarMutation', Mutations::FooMutation

This allows us to rename a mutation and continue to support the old name, when coupled with the deprecated argument.

Example:

mount_aliased_mutation 'UpdateFoo',
                        Mutations::Foo::Update,
                        deprecated: { reason: 'Use fooUpdate', milestone: '13.2' }

Deprecated mutations should be added to Types::DeprecatedMutations and tested for within the unit test of Types::MutationType. The merge request !34798 can be referred to as an example of this, including the method of testing deprecated aliased mutations.

Deprecating EE mutations

EE mutations should follow the same process. For an example of the merge request process, read merge request !42588.

Pagination implementation

To learn more, visit GraphQL pagination.

Validating arguments

For validations of single arguments, use the prepare option as normal.

Sometimes a mutation or resolver may accept a number of optional arguments, but we still want to validate that at least one of the optional arguments is provided. In this situation, consider using the #ready? method within your mutation or resolver to provide the validation. The #ready? method is called before any work is done within the #resolve method.

Example:

def ready?(**args)
  if args.values_at(:body, :position).compact.blank?
    raise Gitlab::Graphql::Errors::ArgumentError,
          'body or position arguments are required'
  end

  # Always remember to call `#super`
  super
end

In the future this may be able to be done using InputUnions if this RFC is merged.

GitLab custom scalars

Types::TimeType

Types::TimeType must be used as the type for all fields and arguments that deal with Ruby Time and DateTime objects.

The type is a custom scalar that:

  • Converts Ruby's Time and DateTime objects into standardized ISO-8601 formatted strings, when used as the type for our GraphQL fields.
  • Converts ISO-8601 formatted time strings into Ruby Time objects, when used as the type for our GraphQL arguments.

This allows our GraphQL API to have a standardized way that it presents time and handles time inputs.

Example:

field :created_at, Types::TimeType, null: true, description: 'Timestamp of when the issue was created.'

Testing

Writing unit tests

Before creating unit tests, review the following examples:

It's faster to test as much of the logic from your GraphQL queries and mutations with unit tests, which are stored in spec/graphql.

Use unit tests to verify that:

  • Types have the expected fields.
  • Resolvers and mutations apply authorizations and return expected data.
  • Edge cases are handled correctly.

Writing integration tests

Integration tests check the full stack for a GraphQL query or mutation and are stored in spec/requests/api/graphql.

For speed, you should test most logic in unit tests instead of integration tests. However, integration tests that check if data is returned verify the following additional items:

  • The mutation is actually queryable within the schema (was mounted in MutationType).
  • The data returned by a resolver or mutation correctly matches the return types of the fields and resolves without errors.

Integration tests can also verify the following items, because they invoke the full stack:

  • An argument or scalar's prepare applies correctly.
  • Logic in a resolver or mutation's #ready? method applies correctly.
  • An argument's default_value applies correctly.
  • Objects resolve performantly and there are no N+1 issues.

When adding a query, you can use the a working graphql query shared example to test if the query renders valid results.

You can construct a query including all available fields using the GraphqlHelpers#all_graphql_fields_for helper. This makes it easy to add a test rendering all possible fields for a query.

If you're adding a field to a query that supports pagination and sorting, visit Testing for details.

To test GraphQL mutation requests, GraphqlHelpers provides two helpers: graphql_mutation which takes the name of the mutation, and a hash with the input for the mutation. This returns a struct with a mutation query, and prepared variables.

You can then pass this struct to the post_graphql_mutation helper, that posts the request with the correct parameters, like a GraphQL client would do.

To access the response of a mutation, you can use the graphql_mutation_response helper.

