GraphQL

Getting Started

Helpful Resources

General resources:

• 📚 Official Introduction to GraphQL
• 📚 Official Introduction to Apollo

GraphQL at GitLab:

• GitLab Unfiltered GraphQL playlist
• GraphQL at GitLab: Deep Dive (video) by Nick Thomas
  • An overview of the history of GraphQL at GitLab (not frontend-specific)
• GitLab Feature Walkthrough with GraphQL and Vue Apollo (video) by Natalia Tepluhina
  • A real-life example of implementing a frontend feature in GitLab using GraphQL
• History of client-side GraphQL at GitLab (video) Illya Klymov and Natalia Tepluhina
• From Vuex to Apollo (video) by Natalia Tepluhina
  • An overview of when Apollo might be a better choice than Vuex, and how one could go about the transition
• 🛠 Vuex -> Apollo Migration: a proof-of-concept project
  • A collection of examples that show the possible approaches for state management with Vue+GraphQL+(Vuex or Apollo) apps

Libraries

We use Apollo (specifically Apollo Client) and Vue Apollo when using GraphQL for frontend development.

If you are using GraphQL in a Vue application, the Usage in Vue section can help you learn how to integrate Vue Apollo.

For other use cases, check out the Usage outside of Vue section.

We use Immer for immutable cache updates; see Immutability and cache updates for more information.

Tooling

• Apollo Client Devtools

Apollo GraphQL VS Code extension

If you use VS Code, the Apollo GraphQL extension supports autocompletion in .graphql files. To set up the GraphQL extension, follow these steps:

1. Generate the schema: bundle exec rake gitlab:graphql:schema:dump

2. Add an apollo.config.js file to the root of your gitlab local directory.

3. Populate the file with the following content:

   module.exports = {
     client: {
       includes: ['./app/assets/javascripts/**/*.graphql', './ee/app/assets/javascripts/**/*.graphql'],
       service: {
         name: 'GitLab',
         localSchemaFile: './tmp/tests/graphql/gitlab_schema.graphql',
       },
     },
   };

4. Restart VS Code.

Exploring the GraphQL API

Our GraphQL API can be explored via GraphiQL at your instance's /-/graphql-explorer or at GitLab.com. Consult the GitLab GraphQL API Reference documentation where needed.

To check all existing queries and mutations, on the right side of GraphiQL, select Documentation explorer. To check the execution of the queries and mutations you've written, in the upper-left corner, select Execute query.

Apollo Client

To save duplicated clients getting created in different apps, we have a default client that should be used. This sets up the Apollo client with the correct URL and also sets the CSRF headers.

Default client accepts two parameters: resolvers and config.

• resolvers parameter is created to accept an object of resolvers for local state management queries and mutations
• config parameter takes an object of configuration settings:
  • cacheConfig field accepts an optional object of settings to customize Apollo cache
  • baseUrl allows us to pass a URL for GraphQL endpoint different from our main endpoint (for example, ${gon.relative_url_root}/api/graphql)
  • fetchPolicy determines how you want your component to interact with the Apollo cache. Defaults to "cache-first".

Multiple client queries for the same object

If you are making multiple queries to the same Apollo client object you might encounter the following error: Cache data may be lost when replacing the someProperty field of a Query object. To address this problem, either ensure all objects of SomeEntityhave an id or a custom merge function. We are already checking id presence for every GraphQL type that has an id, so this shouldn't be the case (unless you see this warning when running unit tests; in this case please ensure your mocked responses contain an id whenever it's requested).

When SomeEntity type doesn't have an id property in the GraphQL schema, to fix this warning we need to define a custom merge function.

We have some client-wide types with merge: true defined in the default client as typePolicies (this means that Apollo will merge existing and incoming responses in the case of subsequent queries). Consider adding SomeEntity there or defining a custom merge function for it.

GraphQL Queries

To save query compilation at runtime, webpack can directly import .graphql files. This allows webpack to pre-process the query at compile time instead of the client doing compilation of queries.

To distinguish queries from mutations and fragments, the following naming convention is recommended:

• all_users.query.graphql for queries;
• add_user.mutation.graphql for mutations;
• basic_user.fragment.graphql for fragments.

If you are using queries for the CustomersDot GraphQL endpoint, end the filename with .customer.query.graphql, .customer.mutation.graphql, or .customer.fragment.graphql.

Fragments

Fragments are a way to make your complex GraphQL queries more readable and re-usable. Here is an example of GraphQL fragment:

fragment DesignListItem on Design {
  id
  image
  event
  filename
  notesCount
}

Fragments can be stored in separate files, imported and used in queries, mutations, or other fragments.

#import "./design_list.fragment.graphql"
#import "./diff_refs.fragment.graphql"

fragment DesignItem on Design {
  ...DesignListItem
  fullPath
  diffRefs {
    ...DesignDiffRefs
  }
}

More about fragments: GraphQL documentation

Global IDs

The GitLab GraphQL API expresses id fields as Global IDs rather than the PostgreSQL primary key id. Global ID is a convention used for caching and fetching in client-side libraries.

To convert a Global ID to the primary key id, you can use getIdFromGraphQLId:

import { getIdFromGraphQLId } from '~/graphql_shared/utils';

const primaryKeyId = getIdFromGraphQLId(data.id);

It is required to query global id for every GraphQL type that has an id in the schema:

query allReleases(...) {
  project(...) {
    id // Project has an ID in GraphQL schema so should fetch it
    releases(...) {
      nodes {
        // Release has no ID property in GraphQL schema
        name
        tagName
        tagPath
        assets {
          count
          links {
            nodes {
              id // Link has an ID in GraphQL schema so should fetch it
              name
            }
          }
        }
      }
      pageInfo {
        // PageInfo no ID property in GraphQL schema
        startCursor
        hasPreviousPage
        hasNextPage
        endCursor
      }
    }
  }
}

Immutability and cache updates

From Apollo version 3.0.0 all the cache updates need to be immutable. It needs to be replaced entirely with a new and updated object.

