gitlab-org--gitlab-foss/doc/development/api_graphql_styleguide.md

82 KiB

stage group info
none unassigned To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

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. In addition, we have a subscription to GraphQL Pro. For details see GraphQL Pro subscription.

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 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.

Limits

Several limits apply to the GraphQL API and some of these can be overridden by developers.

Max page size

By default, connections can only return at most a maximum number of records defined in app/graphql/gitlab_schema.rb per page.

Developers can specify a custom max page size when defining a connection.

Max complexity

Complexity is explained on our client-facing API page.

Fields default to adding 1 to a query's complexity score, but developers can specify a custom complexity when defining a field.

The complexity score of a query can itself be queried for.

Request timeout

Requests time out at 30 seconds.

Breaking changes

The GitLab GraphQL API is versionless which means developers must familiarize themselves with our Deprecation and Removal process.

Breaking changes are:

  • Removing or renaming a field, argument, enum value, or mutation.
  • Changing the type of a field, argument or enum value.
  • Raising the complexity of a field or complexity multipliers in a resolver.
  • Changing a field from being not nullable (null: false) to nullable (null: true), as discussed in Nullable fields.
  • Changing an argument from being optional (required: false) to being required (required: true).
  • Changing the max page size of a connection.
  • Lowering the global limits for query complexity and depth.
  • Anything else that can result in queries hitting a limit that previously was allowed.

See the deprecating schema items section for how to deprecate items.

Breaking change exemptions

Two scenarios exist where schema items are exempt from the deprecation process, and can be removed or changed at any time without notice. These are schema items that either:

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 (for example, 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 (such as 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::Types::ID, null: true
field :title, GraphQL::Types::String, null: true

# we also have a method here that we've defined, that extends `field`
markdown_field :title_html, null: true
field :description, GraphQL::Types::String, 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::Types::ID, a special string type that signifies a unique ID. title and description are regular GraphQL::Types::String types.

Note that the old scalar types GraphQL:ID, GraphQL::INT_TYPE, GraphQL::STRING_TYPE, GraphQL:BOOLEAN_TYPE, and GraphQL::FLOAT_TYPE are no longer allowed. Please use GraphQL::Types::ID, GraphQL::Types::Int, GraphQL::Types::String, GraphQL::Types::Boolean, and GraphQL::Types::Float.

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::Types::ID

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).

Page size limit

By default, the API returns at most a maximum number of records defined in app/graphql/gitlab_schema.rb per page in a connection and this is also the default number of records returned per page if no limiting arguments (first: or last:) are provided by a client.

The max_page_size argument can be used to specify a different page size limit for a connection.

WARNING: It's better to change the frontend client, or product requirements, to not need large amounts of records per page than it is to raise the max_page_size, as the default is set to ensure the GraphQL API remains performant.

For example:

field :tags,
  Types::ContainerRepositoryTagType.connection_type,
  null: true,
  description: 'Tags of the container repository',
  max_page_size: 20

Field complexity

The GitLab GraphQL API uses a complexity score to limit performing overly complex queries. Complexity is described in our client documentation on the topic.

Complexity limits are defined in app/graphql/gitlab_schema.rb.

By default, fields add 1 to a query's complexity score. This can be overridden by providing a custom complexity value for a field.

Developers should specify higher complexity for fields that cause more work to be performed by the server in order to return data. Fields that represent data that can be returned with little-to-no work, for example in most cases; id or title, can be given a complexity of 0.

calls_gitaly

Fields that have the potential to perform a Gitaly call when resolving must be marked as such by passing calls_gitaly: true to field when defining it.

For example:

field :blob, type: Types::Snippets::BlobType,
      description: 'Snippet blob',
      null: false,
      calls_gitaly: true

This increments the complexity score of the field by 1.

If a resolver calls Gitaly, it can be annotated with BaseResolver.calls_gitaly!. This passes calls_gitaly: true to any field that uses this resolver.

