Efficient `IN` operator queries

This document describes a technique for building efficient ordered database queries with the IN SQL operator and the usage of a GitLab utility module to help apply the technique.

NOTE: The described technique makes heavy use of keyset pagination. It's advised to get familiar with the topic first.

Motivation

In GitLab, many domain objects like Issue live under nested hierarchies of projects and groups. To fetch nested database records for domain objects at the group-level, we often perform queries with the IN SQL operator. We are usually interested in ordering the records by some attributes and limiting the number of records using ORDER BY and LIMIT clauses for performance. Pagination may be used to fetch subsequent records.

Example tasks requiring querying nested domain objects from the group level:

Show first 20 issues by creation date or due date from the group gitlab-org.
Show first 20 merge requests by merged at date from the group gitlab-com.

Unfortunately, ordered group-level queries typically perform badly as their executions require heavy I/O, memory, and computations. Let's do an in-depth examination of executing one such query.

Performance problems with `IN` queries

Consider the task of fetching the twenty oldest created issues from the group gitlab-org with the following query:

SELECT "issues".*
FROM "issues"
WHERE "issues"."project_id" IN
    (SELECT "projects"."id"
     FROM "projects"
     WHERE "projects"."namespace_id" IN
         (SELECT traversal_ids[array_length(traversal_ids, 1)] AS id
          FROM "namespaces"
          WHERE (traversal_ids @> ('{9970}'))))
ORDER BY "issues"."created_at" ASC,
         "issues"."id" ASC
LIMIT 20

NOTE: For pagination, ordering by the created_at column is not enough, we must add the id column as a tie-breaker.

The execution of the query can be largely broken down into three steps:

The database accesses both namespaces and projects tables to find all projects from all groups in the group hierarchy.
The database retrieves issues records for each project causing heavy disk I/O. Ideally, an appropriate index configuration should optimize this process.
The database sorts the issues rows in memory by created_at and returns LIMIT 20 rows to the end-user. For large groups, this final step requires both large memory and CPU resources.

Execution plan for this DB query:

 Limit  (cost=90170.07..90170.12 rows=20 width=1329) (actual time=967.597..967.607 rows=20 loops=1)
   Buffers: shared hit=239127 read=3060
   I/O Timings: read=336.879
   ->  Sort  (cost=90170.07..90224.02 rows=21578 width=1329) (actual time=967.596..967.603 rows=20 loops=1)
         Sort Key: issues.created_at, issues.id
         Sort Method: top-N heapsort  Memory: 74kB
         Buffers: shared hit=239127 read=3060
         I/O Timings: read=336.879
         ->  Nested Loop  (cost=1305.66..89595.89 rows=21578 width=1329) (actual time=4.709..797.659 rows=241534 loops=1)
               Buffers: shared hit=239121 read=3060
               I/O Timings: read=336.879
               ->  HashAggregate  (cost=1305.10..1360.22 rows=5512 width=4) (actual time=4.657..5.370 rows=1528 loops=1)
                     Group Key: projects.id
                     Buffers: shared hit=2597
                     ->  Nested Loop  (cost=576.76..1291.32 rows=5512 width=4) (actual time=2.427..4.244 rows=1528 loops=1)
                           Buffers: shared hit=2597
                           ->  HashAggregate  (cost=576.32..579.06 rows=274 width=25) (actual time=2.406..2.447 rows=265 loops=1)
                                 Group Key: namespaces.traversal_ids[array_length(namespaces.traversal_ids, 1)]
                                 Buffers: shared hit=334
                                 ->  Bitmap Heap Scan on namespaces  (cost=141.62..575.63 rows=274 width=25) (actual time=1.933..2.330 rows=265 loops=1)
                                       Recheck Cond: (traversal_ids @> '{9970}'::integer[])
                                       Heap Blocks: exact=243
                                       Buffers: shared hit=334
                                       ->  Bitmap Index Scan on index_namespaces_on_traversal_ids  (cost=0.00..141.55 rows=274 width=0) (actual time=1.897..1.898 rows=265 loops=1)
                                             Index Cond: (traversal_ids @> '{9970}'::integer[])
                                             Buffers: shared hit=91
                           ->  Index Only Scan using index_projects_on_namespace_id_and_id on projects  (cost=0.44..2.40 rows=20 width=8) (actual time=0.004..0.006 rows=6 loops=265)
                                 Index Cond: (namespace_id = (namespaces.traversal_ids)[array_length(namespaces.traversal_ids, 1)])
                                 Heap Fetches: 51
                                 Buffers: shared hit=2263
               ->  Index Scan using index_issues_on_project_id_and_iid on issues  (cost=0.57..10.57 rows=544 width=1329) (actual time=0.114..0.484 rows=158 loops=1528)
                     Index Cond: (project_id = projects.id)
                     Buffers: shared hit=236524 read=3060
                     I/O Timings: read=336.879
 Planning Time: 7.750 ms
 Execution Time: 967.973 ms
(36 rows)