To facilitate the process of updating the cache and returning the new object we use the library Immer. Follow these conventions:

• The updated cache is named data.
• The original cache data is named sourceData.

A typical update process looks like this:

...
const sourceData = client.readQuery({ query });

const data = produce(sourceData, draftState => {
  draftState.commits.push(newCommit);
});

client.writeQuery({
  query,
  data,
});
...

As shown in the code example by using produce, we can perform any kind of direct manipulation of the draftState. Besides, immer guarantees that a new state which includes the changes to draftState is generated.

Usage in Vue

To use Vue Apollo, import the Vue Apollo plugin as well as the default client. This should be created at the same point the Vue application is mounted.

import Vue from 'vue';
import VueApollo from 'vue-apollo';
import createDefaultClient from '~/lib/graphql';
Vue.use(VueApollo);

const apolloProvider = new VueApollo({
  defaultClient: createDefaultClient(),
});

new Vue({
  ...,
  apolloProvider,
  ...
});

Read more about Vue Apollo in the Vue Apollo documentation.

Local state with Apollo

It is possible to manage an application state with Apollo when creating your default client.

Using client-side resolvers

The default state can be set by writing to the cache after setting up the default client. In the example below, we are using query with @client Apollo directive to write the initial data to Apollo cache and then get this state in the Vue component:

// user.query.graphql

query User {
  user @client {
    name
    surname
    age
  }
}

// index.js

import Vue from 'vue';
import VueApollo from 'vue-apollo';
import createDefaultClient from '~/lib/graphql';
import userQuery from '~/user/user.query.graphql'
Vue.use(VueApollo);

const defaultClient = createDefaultClient();

defaultClient.cache.writeQuery({
  query: userQuery,
  data: {
    user: {
      name: 'John',
      surname: 'Doe',
      age: 30
    },
  },
});

const apolloProvider = new VueApollo({
  defaultClient,
});

// App.vue
import userQuery from '~/user/user.query.graphql'

export default {
  apollo: {
    user: {
      query: userQuery
    }
  }
}

Instead of using writeQuery, we can create a type policy that will return user on every attempt of reading the userQuery from the cache:

const defaultClient = createDefaultClient({}, {
  cacheConfig: {
    typePolicies: {
      Query: {
        fields: {
          user: {
            read(data) {
              return data || {
                user: {
                  name: 'John',
                  surname: 'Doe',
                  age: 30
                },
              }
            }
          }
        }
      }
    }
  }
});

Along with creating local data, we can also extend existing GraphQL types with @client fields. This is extremely helpful when we need to mock an API response for fields not yet added to our GraphQL API.

Mocking API response with local Apollo cache

Using local Apollo Cache is helpful when we have a reason to mock some GraphQL API responses, queries, or mutations locally (such as when they're still not added to our actual API).

For example, we have a fragment on DesignVersion used in our queries:

fragment VersionListItem on DesignVersion {
  id
  sha
}

We also must fetch the version author and the created at property to display in the versions dropdown list. But, these changes are still not implemented in our API. We can change the existing fragment to get a mocked response for these new fields:

fragment VersionListItem on DesignVersion {
  id
  sha
  author @client {
    avatarUrl
    name
  }
  createdAt @client
}

Now Apollo tries to find a resolver for every field marked with @client directive. Let's create a resolver for DesignVersion type (why DesignVersion? because our fragment was created on this type).

// resolvers.js

const resolvers = {
  DesignVersion: {
    author: () => ({
      avatarUrl:
        'https://www.gravatar.com/avatar/e64c7d89f26bd1972efa854d13d7dd61?s=80&d=identicon',
      name: 'Administrator',
      __typename: 'User',
    }),
    createdAt: () => '2019-11-13T16:08:11Z',
  },
};

export default resolvers;

We need to pass a resolvers object to our existing Apollo Client:

// graphql.js

import createDefaultClient from '~/lib/graphql';
import resolvers from './graphql/resolvers';

const defaultClient = createDefaultClient(resolvers);

For each attempt to fetch a version, our client fetches id and sha from the remote API endpoint. It then assigns our hardcoded values to the author and createdAt version properties. With this data, frontend developers are able to work on their UI without being blocked by backend. When the response is added to the API, our custom local resolver can be removed. The only change to the query/fragment is to remove the @client directive.

Read more about local state management with Apollo in the Vue Apollo documentation.

Using with Vuex

We do not recommend creating new applications with Vuex and Apollo Client combined

Working on GraphQL-based features when frontend and backend are not in sync

Any feature that requires GraphQL queries/mutations to be created or updated should be carefully planned. Frontend and backend counterparts should agree on a schema that satisfies both client-side and server-side requirements. This enables both departments to start implementing their parts without blocking each other.

Ideally, the backend implementation should be done prior to the frontend so that the client can immediately start querying the API with minimal back and forth between departments. However, we recognize that priorities don't always align. For the sake of iteration and delivering work we're committed to, it might be necessary for the frontend to be implemented ahead of the backend.