Using these helpers, you can build specs like this:

let(:mutation) do
  graphql_mutation(
    :merge_request_set_wip,
    project_path: 'gitlab-org/gitlab-foss',
    iid: '1',
    wip: true
  )
end

it 'returns a successful response' do
   post_graphql_mutation(mutation, current_user: user)

   expect(response).to have_gitlab_http_status(:success)
   expect(graphql_mutation_response(:merge_request_set_wip)['errors']).to be_empty
end

Testing tips and tricks

  • Avoid false positives:

    Authenticating a user with the current_user: argument for post_graphql generates more queries on the first request than on subsequent requests on that same user. If you are testing for N+1 queries using QueryRecorder, use a different user for each request.

    The below example shows how a test for avoiding N+1 queries should look:

    RSpec.describe 'Query.project(fullPath).pipelines' do
      include GraphqlHelpers
    
      let(:project) { create(:project) }
    
      let(:query) do
        %(
          {
            project(fullPath: "#{project.full_path}") {
              pipelines {
                nodes {
                  id
                }
              }
            }
          }
        )
      end
    
      it 'avoids N+1 queries' do
        first_user = create(:user)
        second_user = create(:user)
        create(:ci_pipeline, project: project)
    
        control_count = ActiveRecord::QueryRecorder.new do
          post_graphql(query, current_user: first_user)
        end
    
        create(:ci_pipeline, project: project)
    
        expect do
          post_graphql(query, current_user: second_user)  # use a different user to avoid a false positive from authentication queries
        end.not_to exceed_query_limit(control_count)
      end
    end
  • Mimic the folder structure of app/graphql/types:

    For example, tests for fields on Types::Ci::PipelineType in app/graphql/types/ci/pipeline_type.rb should be stored in spec/requests/api/graphql/ci/pipeline_spec.rb regardless of the query being used to fetch the pipeline data.

Notes about Query flow and GraphQL infrastructure

The GitLab GraphQL infrastructure can be found in lib/gitlab/graphql.

Instrumentation is functionality that wraps around a query being executed. It is implemented as a module that uses the Instrumentation class.

Example: Present

module Gitlab
  module Graphql
    module Present
      #... some code above...

      def self.use(schema_definition)
        schema_definition.instrument(:field, ::Gitlab::Graphql::Present::Instrumentation.new)
      end
    end
  end
end

A Query Analyzer contains a series of callbacks to validate queries before they are executed. Each field can pass through the analyzer, and the final value is also available to you.

Multiplex queries enable multiple queries to be sent in a single request. This reduces the number of requests sent to the server. (there are custom Multiplex Query Analyzers and Multiplex Instrumentation provided by GraphQL Ruby).

Query limits

Queries and mutations are limited by depth, complexity, and recursion to protect server resources from overly ambitious or malicious queries. These values can be set as defaults and overridden in specific queries as needed. The complexity values can be set per object as well, and the final query complexity is evaluated based on how many objects are being returned. This is useful for objects that are expensive (e.g. requiring Gitaly calls).

For example, a conditional complexity method in a resolver:

def self.resolver_complexity(args, child_complexity:)
  complexity = super
  complexity += 2 if args[:labelName]

  complexity
end

More about complexity: GraphQL Ruby documentation.

Documentation and schema

Our schema is located at app/graphql/gitlab_schema.rb. See the schema reference for details.

This generated GraphQL documentation needs to be updated when the schema changes. For information on generating GraphQL documentation and schema files, see updating the schema documentation.

To help our readers, you should also add a new page to our GraphQL API documentation. For guidance, see the GraphQL API page.

Include a changelog entry

All client-facing changes must include a changelog entry.

Laziness

One important technique unique to GraphQL for managing performance is using lazy values. Lazy values represent the promise of a result, allowing their action to be run later, which enables batching of queries in different parts of the query tree. The main example of lazy values in our code is the GraphQL BatchLoader.

To manage lazy values directly, read Gitlab::Graphql::Lazy, and in particular Gitlab::Graphql::Laziness. This contains #force and #delay, which help implement the basic operations of creation and elimination of laziness, where needed.

For dealing with lazy values without forcing them, use Gitlab::Graphql::Lazy.with_value.