For example:

class BranchResolver < BaseResolver
  type ::Types::BranchType, null: true
  calls_gitaly!

  argument name: ::GraphQL::Types::String, required: true

  def resolve(name:)
    object.branch(name)
  end
end

Then when we use it, any field that uses BranchResolver has the correct value for calls_gitaly:.

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
  graphql_name 'MergeRequestPermissions'

  present_using MergeRequestPresenter

  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:

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

feature_flag property

The feature_flag property allows you to toggle the field's visibility in 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::Types::String,
      null: true,
      description: 'Some test field.',
      feature_flag: :my_feature_flag

Frontend and Backend feature flag strategies

Directives

When feature flags are used in the frontend to control the @include and @skip directives, do not use the feature_flag property on the server-side. For the accepted backend workaround, see Toggle the value of a field. It is recommended that the feature flag used in this approach be the same for frontend and backend.

Even if the frontend directives evaluate to @include:false / @skip:true, the guarded entity is sent to the backend and matched against the GraphQL schema. We would then get an exception due to a schema mismatch. See the frontend GraphQL guide for more guidance.

Different versions of a query

See the guide frontend GraphQL guide for different versions of a query, and why it is not the preferred approach

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.

Consider also marking the field as Alpha while the value of the field can be toggled. You can change or remove Alpha fields at any time without needing to deprecate them. This also signals to consumers of the public GraphQL API that the field is not meant to be used yet.

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::Types::String,
      null: true,
      alpha: { milestone: '10.0' },
      description: 'Some test field. Returns `null`' \
                   'if `my_feature_flag` feature flag is disabled.'

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

Deprecating schema items

The GitLab GraphQL API is versionless, which means we maintain backwards compatibility with older versions of the API with every change.

Rather than removing fields, arguments, enum values, or mutations, they must be deprecated instead.

The deprecated parts of the schema can then be removed in a future release in accordance with the GitLab deprecation process.

To deprecate a schema item in GraphQL:

  1. Create a deprecation issue for the item.
  2. Mark the item as deprecated in the schema.

See also:

Create a deprecation issue

Every GraphQL deprecation should have a deprecation issue created using the Deprecations issue template to track its deprecation and removal.

Apply these two labels to the deprecation issue:

  • ~GraphQL
  • ~deprecation

Mark the item as deprecated

Fields, arguments, enum values, and mutations 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::Types::String, 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, argument, or enum value being replaced, the reason must indicate the replacement. For example, the following is a reason for a replaced field:

Use `otherFieldName`

Examples:

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, argument, or enum value being deprecated is not being replaced, a descriptive deprecation reason should be given.

Deprecate Global IDs

We use the rails/globalid gem to generate and parse Global IDs, so as such they are coupled to model names. When we rename a model, its Global ID changes.

If the Global ID is used as an argument type anywhere in the schema, then the Global ID change would normally constitute a breaking change.

To continue to support clients using the old Global ID argument, we add a deprecation to Gitlab::GlobalId::Deprecations.

NOTE: If the Global ID is only exposed as a field then we do not need to deprecate it. We consider the change to the way a Global ID is expressed in a field to be backwards-compatible. We expect that clients don't parse these values: they are meant to be treated as opaque tokens, and any structure in them is incidental and not to be relied on.

Example scenario:

This example scenario is based on this merge request.

A model named PrometheusService is to be renamed Integrations::Prometheus. The old model name is used to create a Global ID type that is used as an argument for a mutation:

# Mutations::UpdatePrometheus:

argument :id, Types::GlobalIDType[::PrometheusService],
              required: true,
              description: "The ID of the integration to mutate."

Clients call the mutation by passing a Global ID string that looks like "gid://gitlab/PrometheusService/1", named as PrometheusServiceID, as the input.id argument:

mutation updatePrometheus($id: PrometheusServiceID!, $active: Boolean!) {
  prometheusIntegrationUpdate(input: { id: $id, active: $active }) {
    errors
    integration {
      active
    }
  }
}

We rename the model to Integrations::Prometheus, and then update the codebase with the new name. When we come to update the mutation, we pass the renamed model to Types::GlobalIDType[]:

# Mutations::UpdatePrometheus:

argument :id, Types::GlobalIDType[::Integrations::Prometheus],
              required: true,
              description: "The ID of the integration to mutate."