The performance of the query depends on the number of rows in the database. On average, we can say the following:

Number of groups in the group-hierarchy: less than 1 000
Number of projects: less than 5 000
Number of issues: less than 100 000

From the list, it's apparent that the number of issues records has the largest impact on the performance. As per typical usage, we can say that the number of issue records grows at a faster rate than the namespaces and the projects records.

This problem affects most of our group-level features where records are listed in a specific order, such as group-level issues, merge requests pages, and APIs. For very large groups the database queries can easily time out, causing HTTP 500 errors.

Optimizing ordered `IN` queries

In the talk "How to teach an elephant to dance rock and roll", Maxim Boguk demonstrated a technique to optimize a special class of ordered IN queries, such as our ordered group-level queries.

A typical ordered IN query may look like this:

SELECT t.* FROM t
WHERE t.fkey IN (value_set)
ORDER BY t.pkey
LIMIT N;

Here's the key insight used in the technique: we need at most |value_set| + N record lookups, rather than retrieving all records satisfying the condition t.fkey IN value_set (|value_set| is the number of values in value_set).

We adopted and generalized the technique for use in GitLab by implementing utilities in the Gitlab::Pagination::Keyset::InOperatorOptimization class to facilitate building efficient IN queries.

Requirements

The technique is not a drop-in replacement for the existing group-level queries using IN operator. The technique can only optimize IN queries that satisfy the following requirements:

LIMIT is present, which usually means that the query is paginated (offset or keyset pagination).
The column used with the IN query and the columns in the ORDER BY clause are covered with a database index. The columns in the index must be in the following order: column_for_the_in_query, order by column 1, and order by column 2.
The columns in the ORDER BY clause are distinct (the combination of the columns uniquely identifies one particular row in the table).

WARNING: This technique does not improve the performance of the COUNT(*) queries.

The `InOperatorOptimization` module

The Gitlab::Pagination::Keyset::InOperatorOptimization module implements utilities for applying a generalized version of the efficient IN query technique described in the previous section.

To build optimized, ordered IN queries that meet the requirements, use the utility class QueryBuilder from the module.

NOTE: The generic keyset pagination module introduced in the merge request 51481 plays a fundamental role in the generalized implementation of the technique in Gitlab::Pagination::Keyset::InOperatorOptimization.

Basic usage of `QueryBuilder`

To illustrate a basic usage, we build a query that fetches 20 issues with the oldest created_at from the group gitlab-org.

The following ActiveRecord query would produce a query similar to the unoptimized query that we examined earlier:

scope = Issue
  .where(project_id: Group.find(9970).all_projects.select(:id)) # `gitlab-org` group and its subgroups
  .order(:created_at, :id)
  .limit(20)

Instead, use the query builder InOperatorOptimization::QueryBuilder to produce an optimized version:

scope = Issue.order(:created_at, :id)
array_scope = Group.find(9970).all_projects.select(:id)
array_mapping_scope = -> (id_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)) }
finder_query = -> (created_at_expression, id_expression) { Issue.where(Issue.arel_table[:id].eq(id_expression)) }

Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder.new(
  scope: scope,
  array_scope: array_scope,
  array_mapping_scope: array_mapping_scope,
  finder_query: finder_query
).execute.limit(20)

scope represents the original ActiveRecord::Relation object without the IN query. The relation should define an order which must be supported by the keyset pagination library.
array_scope contains the ActiveRecord::Relation object, which represents the original IN (subquery). The select values must contain the columns by which the subquery is "connected" to the main query: the id of the project record.
array_mapping_scope defines a lambda returning an ActiveRecord::Relation object. The lambda matches (=) single select values from the array_scope. The lambda yields as many arguments as the select values defined in the array_scope. The arguments are Arel SQL expressions.
finder_query loads the actual record row from the database. It must also be a lambda, where the order by column expressions is available for locating the record. In this example, the yielded values are created_at and id SQL expressions. Finding a record is very fast via the primary key, so we don't use the created_at value. Providing the finder_query lambda is optional. If it's not given, the IN operator optimization only makes the ORDER BY columns available to the end-user and not the full database row.