Implementing frontend queries and mutations ahead of the backend

In such case, the frontend defines GraphQL schemas or fields that do not correspond to any backend resolver yet. This is fine as long as the implementation is properly feature-flagged so it does not translate to public-facing errors in the product. However, we do validate client-side queries/mutations against the backend GraphQL schema with the graphql-verify CI job. You must confirm your changes pass the validation if they are to be merged before the backend actually supports them. Below are a few suggestions to go about this.

Using the @client directive

The preferred approach is to use the @client directive on any new query, mutation, or field that isn't yet supported by the backend. Any entity with the directive is skipped by the graphql-verify validation job.

Additionally Apollo attempts to resolve them client-side, which can be used in conjunction with Mocking API response with local Apollo cache. This provides a convenient way of testing your feature with fake data defined client-side. When opening a merge request for your changes, it can be a good idea to provide local resolvers as a patch that reviewers can apply in their GDK to easily smoke-test your work.

Make sure to track the removal of the directive in a follow-up issue, or as part of the backend implementation plan.

Adding an exception to the list of known failures

GraphQL queries/mutations validation can be completely turned off for specific files by adding their paths to the config/known_invalid_graphql_queries.yml file, much like you would disable ESLint for some files via an .eslintignore file. Bear in mind that any file listed in here is not validated at all. So if you're only adding fields to an existing query, use the @client directive approach so that the rest of the query is still validated.

Again, make sure that those overrides are as short-lived as possible by tracking their removal in the appropriate issue.

Feature-flagged queries

In cases where the backend is complete and the frontend is being implemented behind a feature flag, a couple options are available to leverage the feature flag in the GraphQL queries.

The @include directive

The @include (or its opposite, @skip) can be used to control whether an entity should be included in the query. If the @include directive evaluates to false, the entity's resolver is not hit and the entity is excluded from the response. For example:

query getAuthorData($authorNameEnabled: Boolean = false) {
  username
  name @include(if: $authorNameEnabled)
}

Then in the Vue (or JavaScript) call to the query we can pass in our feature flag. This feature flag needs to be already set up correctly. See the feature flag documentation for the correct way to do this.

export default {
  apollo: {
    user: {
      query: QUERY_IMPORT,
      variables() {
        return {
          authorNameEnabled: gon?.features?.authorNameEnabled,
        };
      },
    }
  },
};

Note that, even if the directive evaluates to false, the guarded entity is sent to the backend and matched against the GraphQL schema. So this approach requires that the feature-flagged entity exists in the schema, even if the feature flag is disabled. When the feature flag is turned off, it is recommended that the resolver returns null at the very least using the same feature flag as the frontend. See the API GraphQL guide.

Different versions of a query

There's another approach that involves duplicating the standard query, and it should be avoided. The copy includes the new entities while the original remains unchanged. It is up to the production code to trigger the right query based on the feature flag's status. For example:

export default {
  apollo: {
    user: {
      query() {
        return this.glFeatures.authorNameEnabled ? NEW_QUERY : ORIGINAL_QUERY,
      }
    }
  },
};

Avoiding multiple query versions

The multiple version approach is not recommended as it results in bigger merge requests and requires maintaining two similar queries for as long as the feature flag exists. Multiple versions can be used in cases where the new GraphQL entities are not yet part of the schema, or if they are feature-flagged at the schema level (new_entity: :feature_flag).

Manually triggering queries

Queries on a component's apollo property are made automatically when the component is created. Some components instead want the network request made on-demand, for example a dropdown list with lazy-loaded items.

There are two ways to do this:

1. Use the skip property

export default {
  apollo: {
    user: {
      query: QUERY_IMPORT,
      skip() {
        // only make the query when dropdown is open
        return !this.isOpen;
      },
    }
  },
};

1. Using addSmartQuery

You can manually create the Smart Query in your method.

handleClick() {
  this.$apollo.addSmartQuery('user', {
    // this takes the same values as you'd have in the `apollo` section
    query: QUERY_IMPORT,
  }),
};

Working with pagination

The GitLab GraphQL API uses Relay-style cursor pagination for connection types. This means a "cursor" is used to keep track of where in the data set the next items should be fetched from. GraphQL Ruby Connection Concepts is a good overview and introduction to connections.

Every connection type (for example, DesignConnection and DiscussionConnection) has a field pageInfo that contains an information required for pagination:

pageInfo {
  endCursor
  hasNextPage
  hasPreviousPage
  startCursor
}

Here:

• startCursor and endCursor display the cursor of the first and last items respectively.
• hasPreviousPage and hasNextPage allow us to check if there are more pages available before or after the current page.

When we fetch data with a connection type, we can pass cursor as after or before parameter, indicating a starting or ending point of our pagination. They should be followed with first or last parameter respectively to indicate how many items we want to fetch after or before a given endpoint.

For example, here we're fetching 10 designs after a cursor (let us call this projectQuery):

#import "~/graphql_shared/fragments/page_info.fragment.graphql"

query {
  project(fullPath: "root/my-project") {
    id
    issue(iid: "42") {
      designCollection {
        designs(atVersion: null, after: "Ihwffmde0i", first: 10) {
          edges {
            node {
              id
            }
          }
          pageInfo {
            ...PageInfo
          }
        }
      }
    }
  }
}

Note that we are using the page_info.fragment.graphql to populate the pageInfo information.

Using fetchMore method in components

This approach makes sense to use with user-handled pagination. For example, when the scrolling to fetch more data or explicitly clicking a Next Page button. When we need to fetch all the data initially, it is recommended to use a (non-smart) query, instead.

When making an initial fetch, we usually want to start a pagination from the beginning. In this case, we can either:

• Skip passing a cursor.
• Pass null explicitly to after.