This would cause a breaking change to the mutation, as the API now rejects clients who pass an id argument as "gid://gitlab/PrometheusService/1", or that specify the argument type as PrometheusServiceID in the query signature.

To allow clients to continue to interact with the mutation unchanged, edit the DEPRECATIONS constant in Gitlab::GlobalId::Deprecations and add a new Deprecation to the array:

DEPRECATIONS = [
  Deprecation.new(old_model_name: 'PrometheusService', new_model_name: 'Integrations::Prometheus', milestone: '14.0')
].freeze

Then follow our regular deprecation process. To later remove support for the former argument style, remove the Deprecation:

DEPRECATIONS = [].freeze

During the deprecation period the API will accept either of these formats for the argument value:

  • "gid://gitlab/PrometheusService/1"
  • "gid://gitlab/Integrations::Prometheus/1"

The API will also accept these types in the query signature for the argument:

  • PrometheusServiceID
  • IntegrationsPrometheusID

NOTE: Although queries that use the old type (PrometheusServiceID in this example) will be considered valid and executable by the API, validator tools will consider them to be invalid. This is because we are deprecating using a bespoke method outside of the @deprecated directive, so validators are not aware of the support.

The documentation will mention that the old Global ID style is now deprecated.

Marking schema items as Alpha

Fields, arguments, enum values, and mutations can be marked as being in alpha.

An item marked as "alpha" is exempt from the deprecation process and can be removed at any time without notice.

This leverages GraphQL deprecations to cause the schema item to appear as deprecated, and will be described as being in "alpha" in our generated docs and its GraphQL description.

To mark a schema item as being in "alpha", use the alpha: keyword. You must provide the milestone: that introduced the alpha item.

For example:

field :token, GraphQL::Types::String, null: true,
      alpha: { milestone: '10.0' },
      description: 'Token for login.'

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::Types::String, 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::Types::Int, null: true, description: 'X-axis value of the chart datum.'
    field :y, GraphQL::Types::Int, 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::Types::ID, description: 'ID of the issue.'
field :confidential, GraphQL::Types::Boolean, description: 'Indicates the issue is confidential.'
field :closed_at, Types::TimeType, description: 'Timestamp of when the issue was closed.'

You can view descriptions of fields and arguments in:

Description style guide

Language and punctuation

Use {x} of the {y} where possible, where {x} is the item you're describing, and {y} is the resource it applies to. For example:

ID of the issue.

Do not start descriptions with The or A, for consistency and conciseness.

End all descriptions with a period (.).

Booleans

For a boolean field (GraphQL::Types::Boolean), start with a verb that describes what it does. For example:

Indicates the issue is confidential.

If necessary, provide the default. For example:

Sets the issue to confidential. Default is false.

Types::TimeType field description

For Types::TimeType GraphQL fields, include the word timestamp. This lets the reader know that the format of the property is Time, rather than just Date.

For example:

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::Types::String,
          required: false,
          description: copy_field_description(Types::MergeRequestType, :title)

Documentation references

Sometimes we want to refer to external URLs in our descriptions. To make this easier, and provide proper markup in the generated reference documentation, we provide a see property on fields. For example:

field :genus,
      type: GraphQL::Types::String,
      null: true,
      description: 'A taxonomic genus.'
      see: { 'Wikipedia page on genera' => 'https://wikipedia.org/wiki/Genus' }

This renders in our documentation as:

A taxonomic genus. See: [Wikipedia page on genera](https://wikipedia.org/wiki/Genus)

Multiple documentation references can be provided. The syntax for this property is a HashMap where the keys are textual descriptions, and the values are URLs.

