GraphQL pagination

Types of pagination

GitLab uses two primary types of pagination: offset and keyset (sometimes called cursor-based) pagination. The GraphQL API mainly uses keyset pagination, falling back to offset pagination when needed.

Offset pagination

This is the traditional, page-by-page pagination, that is most common, and used across much of GitLab. You can recognize it by a list of page numbers near the bottom of a page, which, when clicked, take you to that page of results.

For example, when you click Page 100, we send 100 to the backend. For example, if each page has say 20 items, the backend calculates 20 * 100 = 2000, and it queries the database by offsetting (skipping) the first 2000 records and pulls the next 20.

page number * page size = where to find my records

There are a couple of problems with this:

• Performance. When we query for page 100 (which gives an offset of 2000), then the database has to scan through the table to that specific offset, and then pick up the next 20 records. As the offset increases, the performance degrades quickly. Read more in The SQL I Love <3. Efficient pagination of a table with 100M records.

• Data stability. When you get the 20 items for page 100 (at offset 2000), GitLab shows those 20 items. If someone then deletes or adds records in page 99 or before, the items at offset 2000 become a different set of items. You can even get into a situation where, when paginating, you could skip over items, because the list keeps changing. Read more in Pagination: You're (Probably) Doing It Wrong.

Keyset pagination

Given any specific record, if you know how to calculate what comes after it, you can query the database for those specific records.

For example, suppose you have a list of issues sorted by creation date. If you know the first item on a page has a specific date (say Jan 1), you can ask for all records that were created after that date and take the first 20. It no longer matters if many are deleted or added, as you always ask for the ones after that date, and so get the correct items.

Unfortunately, there is no easy way to know if the issue created on Jan 1 is on page 20 or page 100.

Some of the benefits and tradeoffs of keyset pagination are

• Performance is much better.

• More data stability for end-users since records are not missing from lists due to deletions or insertions.

• It's the best way to do infinite scrolling.

• It's more difficult to program and maintain. Easy for updated_at and sort_order, complicated (or impossible) for complex sorting scenarios.

Implementation

When pagination is supported for a query, GitLab defaults to using keyset pagination. You can see where this is configured in pagination/connections.rb. If a query returns ActiveRecord::Relation, keyset pagination is automatically used.

This was a conscious decision to support performance and data stability.

However, there are some cases where we have to use the offset pagination connection, OffsetActiveRecordRelationConnection, such as when sorting by label priority in issues, due to the complexity of the sort.

Keyset pagination

The keyset pagination implementation is a subclass of GraphQL::Pagination::ActiveRecordRelationConnection, which is a part of the graphql gem. This is installed as the default for all ActiveRecord::Relation. However, instead of using a cursor based on an offset (which is the default), GitLab uses a more specialized cursor.

The cursor is created by encoding a JSON object which contains the relevant ordering fields. For example:

ordering = {"id"=>"72410125", "created_at"=>"2020-10-08 18:05:21.953398000 UTC"}
json = ordering.to_json
cursor = Base64Bp.urlsafe_encode64(json, padding: false)

"eyJpZCI6IjcyNDEwMTI1IiwiY3JlYXRlZF9hdCI6IjIwMjAtMTAtMDggMTg6MDU6MjEuOTUzMzk4MDAwIFVUQyJ9"

json = Base64Bp.urlsafe_decode64(cursor)
Gitlab::Json.parse(json)

{"id"=>"72410125", "created_at"=>"2020-10-08 18:05:21.953398000 UTC"}

The benefits of storing the order attribute values in the cursor:

• If only the ID of the object were stored, the object and its attributes could be queried. That would require an additional query, and if the object is no longer there, then the needed attributes are not available.
• If an attribute is NULL, then one SQL query can be used. If it's not NULL, then a different SQL query can be used.

Based on whether the main attribute field being sorted on is NULL in the cursor, the proper query condition is built. The last ordering field is considered to be unique (a primary key), meaning the column never contains NULL values.

Limitations of query complexity

We only support two ordering fields, and one of those fields needs to be the primary key.

Here are two examples of pseudocode for the query:

• Two-condition query. X represents the values from the cursor. C represents the columns in the database, sorted in ascending order, using an :after cursor, and with NULL values sorted last.