After data is fetched, we can use the update-hook as an opportunity to customize the data that is set in the Vue component property. This allows us to get a hold of the pageInfo object among other data.

In the result-hook, we can inspect the pageInfo object to see if we need to fetch the next page. Note that we also keep a requestCount to ensure that the application does not keep requesting the next page, indefinitely:

data() {
  return {
    pageInfo: null,
    requestCount: 0,
  }
},
apollo: {
  designs: {
    query: projectQuery,
    variables() {
      return {
        // ... The rest of the design variables
        first: 10,
      };
    },
    update(data) {
      const { id = null, issue = {} } = data.project || {};
      const { edges = [], pageInfo } = issue.designCollection?.designs || {};

      return {
        id,
        edges,
        pageInfo,
      };
    },
    result() {
      const { pageInfo } = this.designs;

      // Increment the request count with each new result
      this.requestCount += 1;
      // Only fetch next page if we have more requests and there is a next page to fetch
      if (this.requestCount < MAX_REQUEST_COUNT && pageInfo?.hasNextPage) {
        this.fetchNextPage(pageInfo.endCursor);
      }
    },
  },
},

When we want to move to the next page, we use an Apollo fetchMore method, passing a new cursor (and, optionally, new variables) there.

fetchNextPage(endCursor) {
  this.$apollo.queries.designs.fetchMore({
    variables: {
      // ... The rest of the design variables
      first: 10,
      after: endCursor,
    },
  });
}

Defining field merge policy

We would also need to define a field policy to specify how do we want to merge the existing results with the incoming results. For example, if we have Previous/Next buttons, it makes sense to replace the existing result with the incoming one:

const apolloProvider = new VueApollo({
  defaultClient: createDefaultClient(
    {},
    {
      cacheConfig: {
        typePolicies: {
          DesignCollection: {
            fields: {
              designs: {
                merge(existing, incoming) {
                  if (!incoming) return existing;
                  if (!existing) return incoming;

                  // We want to save only incoming nodes and replace existing ones
                  return incoming
                }
              }
            }
          }
        }
      },
    },
  ),
});

When we have an infinite scroll, it would make sense to add the incoming designs nodes to existing ones instead of replacing. In this case, merge function would be slightly different:

const apolloProvider = new VueApollo({
  defaultClient: createDefaultClient(
    {},
    {
      cacheConfig: {
        typePolicies: {
          DesignCollection: {
            fields: {
              designs: {
                merge(existing, incoming) {
                  if (!incoming) return existing;
                  if (!existing) return incoming;

                  const { nodes, ...rest } = incoming;
                  // We only need to merge the nodes array.
                  // The rest of the fields (pagination) should always be overwritten by incoming
                  let result = rest;
                  result.nodes = [...existing.nodes, ...nodes];
                  return result;
                }
              }
            }
          }
        }
      },
    },
  ),
});

apollo-client provides a few field policies to be used with paginated queries. Here's another way to achieve infinite scroll pagination with the concatPagination policy:

import { concatPagination } from '@apollo/client/utilities';
import Vue from 'vue';
import VueApollo from 'vue-apollo';
import createDefaultClient from '~/lib/graphql';

Vue.use(VueApollo);

export default new VueApollo({
  defaultClient: createDefaultClient(
    {},
    {
      cacheConfig: {
        typePolicies: {
          Project: {
            fields: {
              dastSiteProfiles: {
                keyArgs: ['fullPath'], // You might need to set the keyArgs option to enforce the cache's integrity
              },
            },
          },
          DastSiteProfileConnection: {
            fields: {
              nodes: concatPagination(),
            },
          },
        },
      },
    },
  ),
});

This is similar to the DesignCollection example above as new page results are appended to the previous ones.

For some cases, it's hard to define the correct keyArgs for the field because all the fields are updated. In this case, we can set keyArgs to false. This instructs Apollo Client to not perform any automatic merge, and fully rely on the logic we put into the merge function.

For example, we have a query like this:

query searchGroupsWhereUserCanTransfer {
  currentUser {
    id
    groups(after: 'somecursor') {
      nodes {
        id
        fullName
      }
      pageInfo {
        ...PageInfo
      }
    }
  }
}

Here, the groups field doesn't have a good candidate for keyArgs: we don't want to account for after argument because it will change on requesting subsequent pages. Setting keyArgs to false makes the update work as intended:

typePolicies: {
  UserCore: {
    fields: {
      groups: {
        keyArgs: false,
      },
    },
  },
  GroupConnection: {
    fields: {
      nodes: concatPagination(),
    },
  },
}

Using a recursive query in components

When it is necessary to fetch all paginated data initially an Apollo query can do the trick for us. If we need to fetch the next page based on user interactions, it is recommend to use a smartQuery along with the fetchMore-hook.

When the query resolves we can update the component data and inspect the pageInfo object. This allows us to see if we need to fetch the next page, calling the method recursively.