Authorization

See: GraphQL Authorization

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::Types::String], required: false, as: :slug

  alias_method :blog, :object

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

While you can use the same resolver class in two different places, such as in two different fields where the same object is exposed, you should never re-use resolver objects 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 are 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 are suppressed and the client receives 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, because 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, which 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::Types::String], required: false

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

  when_single do
    argument :name, ::GraphQL::Types::String, 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

By default, fields defined in #preloads are preloaded if that field is selected in the query. Occasionally, finer control may be needed to avoid preloading too much or incorrect content.

Extending the above example, we might want to preload a different association if certain fields are requested together. This can be done by overriding #filtered_preloads:

class MyThingResolver < BaseResolver
  # ...

  def filtered_preloads
    return [:alternate_attribute] if lookahead.selects?(:field_one) && lookahead.selects?(:field_two)

    super
  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 (required - all resolvers must include a type annotation)
  • extras
  • description
  • Gitaly annotations (with calls_gitaly!)

Example:

module Resolvers
  MyResolver < BaseResolver
    type Types::MyType, null: true
    extras [:lookahead]
    description 'Retrieve a single MyType'
    calls_gitaly!
  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 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, you might want to only update one aspect of an object, and thus only need a fine-grained mutation, for example MergeRequest::SetDraft.

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

Due to changes in the 1.13 version of the graphql-ruby gem, graphql_name should be the first line of the class to ensure that type names are generated correctly. The Graphql::GraphqlNamePosition cop enforces this. See issue #27536 for further context.

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::Types::String,
         required: true,
         description: "A description of the argument."

Nullability

Arguments can be marked as required: true which means the value must be present and not null. If a required argument's value can be null, use the required: :nullable declaration.

Example:

argument :due_date,
         Types::TimeType,
         required: :nullable,
         description: 'The desired due date for the issue. Due date is removed if null.'

In the above example, the due_date argument must be given, but unlike the GraphQL spec, the value can be null. This allows 'unsetting' the due date in a single mutation rather than creating a new mutation for removing the due date.

{ due_date: null } # => OK
{ due_date: "2025-01-10" } # => OK
{  } # => invalid (not given)

Keywords

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

Example:

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

Input Types

graphql-ruby wraps up arguments into an input type.

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

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

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

argument :draft,
         GraphQL::Types::Boolean,
         required: false,
         description: <<~DESC
           Whether or not to set the merge request as a draft.
         DESC

These arguments automatically generate an input type called MergeRequestSetDraftInput 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.

See also Deprecate Global IDs.

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 in 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::SetDraft 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
    graphql_name 'Mutation'

    include Gitlab::Graphql::MountMutation

    mount_mutation Mutations::MergeRequests::SetDraft
  end
end

Generates a field called mergeRequestSetDraft that Mutations::MergeRequests::SetDraft 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 SetDraft < Base
      graphql_name 'MergeRequestSetDraft'

      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, because 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, because 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 in 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 in 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.

Subscriptions

We use subscriptions to push updates to clients. We use the Action Cable implementation to deliver the messages over websockets.

When a client subscribes to a subscription, we store their query in-memory within Puma workers. Then when the subscription is triggered, the Puma workers execute the stored GraphQL queries and push the results to the clients.

NOTE: We cannot test subscriptions using GraphiQL, because they require an Action Cable client, which GraphiQL does not support at the moment.

Building subscriptions

All fields under Types::SubscriptionType are subscriptions that clients can subscribe to. These fields require a subscription class, which is a descendant of Subscriptions::BaseSubscription and is stored under app/graphql/subscriptions.

The arguments required to subscribe and the fields that are returned are defined in the subscription class. Multiple fields can share the same subscription class if they have the same arguments and return the same fields.

This class runs during the initial subscription request and subsequent updates. You can read more about this in the GraphQL Ruby guides.

Authorization

You should implement the #authorized? method of the subscription class so that the initial subscription and subsequent updates are authorized.

When a user is not authorized, you should call the unauthorized! helper so that execution is halted and the user is unsubscribed. Returning false results in redaction of the response but we leak information that some updates are happening. This is due to a bug in the GraphQL gem.