  X1 IS NOT NULL
    AND
      (C1 > X1)
        OR
      (C1 IS NULL)
        OR
      (C1 = X1
        AND
       C2 > X2)
  
  X1 IS NULL
    AND
      (C1 IS NULL
        AND
       C2 > X2)

  Below is an example based on the relation Issue.order(relative_position: :asc).order(id: :asc) with an after cursor of relative_position: 1500, id: 500:

  when cursor[relative_position] is not NULL
  
      ("issues"."relative_position" > 1500)
      OR (
        "issues"."relative_position" = 1500
        AND
        "issues"."id" > 500
      )
      OR ("issues"."relative_position" IS NULL)
  
  when cursor[relative_position] is NULL
  
      "issues"."relative_position" IS NULL
      AND
      "issues"."id" > 500

• Three-condition query. The example below is not complete, but shows the complexity of adding one more condition. X represents the values from the cursor. C represents the columns in the database, sorted in ascending order, using an :after cursor, and with NULL values sorted last.

  X1 IS NOT NULL
    AND
      (C1 > X1)
        OR
      (C1 IS NULL)
        OR
      (C1 = X1 AND C2 > X2)
        OR
      (C1 = X1
        AND
          X2 IS NOT NULL
            AND
              ((C2 > X2)
                 OR
               (C2 IS NULL)
                 OR
               (C2 = X2 AND C3 > X3)
        OR
          X2 IS NULL.....

By using Gitlab::Graphql::Pagination::Keyset::QueryBuilder, we're able to build the necessary SQL conditions and apply them to the Active Record relation.

Complex queries can be difficult or impossible to use. For example, in issuable.rb, the order_due_date_and_labels_priority method creates a very complex query.

These types of queries are not supported. In these instances, you can use offset pagination.

Offset pagination

There are times when the complexity of sorting is more than our keyset pagination can handle.

For example, in IssuesResolver, when sorting by priority_asc, we can't use keyset pagination as the ordering is much too complex. For more information, read issuable.rb.

In cases like this, we can fall back to regular offset pagination by returning a Gitlab::Graphql::Pagination::OffsetActiveRecordRelationConnection instead of an ActiveRecord::Relation:

    def resolve(parent, finder, **args)
      issues = apply_lookahead(Gitlab::Graphql::Loaders::IssuableLoader.new(parent, finder).batching_find_all)

      if non_stable_cursor_sort?(args[:sort])
        # Certain complex sorts are not supported by the stable cursor pagination yet.
        # In these cases, we use offset pagination, so we return the correct connection.
        Gitlab::Graphql::Pagination::OffsetActiveRecordRelationConnection.new(issues)
      else
        issues
      end
    end

External pagination

There may be times where you need to return data through the GitLab API that is stored in another system. In these cases you may have to paginate a third-party's API.

An example of this is with our Error Tracking implementation, where we proxy Sentry errors through the GitLab API. We do this by calling the Sentry API which enforces its own pagination rules. This means we cannot access the collection within GitLab to perform our own custom pagination.

For consistency, we manually set the pagination cursors based on values returned by the external API, using Gitlab::Graphql::ExternallyPaginatedArray.new(previous_cursor, next_cursor, *items).

You can see an example implementation in the following files:

• types/error__tracking/sentry_error_collection_type.rb which adds an extension to field :errors.
• resolvers/error_tracking/sentry_errors_resolver.rb which returns the data from the resolver.

Testing

Any GraphQL field that supports pagination and sorting should be tested using the sorted paginated query shared example found in graphql/sorted_paginated_query_shared_examples.rb. It helps verify that your sort keys are compatible and that cursors work properly.

This is particularly important when using keyset pagination, as some sort keys might not be supported.

Add a section to your request specs like this:

describe 'sorting and pagination' do
  ...
end

You can then use issues_spec.rb as an example to construct your tests.

graphql/sorted_paginated_query_shared_examples.rb also contains some documentation on how to use the shared examples.

The shared example requires certain let variables and methods to be set up:

describe 'sorting and pagination' do
  let(:sort_project) { create(:project, :public) }
  let(:data_path)    { [:project, :issues] }

  def pagination_query(params, page_info)
    graphql_query_for(
      'project',
      { 'fullPath' => sort_project.full_path },
      query_graphql_field('issues', params, "#{page_info} edges { node { id } }")
    )
  end

  def pagination_results_data(data)
    data.map { |issue| issue.dig('node', 'iid').to_i }
  end

  context 'when sorting by weight' do
    ...
    context 'when ascending' do
      it_behaves_like 'sorted paginated query' do
        let(:sort_param)       { 'WEIGHT_ASC' }
        let(:first_param)      { 2 }
        let(:expected_results) { [weight_issue3.iid, weight_issue5.iid, weight_issue1.iid, weight_issue4.iid, weight_issue2.iid] }
      end
    end