Note that we also keep a requestCount to ensure that the application does not keep requesting the next page, indefinitely.

data() {
  return {
    requestCount: 0,
    isLoading: false,
    designs: {
      edges: [],
      pageInfo: null,
    },
  }
},
created() {
  this.fetchDesigns();
},
methods: {
  handleError(error) {
    this.isLoading = false;
    // Do something with `error`
  },
  fetchDesigns(endCursor) {
    this.isLoading = true;

    return this.$apollo
      .query({
        query: projectQuery,
        variables() {
          return {
            // ... The rest of the design variables
            first: 10,
            endCursor,
          };
        },
      })
      .then(({ data }) => {
        const { id = null, issue = {} } = data.project || {};
        const { edges = [], pageInfo } = issue.designCollection?.designs || {};

        // Update data
        this.designs = {
          id,
          edges: [...this.designs.edges, ...edges];
          pageInfo: pageInfo;
        };

        // Increment the request count with each new result
        this.requestCount += 1;
        // Only fetch next page if we have more requests and there is a next page to fetch
        if (this.requestCount < MAX_REQUEST_COUNT && pageInfo?.hasNextPage) {
          this.fetchDesigns(pageInfo.endCursor);
        } else {
          this.isLoading = false;
        }
      })
      .catch(this.handleError);
  },
},

Pagination and optimistic updates

When Apollo caches paginated data client-side, it includes pageInfo variables in the cache key. If you wanted to optimistically update that data, you'd have to provide pageInfo variables when interacting with the cache via .readQuery() or .writeQuery(). This can be tedious and counter-intuitive.

To make it easier to deal with cached paginated queries, Apollo provides the @connection directive. The directive accepts a key parameter that is used as a static key when caching the data. You'd then be able to retrieve the data without providing any pagination-specific variables.

Here's an example of a query using the @connection directive:

#import "~/graphql_shared/fragments/page_info.fragment.graphql"

query DastSiteProfiles($fullPath: ID!, $after: String, $before: String, $first: Int, $last: Int) {
  project(fullPath: $fullPath) {
    siteProfiles: dastSiteProfiles(after: $after, before: $before, first: $first, last: $last)
      @connection(key: "dastSiteProfiles") {
      pageInfo {
        ...PageInfo
      }
      edges {
        cursor
        node {
          id
          # ...
        }
      }
    }
  }
}

In this example, Apollo stores the data with the stable dastSiteProfiles cache key.

To retrieve that data from the cache, you'd then only need to provide the $fullPath variable, omitting pagination-specific variables like after or before:

const data = store.readQuery({
  query: dastSiteProfilesQuery,
  variables: {
    fullPath: 'namespace/project',
  },
});

Read more about the @connection directive in Apollo's documentation.

Managing performance

The Apollo client batches queries by default. Given 3 deferred queries, Apollo groups them into one request, sends the single request to the server, and responds after all 3 queries have completed.

If you need to have queries sent as individual requests, additional context can be provided to tell Apollo to do this.

export default {
  apollo: {
    user: {
      query: QUERY_IMPORT,
      context: {
        isSingleRequest: true,
      }
    }
  },
};

Polling and Performance

While the Apollo client has support for simple polling, for performance reasons, our ETag-based caching is preferred to hitting the database each time.

After the ETag resource is set up to be cached from backend, there are a few changes to make on the frontend.

First, get your ETag resource from the backend, which should be in the form of a URL path. In the example of the pipelines graph, this is called the graphql_resource_etag, which is used to create new headers to add to the Apollo context:

/* pipelines/components/graph/utils.js */

/* eslint-disable @gitlab/require-i18n-strings */
const getQueryHeaders = (etagResource) => {
  return {
    fetchOptions: {
      method: 'GET',
    },
    headers: {
      /* This will depend on your feature */
      'X-GITLAB-GRAPHQL-FEATURE-CORRELATION': 'verify/ci/pipeline-graph',
      'X-GITLAB-GRAPHQL-RESOURCE-ETAG': etagResource,
      'X-REQUESTED-WITH': 'XMLHttpRequest',
    },
  };
};
/* eslint-enable @gitlab/require-i18n-strings */

/* component.vue */

apollo: {
  pipeline: {
    context() {
      return getQueryHeaders(this.graphqlResourceEtag);
    },
    query: getPipelineDetails,
    pollInterval: 10000,
    ..
  },
},

Here, the apollo query is watching for changes in graphqlResourceEtag. If your ETag resource dynamically changes, you should make sure the resource you are sending in the query headers is also updated. To do this, you can store and update the ETag resource dynamically in the local cache.

You can see an example of this in the pipeline status of the pipeline editor. The pipeline editor watches for changes in the latest pipeline. When the user creates a new commit, we update the pipeline query to poll for changes in the new pipeline.

# pipeline_etag.query.graphql

query getPipelineEtag {
  pipelineEtag @client
}

/* pipeline_editor/components/header/pipeline_status.vue */

import getPipelineEtag from '~/ci/pipeline_editor/graphql/queries/client/pipeline_etag.query.graphql';

apollo: {
  pipelineEtag: {
    query: getPipelineEtag,
  },
  pipeline: {
    context() {
      return getQueryHeaders(this.pipelineEtag);
    },
    query: getPipelineQuery,
    pollInterval: POLL_INTERVAL,
  },
}

/* pipeline_editor/components/commit/commit_section.vue */

await this.$apollo.mutate({
  mutation: commitCIFile,
  update(store, { data }) {
    const pipelineEtag = data?.commitCreate?.commit?.commitPipelinePath;

    if (pipelineEtag) {
      store.writeQuery({ query: getPipelineEtag, data: { pipelineEtag } });
    }
  },
});

ETags depend on the request being a GET instead of GraphQL's usual POST. Our default link library does not support GET requests, so we must let our default Apollo client know to use a different library. Keep in mind, this means your app cannot batch queries.

/* componentMountIndex.js */

const apolloProvider = new VueApollo({
  defaultClient: createDefaultClient(
    {},
    {
      useGet: true,
    },
  ),
});

Finally, we can add a visibility check so that the component pauses polling when the browser tab is not active. This should lessen the request load on the page.