The following database index on the issues table must be present to make the query execute efficiently:

"idx_issues_on_project_id_and_created_at_and_id" btree (project_id, created_at, id)

The SQL query:

SELECT "issues".*
FROM
  (WITH RECURSIVE "array_cte" AS MATERIALIZED
     (SELECT "projects"."id"
 FROM "projects"
 WHERE "projects"."namespace_id" IN
     (SELECT traversal_ids[array_length(traversal_ids, 1)] AS id
      FROM "namespaces"
      WHERE (traversal_ids @> ('{9970}')))),
                  "recursive_keyset_cte" AS (  -- initializer row start
                                               (SELECT NULL::issues AS records,
                                                       array_cte_id_array,
                                                       issues_created_at_array,
                                                       issues_id_array,
                                                       0::bigint AS COUNT
                                                FROM
                                                  (SELECT ARRAY_AGG("array_cte"."id") AS array_cte_id_array,
                                                          ARRAY_AGG("issues"."created_at") AS issues_created_at_array,
                                                          ARRAY_AGG("issues"."id") AS issues_id_array
                                                   FROM
                                                     (SELECT "array_cte"."id"
                                                      FROM array_cte) array_cte
                                                   LEFT JOIN LATERAL
                                                     (SELECT "issues"."created_at",
                                                             "issues"."id"
                                                      FROM "issues"
                                                      WHERE "issues"."project_id" = "array_cte"."id"
                                                      ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
                                                      LIMIT 1) issues ON TRUE
                                                   WHERE "issues"."created_at" IS NOT NULL
                                                     AND "issues"."id" IS NOT NULL) array_scope_lateral_query
                                                LIMIT 1)
                                                -- initializer row finished
                                             UNION ALL
                                               (SELECT
                                                  -- result row start
                                                  (SELECT issues -- record finder query as the first column
                                                   FROM "issues"
                                                   WHERE "issues"."id" = recursive_keyset_cte.issues_id_array[position]
                                                   LIMIT 1),
                                                   array_cte_id_array,
                                                   recursive_keyset_cte.issues_created_at_array[:position_query.position-1]||next_cursor_values.created_at||recursive_keyset_cte.issues_created_at_array[position_query.position+1:],
                                                   recursive_keyset_cte.issues_id_array[:position_query.position-1]||next_cursor_values.id||recursive_keyset_cte.issues_id_array[position_query.position+1:],
                                                   recursive_keyset_cte.count + 1
                                                -- result row finished
                                                FROM recursive_keyset_cte,
                                                     LATERAL
                                                  -- finding the cursor values of the next record start
                                                  (SELECT created_at,
                                                          id,
                                                          position
                                                   FROM UNNEST(issues_created_at_array, issues_id_array) WITH
                                                   ORDINALITY AS u(created_at, id, position)
                                                   WHERE created_at IS NOT NULL
                                                     AND id IS NOT NULL
                                                   ORDER BY "created_at" ASC, "id" ASC
                                                   LIMIT 1) AS position_query,
                                                  -- finding the cursor values of the next record end
                                                  -- finding the next cursor values (next_cursor_values_query) start
                                                             LATERAL
                                                  (SELECT "record"."created_at",
                                                          "record"."id"
                                                   FROM (
                                                         VALUES (NULL,
                                                                 NULL)) AS nulls
                                                   LEFT JOIN
                                                     (SELECT "issues"."created_at",
                                                             "issues"."id"
                                                      FROM (
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[position]
                                                                 AND recursive_keyset_cte.issues_created_at_array[position] IS NULL
                                                                 AND "issues"."created_at" IS NULL
                                                                 AND "issues"."id" > recursive_keyset_cte.issues_id_array[position]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[position]
                                                                 AND recursive_keyset_cte.issues_created_at_array[position] IS NOT NULL
                                                                 AND "issues"."created_at" IS NULL
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[position]
                                                                 AND recursive_keyset_cte.issues_created_at_array[position] IS NOT NULL
                                                                 AND "issues"."created_at" > recursive_keyset_cte.issues_created_at_array[position]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[position]
                                                                 AND recursive_keyset_cte.issues_created_at_array[position] IS NOT NULL
                                                                 AND "issues"."created_at" = recursive_keyset_cte.issues_created_at_array[position]
                                                                 AND "issues"."id" > recursive_keyset_cte.issues_id_array[position]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)) issues
                                                      ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
                                                      LIMIT 1) record ON TRUE
                                                   LIMIT 1) AS next_cursor_values))
                                                  -- finding the next cursor values (next_cursor_values_query) END
SELECT (records).*
   FROM "recursive_keyset_cte" AS "issues"
   WHERE (COUNT <> 0)) issues -- filtering out the initializer row
LIMIT 20