Triggering subscriptions

Define a method under the GraphqlTriggers module to trigger a subscription. Do not call GitlabSchema.subscriptions.trigger directly in application code so that we have a single source of truth and we do not trigger a subscription with different arguments and objects.

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 in your mutation or resolver to provide the validation. The #ready? method is called before any work is done in 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 OneOf Input Objects 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

For testing mutations and resolvers, consider the unit of test a full GraphQL request, not a call to a resolver. The reasons for this are that we want to avoid lots of coupling to the framework, since this makes upgrades to dependencies much more difficult.

You should:

  • Prefer request specs (either using the full API endpoint or going through GitlabSchema.execute) to unit specs for resolvers and mutations.
  • Prefer GraphqlHelpers#execute_query and GraphqlHelpers#run_with_clean_state to GraphqlHelpers#resolve and GraphqlHelpers#resolve_field.

For example:

# Good:
gql_query = %q(some query text...)
post_graphql(gql_query, current_user: current_user)
# or:
GitlabSchema.execute(gql_query, context: { current_user: current_user })

# Deprecated: avoid
resolve(described_class, obj: project, ctx: { current_user: current_user })

Writing unit tests (deprecated)

WARNING: Avoid writing unit tests if the same thing can be tested with a full GraphQL request.

Before creating unit tests, review the following examples:

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, consider calling GitlabSchema.execute directly, or making use of smaller test schemas that only contain the types under test.

However, full request integration tests that check if data is returned verify the following additional items:

  • The mutation is actually queryable in 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.
  • The arguments coerce correctly on input, and the fields serialize correctly on output.

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 successfully, 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

  • Become familiar with the methods in the GraphqlHelpers support module. Many of these methods make writing GraphQL tests easier.

  • Use traversal helpers like GraphqlHelpers#graphql_data_at and GraphqlHelpers#graphql_dig_at to access result fields. For example:

    result = GitlabSchema.execute(query)
    
    mr_iid = graphql_dig_at(result.to_h, :data, :project, :merge_request, :iid)
    
  • Use GraphqlHelpers#a_graphql_entity_for to match against results. For example:

    post_graphql(some_query)
    
    # checks that it is a hash containing { id => global_id_of(issue) }
    expect(graphql_data_at(:project, :issues, :nodes))
      .to contain_exactly(a_graphql_entity_for(issue))
    
    # Additional fields can be passed, either as names of methods, or with values
    expect(graphql_data_at(:project, :issues, :nodes))
      .to contain_exactly(a_graphql_entity_for(issue, :iid, :title, created_at: some_time))
    
  • Use GraphqlHelpers#empty_schema to create an empty schema, rather than creating one by hand. For example:

    # good
    let(:schema) { empty_schema }
    
    # bad
    let(:query_type) { GraphQL::ObjectType.new }
    let(:schema) { GraphQL::Schema.define(query: query_type, mutation: nil)}
    
  • Use GraphqlHelpers#query_double(schema: nil) of double('query', schema: nil). For example:

    # good
    let(:query) { query_double(schema: GitlabSchema) }
    
    # bad
    let(:query) { double('Query', schema: GitlabSchema) }
    
  • 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.

  • There can be possible cyclic dependencies within our GraphQL types. See Adding field with resolver on a Type causes "Can't determine the return type " error on a different Type and Fix unresolved name due to cyclic definition.

    In particular, this can happen with connection_type. Normally we might use the following in a resolver:

    type Types::IssueType.connection_type, null: true
    

    However this might cause a cyclic definition, which can result in errors like:

    NameError: uninitialized constant Resolvers::GroupIssuesResolver
    

    To fix this, we must create a new file that encapsulates the connection type, and then reference it using double quotes. This gives a delayed resolution, and the proper connection type. For example:

    module Types
      # rubocop: disable Graphql/AuthorizeTypes
      class IssueConnectionType < CountableConnectionType
      end
    end
    
    Types::IssueConnectionType.prepend_mod_with('Types::IssueConnectionType')
    

    in types/issue_connection_type.rb defines a new Types::IssueConnectionType, and is then referenced in app/graphql/resolvers/base_issues_resolver.rb

    type "Types::IssueConnection", null: true
    

    Only use this style if you are having spec failures. This is not intended to be a new pattern that we use. This issue may disappear after we've upgraded to 2.x.