/* component.vue */

import { toggleQueryPollingByVisibility } from '~/pipelines/components/graph/utils';

export default {
  mounted() {
    toggleQueryPollingByVisibility(this.$apollo.queries.pipeline, POLL_INTERVAL);
  },
};

You can use this MR as a reference on how to fully implement ETag caching on the frontend.

Once subscriptions are mature, this process can be replaced by using them and we can remove the separate link library and return to batching queries.

How to test ETag caching

You can test that your implementation works by checking requests on the network tab. If there are no changes in your ETag resource, all polled requests should:

• Be GET requests instead of POST requests.
• Have an HTTP status of 304 instead of 200.

Make sure that caching is not disabled in your developer tools when testing.

If you are using Chrome and keep seeing 200 HTTP status codes, it might be this bug: Developer tools show 200 instead of 304. In this case, inspect the response headers' source to confirm that the request was actually cached and did return with a 304 status code.

Subscriptions

We use subscriptions to receive real-time updates from GraphQL API via websockets. Currently, the number of existing subscriptions is limited, you can check a list of available ones in GraphqiQL explorer

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

Refer to the Real-time widgets developer guide for a comprehensive introduction to subscriptions.

Best Practices

When to use (and not use) update hook in mutations

Apollo Client's .mutate() method exposes an update hook that is invoked twice during the mutation lifecycle:

• Once at the beginning. That is, before the mutation has completed.
• Once after the mutation has completed.

You should use this hook only if you're adding or removing an item from the store (that is, ApolloCache). If you're updating an existing item, it is usually represented by a global id.

In that case, presence of this id in your mutation query definition makes the store update automatically. Here's an example of a typical mutation query with id present in it:

mutation issueSetWeight($input: IssueSetWeightInput!) {
  issuableSetWeight: issueSetWeight(input: $input) {
    issuable: issue {
      id
      weight
    }
    errors
  }
}

Testing

Generating the GraphQL schema

Some of our tests load the schema JSON files. To generate these files, run:

bundle exec rake gitlab:graphql:schema:dump

You should run this task after pulling from upstream, or when rebasing your branch. This is run automatically as part of gdk update.

NOTE: If you use the RubyMine IDE, and have marked the tmp directory as "Excluded", you should "Mark Directory As -> Not Excluded" for gitlab/tmp/tests/graphql. This will allow the JS GraphQL plugin to automatically find and index the schema.

Mocking Apollo Client

To test the components with Apollo operations, we need to mock an Apollo Client in our unit tests. We use mock-apollo-client library to mock Apollo client and createMockApollo helper we created on top of it.

We need to inject VueApollo into the Vue instance by calling Vue.use(VueApollo). This will install VueApollo globally for all the tests in the file. It is recommended to call Vue.use(VueApollo) just after the imports.

import VueApollo from 'vue-apollo';
import Vue from 'vue';

Vue.use(VueApollo);

describe('Some component with Apollo mock', () => {
  let wrapper;

  function createComponent(options = {}) {
    wrapper = shallowMount(...);
  }
})

After this, we need to create a mocked Apollo provider:

import createMockApollo from 'helpers/mock_apollo_helper';

describe('Some component with Apollo mock', () => {
  let wrapper;
  let mockApollo;

  function createComponent(options = {}) {
    mockApollo = createMockApollo(...)

    wrapper = shallowMount(SomeComponent, {
      apolloProvider: mockApollo
    });
  }

  afterEach(() => {
    // we need to ensure we don't have provider persisted between tests
    mockApollo = null
  })
})

Now, we need to define an array of handlers for every query or mutation. Handlers should be mock functions that return either a correct query response, or an error:

import getDesignListQuery from '~/design_management/graphql/queries/get_design_list.query.graphql';
import permissionsQuery from '~/design_management/graphql/queries/design_permissions.query.graphql';
import moveDesignMutation from '~/design_management/graphql/mutations/move_design.mutation.graphql';

describe('Some component with Apollo mock', () => {
  let wrapper;
  let mockApollo;

  function createComponent(options = {
    designListHandler: jest.fn().mockResolvedValue(designListQueryResponse)
  }) {
    mockApollo = createMockApollo([
       [getDesignListQuery, options.designListHandler],
       [permissionsQuery, jest.fn().mockResolvedValue(permissionsQueryResponse)],
       [moveDesignMutation, jest.fn().mockResolvedValue(moveDesignMutationResponse)],
    ])

    wrapper = shallowMount(SomeComponent, {
      apolloProvider: mockApollo
    });
  }
})

When mocking resolved values, ensure the structure of the response is the same as the actual API response. For example, root property should be data:

const designListQueryResponse = {
  data: {
    project: {
      id: '1',
      issue: {
        id: 'issue-1',
        designCollection: {
          copyState: 'READY',
          designs: {
            nodes: [
              {
                id: '3',
                event: 'NONE',
                filename: 'fox_3.jpg',
                notesCount: 1,
                image: 'image-3',
                imageV432x230: 'image-3',
                currentUserTodos: {
                  nodes: [],
                },
              },
            ],
          },
          versions: {
            nodes: [],
          },
        },
      },
    },
  },
};

When testing queries, keep in mind they are promises, so they need to be resolved to render a result. Without resolving, we can check the loading state of the query:

it('renders a loading state', () => {
  const wrapper = createComponent();

  expect(wrapper.findComponent(LoadingSpinner).exists()).toBe(true)
});

it('renders designs list', async () => {
  const wrapper = createComponent();

  await waitForPromises()

  expect(findDesigns()).toHaveLength(3);
});

If we need to test a query error, we need to mock a rejected value as request handler:

it('renders error if query fails', async () => {
  const wrapper = createComponent({
    designListHandler: jest.fn.mockRejectedValue('Houston, we have a problem!')
  });

  await waitForPromises()

  expect(wrapper.find('.test-error').exists()).toBe(true)
})

Mutations could be tested the same way:

  const moveDesignHandlerSuccess = jest.fn().mockResolvedValue(moveDesignMutationResponse)

  function createComponent(options = {
    designListHandler: jest.fn().mockResolvedValue(designListQueryResponse),
    moveDesignHandler: moveDesignHandlerSuccess
  }) {
    mockApollo = createMockApollo([
       [getDesignListQuery, options.designListHandler],
       [permissionsQuery, jest.fn().mockResolvedValue(permissionsQueryResponse)],
       [moveDesignMutation, moveDesignHandler],
    ])

    wrapper = shallowMount(SomeComponent, {
      apolloProvider: mockApollo
    });
  }

it('calls a mutation with correct parameters and reorders designs', async () => {
  const wrapper = createComponent();

  wrapper.find(VueDraggable).vm.$emit('change', {
    moved: {
      newIndex: 0,
      element: designToMove,
    },
  });

  expect(moveDesignHandlerSuccess).toHaveBeenCalled();

  await waitForPromises();

  expect(
    findDesigns()
      .at(0)
      .props('id'),
  ).toBe('2');
});

To mock multiple query response states, success and failure, Apollo Client's native retry behavior can combine with Jest's mock functions to create a series of responses. These do not need to be advanced manually, but they do need to be awaited in specific fashion.

describe('when query times out', () => {
  const advanceApolloTimers = async () => {
    jest.runOnlyPendingTimers();
    await waitForPromises()
  };

  beforeEach(async () => {
    const failSucceedFail = jest
      .fn()
      .mockResolvedValueOnce({ errors: [{ message: 'timeout' }] })
      .mockResolvedValueOnce(mockPipelineResponse)
      .mockResolvedValueOnce({ errors: [{ message: 'timeout' }] });

    createComponentWithApollo(failSucceedFail);
    await waitForPromises();
  });

  it('shows correct errors and does not overwrite populated data when data is empty', async () => {
    /* fails at first, shows error, no data yet */
    expect(getAlert().exists()).toBe(true);
    expect(getGraph().exists()).toBe(false);

    /* succeeds, clears error, shows graph */
    await advanceApolloTimers();
    expect(getAlert().exists()).toBe(false);
    expect(getGraph().exists()).toBe(true);

    /* fails again, alert returns but data persists */
    await advanceApolloTimers();
    expect(getAlert().exists()).toBe(true);
    expect(getGraph().exists()).toBe(true);
  });
});

Previously, we've used { mocks: { $apollo ...}} on mount to test Apollo functionality. This approach is discouraged - proper $apollo mocking leaks a lot of implementation details to the tests. Consider replacing it with mocked Apollo provider

wrapper = mount(SomeComponent, {
  mocks: {
    // avoid! Mock real graphql queries and mutations instead
    $apollo: {
      mutate: jest.fn(),
      queries: {
        groups: {
          loading,
        },
      },
    },
  },
});

Testing @client queries

Using mock resolvers

If your application contains @client queries, you get the following Apollo Client warning when passing only handlers:

Unexpected call of console.warn() with:
Warning: mock-apollo-client - The query is entirely client-side (using @client directives) and resolvers have been configured. The request handler will not be called.

To fix this you should define mock resolvers instead of mock handlers. For example, given the following @client query:

query getBlobContent($path: String, $ref: String!) {
  blobContent(path: $path, ref: $ref) @client {
    rawData
  }
}

And its actual client-side resolvers:

import Api from '~/api';

export const resolvers = {
  Query: {
    blobContent(_, { path, ref }) {
      return {
        __typename: 'BlobContent',
        rawData: Api.getRawFile(path, { ref }).then(({ data }) => {
          return data;
        }),
      };
    },
  },
};

export default resolvers;

We can use a mock resolver that returns data with the same shape, while mock the result with a mock function:

let mockApollo;
let mockBlobContentData; // mock function, jest.fn();

const mockResolvers = {
  Query: {
    blobContent() {
      return {
        __typename: 'BlobContent',
        rawData: mockBlobContentData(), // the mock function can resolve mock data
      };
    },
  },
};

const createComponentWithApollo = ({ props = {} } = {}) => {
  mockApollo = createMockApollo([], mockResolvers); // resolvers are the second parameter

  wrapper = shallowMount(MyComponent, {
    propsData: {},
    apolloProvider: mockApollo,
    // ...
  })
};

After which, you can resolve or reject the value needed.

beforeEach(() => {
  mockBlobContentData = jest.fn();
});

it('shows data', async() => {
  mockBlobContentData.mockResolvedValue(data); // you may resolve or reject to mock the result

  createComponentWithApollo();

  await waitForPromises(); // wait on the resolver mock to execute

  expect(findContent().text()).toBe(mockCiYml);
});

Using cache.writeQuery

Sometimes we want to test a result hook of the local query. In order to have it triggered, we need to populate a cache with correct data to be fetched with this query:

query fetchLocalUser {
  fetchLocalUser @client {
    name
  }
}

import fetchLocalUserQuery from '~/design_management/graphql/queries/fetch_local_user.query.graphql';

describe('Some component with Apollo mock', () => {
  let wrapper;
  let mockApollo;

  function createComponent(options = {
    designListHandler: jest.fn().mockResolvedValue(designListQueryResponse)
  }) {
    mockApollo = createMockApollo([...])
    mockApollo.clients.defaultClient.cache.writeQuery({
      query: fetchLocalUserQuery,
      data: {
        fetchLocalUser: {
          __typename: 'User',
          name: 'Test',
        },
      },
    });

    wrapper = shallowMount(SomeComponent, {
      apolloProvider: mockApollo
    });
  }
})