  • There can be instances where a spec fails because the class is not loaded correctly. It relates to the circular dependencies problem and Adding field with resolver on a Type causes "Can't determine the return type " error on a different Type.

    Unfortunately, the errors generated don't really indicate what the problem is. For example, remove the quotes from the Rspec.descrbe in ee/spec/graphql/resolvers/compliance_management/merge_requests/compliance_violation_resolver_spec.rb. Then run rspec ee/spec/graphql/resolvers/compliance_management/merge_requests/compliance_violation_resolver_spec.rb.

    This generates errors with the expectations. For example:

    1) Resolvers::ComplianceManagement::MergeRequests::ComplianceViolationResolver#resolve user is authorized filtering the results when given an array of project IDs finds the filtered compliance violations
    Failure/Error: expect(subject).to contain_exactly(compliance_violation)
    
       expected collection contained:  [#<MergeRequests::ComplianceViolation id: 4, violating_user_id: 26, merge_request_id: 4, reason: "approved_by_committer", severity_level: "low">]
       actual collection contained:    [#<MergeRequests::ComplianceViolation id: 4, violating_user_id: 26, merge_request_id: 4, reason: "app...er_id: 27, merge_request_id: 5, reason: "approved_by_merge_request_author", severity_level: "high">]
       the extra elements were:        [#<MergeRequests::ComplianceViolation id: 5, violating_user_id: 27, merge_request_id: 5, reason: "approved_by_merge_request_author", severity_level: "high">]
    # ./ee/spec/graphql/resolvers/compliance_management/merge_requests/compliance_violation_resolver_spec.rb:55:in `block (6 levels) in <top (required)>'
    

    However, this is not a case of the wrong result being generated, it's because of the loading order of the ComplianceViolationResolver class.

    The only way we've found to fix this is by quoting the class name in the spec. For example, changing

    RSpec.describe Resolvers::ComplianceManagement::MergeRequests::ComplianceViolationResolver do
    

    into:

    RSpec.describe 'Resolvers::ComplianceManagement::MergeRequests::ComplianceViolationResolver' do
    

    See this merge request for some discussion.

    Only use this style if you are having spec failures. This is not intended to be a new pattern that we use. This issue may disappear after we've upgraded to 2.x.

  • When testing resolvers using GraphqlHelpers#resolve, arguments for the resolver can be handled two ways.

    1. 95% of the resolver specs use arguments that are Ruby objects, as opposed to when using the GraphQL API only strings and integers are used. This works fine in most cases.

    2. If your resolver takes arguments that use a prepare proc, such as a resolver that accepts timeframe arguments (TimeFrameArguments), you must pass the arg_style: :internal_prepared parameter into the resolve method. This tells the code to convert the arguments into strings and integers and pass them through regular argument handling, ensuring that the prepare proc is called correctly. For example in iterations_resolver_spec.rb:

      def resolve_group_iterations(args = {}, obj = group, context = { current_user: current_user })
        resolve(described_class, obj: obj, args: args, ctx: context, arg_style: :internal_prepared)
      end
      

      One additional caveat is that if you are passing enums as a resolver argument, you must use the external representation of the enum, rather than the internal. For example:

      # good
      resolve_group_iterations({ search: search, in: ['CADENCE_TITLE'] })
      
      # bad
      resolve_group_iterations({ search: search, in: [:cadence_title] })
      

    The use of :internal_prepared was added as a bridge for the GraphQL gem upgrade. Testing resolvers directly will be removed eventually, and writing unit tests for resolvers/mutations is already deprecated

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 (such as 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.

Monitoring GraphQL

See the Monitoring GraphQL guide for tips on how to inspect logs of GraphQL requests and monitor the performance of your GraphQL queries.