When you need to configure the mocked apollo client's caching behavior, provide additional cache options when creating a mocked client instance and the provided options will merge with the default cache option:

const defaultCacheOptions = {
  fragmentMatcher: { match: () => true },
  addTypename: false,
};

mockApollo = createMockApollo(
  requestHandlers,
  {},
  {
    dataIdFromObject: (object) =>
      // eslint-disable-next-line no-underscore-dangle
      object.__typename === 'Requirement' ? object.iid : defaultDataIdFromObject(object),
  },
);

Handling errors

The GitLab GraphQL mutations have two distinct error modes: Top-level and errors-as-data.

When utilising a GraphQL mutation, consider handling both of these error modes to ensure that the user receives the appropriate feedback when an error occurs.

Top-level errors

These errors are located at the "top level" of a GraphQL response. These are non-recoverable errors including argument errors and syntax errors, and should not be presented directly to the user.

Handling top-level errors

Apollo is aware of top-level errors, so we are able to leverage Apollo's various error-handling mechanisms to handle these errors. For example, handling Promise rejections after invoking the mutate method, or handling the error event emitted from the ApolloMutation component.

Because these errors are not intended for users, error messages for top-level errors should be defined client-side.

Errors-as-data

These errors are nested in the data object of a GraphQL response. These are recoverable errors that, ideally, can be presented directly to the user.

Handling errors-as-data

First, we must add errors to our mutation object:

mutation createNoteMutation($input: String!) {
  createNoteMutation(input: $input) {
    note {
      id
+     errors
    }
  }

Now, when we commit this mutation and errors occur, the response includes errors for us to handle:

{
  data: {
    mutationName: {
      errors: ["Sorry, we were not able to update the note."]
    }
  }
}

When handling errors-as-data, use your best judgement to determine whether to present the error message in the response, or another message defined client-side, to the user.

Usage outside of Vue

It is also possible to use GraphQL outside of Vue by directly importing and using the default client with queries.

import createDefaultClient from '~/lib/graphql';
import query from './query.graphql';

const defaultClient = createDefaultClient();

defaultClient.query({ query })
  .then(result => console.log(result));

When using Vuex, disable the cache when:

• The data is being cached elsewhere
• The use case does not need caching if the data is being cached elsewhere, or if there is no need for it for the given use case.

import createDefaultClient from '~/lib/graphql';
import fetchPolicies from '~/graphql_shared/fetch_policy_constants';

const defaultClient = createDefaultClient(
  {},
  {
    fetchPolicy: fetchPolicies.NO_CACHE,
  },
);

Making initial queries early with GraphQL startup calls

To improve performance, sometimes we want to make initial GraphQL queries early. In order to do this, we can add them to startup calls with the following steps:

• Move all the queries you need initially in your application to app/graphql/queries;

• Add __typename property to every nested query level:

  query getPermissions($projectPath: ID!) {
    project(fullPath: $projectPath) {
      __typename
      userPermissions {
        __typename
        pushCode
        forkProject
        createMergeRequestIn
      }
    }
  }

• If queries contain fragments, you need to move fragments to the query file directly instead of importing them:

  fragment PageInfo on PageInfo {
    __typename
    hasNextPage
    hasPreviousPage
    startCursor
    endCursor
  }
  
  query getFiles(
    $projectPath: ID!
    $path: String
    $ref: String!
  ) {
    project(fullPath: $projectPath) {
      __typename
      repository {
        __typename
        tree(path: $path, ref: $ref) {
          __typename
            pageInfo {
              ...PageInfo
            }
          }
        }
      }
    }
  }

• If the fragment is used only once, we can also remove the fragment altogether:

  query getFiles(
    $projectPath: ID!
    $path: String
    $ref: String!
  ) {
    project(fullPath: $projectPath) {
      __typename
      repository {
        __typename
        tree(path: $path, ref: $ref) {
          __typename
            pageInfo {
              __typename
              hasNextPage
              hasPreviousPage
              startCursor
              endCursor
            }
          }
        }
      }
    }
  }

• Add startup calls with correct variables to the HAML file that serves as a view for your application. To add GraphQL startup calls, we use add_page_startup_graphql_call helper where the first parameter is a path to the query, the second one is an object containing query variables. Path to the query is relative to app/graphql/queries folder: for example, if we need a app/graphql/queries/repository/files.query.graphql query, the path is repository/files.

Troubleshooting

Mocked client returns empty objects instead of mock response

If your unit test is failing because the response contains empty objects instead of mock data, add __typename field to the mocked responses.

Alternatively, GraphQL query fixtures automatically adds the __typename for you upon generation.

Warning about losing cache data

Sometimes you can see a warning in the console: Cache data may be lost when replacing the someProperty field of a Query object. To address this problem, either ensure all objects of SomeEntityhave an id or a custom merge function. Check section about multiple queries to resolve an issue.

- current_route_path = request.fullpath.match(/-\/tree\/[^\/]+\/(.+$)/).to_a[1]
- add_page_startup_graphql_call('repository/path_last_commit', { projectPath: @project.full_path, ref: current_ref, path: current_route_path || "" })
- add_page_startup_graphql_call('repository/permissions', { projectPath: @project.full_path })
- add_page_startup_graphql_call('repository/files', { nextPageCursor: "", pageSize: 100, projectPath: @project.full_path, ref: current_ref, path: current_route_path || "/"})