Using the `IN` query optimization

Adding more filters

In this example, let's add an extra filter by milestone_id.

Be careful when adding extra filters to the query. If the column is not covered by the same index, then the query might perform worse than the non-optimized query. The milestone_id column in the issues table is currently covered by a different index:

"index_issues_on_milestone_id" btree (milestone_id)

Adding the milestone_id = X filter to the scope argument or to the optimized scope causes bad performance.

Example (bad):

Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder.new(
  scope: scope,
  array_scope: array_scope,
  array_mapping_scope: array_mapping_scope,
  finder_query: finder_query
).execute
  .where(milestone_id: 5)
  .limit(20)

To address this concern, we could define another index:

"idx_issues_on_project_id_and_milestone_id_and_created_at_and_id" btree (project_id, milestone_id, created_at, id)

Adding more indexes to the issues table could significantly affect the performance of the UPDATE queries. In this case, it's better to rely on the original query. It means that if we want to use the optimization for the unfiltered page we need to add extra logic in the application code:

if optimization_possible? # no extra params or params covered with the same index as the ORDER BY clause
  run_optimized_query
else
  run_normal_in_query
end

Multiple `IN` queries

Let's assume that we want to extend the group-level queries to include only incident and test case issue types.

The original ActiveRecord query would look like this:

scope = Issue
  .where(project_id: Group.find(9970).all_projects.select(:id)) # `gitlab-org` group and its subgroups
  .where(issue_type: [:incident, :test_case]) # 1, 2
  .order(:created_at, :id)
  .limit(20)

To construct the array scope, we need to take the Cartesian product of the project_id IN and the issue_type IN queries. issue_type is an ActiveRecord enum, so we need to construct the following table:

`project_id`	`issue_type_value`
2	1
2	2
5	1
5	2
10	1
10	2
9	1
9	2

For the issue_types query we can construct a value list without querying a table:

value_list = Arel::Nodes::ValuesList.new([[WorkItems::Type.base_types[:incident]],[WorkItems::Type.base_types[:test_case]]])
issue_type_values = Arel::Nodes::Grouping.new(value_list).as('issue_type_values (value)').to_sql

array_scope = Group
  .find(9970)
  .all_projects
  .from("#{Project.table_name}, #{issue_type_values}")
  .select(:id, :value)

Building the array_mapping_scope query requires two arguments: id and issue_type_value:

array_mapping_scope = -> (id_expression, issue_type_value_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)).where(Issue.arel_table[:issue_type].eq(issue_type_value_expression)) }

The scope and the finder queries don't change:

scope = Issue.order(:created_at, :id)
finder_query = -> (created_at_expression, id_expression) { Issue.where(Issue.arel_table[:id].eq(id_expression)) }

Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder.new(
  scope: scope,
  array_scope: array_scope,
  array_mapping_scope: array_mapping_scope,
  finder_query: finder_query
).execute.limit(20)

The SQL query:

SELECT "issues".*
FROM
  (WITH RECURSIVE "array_cte" AS MATERIALIZED
     (SELECT "projects"."id", "value"
      FROM projects, (
                      VALUES (1), (2)) AS issue_type_values (value)
      WHERE "projects"."namespace_id" IN
          (WITH RECURSIVE "base_and_descendants" AS (
                                                       (SELECT "namespaces".*
                                                        FROM "namespaces"
                                                        WHERE "namespaces"."type" = 'Group'
                                                          AND "namespaces"."id" = 9970)
                                                     UNION
                                                       (SELECT "namespaces".*
                                                        FROM "namespaces", "base_and_descendants"
                                                        WHERE "namespaces"."type" = 'Group'
                                                          AND "namespaces"."parent_id" = "base_and_descendants"."id")) SELECT "id"
           FROM "base_and_descendants" AS "namespaces")),
                  "recursive_keyset_cte" AS (
                                               (SELECT NULL::issues AS records,
                                                       array_cte_id_array,
                                                       array_cte_value_array,
                                                       issues_created_at_array,
                                                       issues_id_array,
                                                       0::bigint AS COUNT
                                                FROM
                                                  (SELECT ARRAY_AGG("array_cte"."id") AS array_cte_id_array,
                                                          ARRAY_AGG("array_cte"."value") AS array_cte_value_array,
                                                          ARRAY_AGG("issues"."created_at") AS issues_created_at_array,
                                                          ARRAY_AGG("issues"."id") AS issues_id_array
                                                   FROM
                                                     (SELECT "array_cte"."id",
                                                             "array_cte"."value"
                                                      FROM array_cte) array_cte
                                                   LEFT JOIN LATERAL
                                                     (SELECT "issues"."created_at",
                                                             "issues"."id"
                                                      FROM "issues"
                                                      WHERE "issues"."project_id" = "array_cte"."id"
                                                        AND "issues"."issue_type" = "array_cte"."value"
                                                      ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
                                                      LIMIT 1) issues ON TRUE
                                                   WHERE "issues"."created_at" IS NOT NULL
                                                     AND "issues"."id" IS NOT NULL) array_scope_lateral_query
                                                LIMIT 1)
                                             UNION ALL
                                               (SELECT
                                                  (SELECT issues
                                                   FROM "issues"
                                                   WHERE "issues"."id" = recursive_keyset_cte.issues_id_array[POSITION]
                                                   LIMIT 1), array_cte_id_array,
                                                             array_cte_value_array,
                                                             recursive_keyset_cte.issues_created_at_array[:position_query.position-1]||next_cursor_values.created_at||recursive_keyset_cte.issues_created_at_array[position_query.position+1:], recursive_keyset_cte.issues_id_array[:position_query.position-1]||next_cursor_values.id||recursive_keyset_cte.issues_id_array[position_query.position+1:], recursive_keyset_cte.count + 1
                                                FROM recursive_keyset_cte,
                                                     LATERAL
                                                  (SELECT created_at,
                                                          id,
                                                          POSITION
                                                   FROM UNNEST(issues_created_at_array, issues_id_array) WITH
                                                   ORDINALITY AS u(created_at, id, POSITION)
                                                   WHERE created_at IS NOT NULL
                                                     AND id IS NOT NULL
                                                   ORDER BY "created_at" ASC, "id" ASC
                                                   LIMIT 1) AS position_query,
                                                             LATERAL
                                                  (SELECT "record"."created_at",
                                                          "record"."id"
                                                   FROM (
                                                         VALUES (NULL,
                                                                 NULL)) AS nulls
                                                   LEFT JOIN
                                                     (SELECT "issues"."created_at",
                                                             "issues"."id"
                                                      FROM (
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[POSITION]
                                                                 AND "issues"."issue_type" = recursive_keyset_cte.array_cte_value_array[POSITION]
                                                                 AND recursive_keyset_cte.issues_created_at_array[POSITION] IS NULL
                                                                 AND "issues"."created_at" IS NULL
                                                                 AND "issues"."id" > recursive_keyset_cte.issues_id_array[POSITION]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[POSITION]
                                                                 AND "issues"."issue_type" = recursive_keyset_cte.array_cte_value_array[POSITION]
                                                                 AND recursive_keyset_cte.issues_created_at_array[POSITION] IS NOT NULL
                                                                 AND "issues"."created_at" IS NULL
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[POSITION]
                                                                 AND "issues"."issue_type" = recursive_keyset_cte.array_cte_value_array[POSITION]
                                                                 AND recursive_keyset_cte.issues_created_at_array[POSITION] IS NOT NULL
                                                                 AND "issues"."created_at" > recursive_keyset_cte.issues_created_at_array[POSITION]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)
                                                            UNION ALL
                                                              (SELECT "issues"."created_at",
                                                                      "issues"."id"
                                                               FROM "issues"
                                                               WHERE "issues"."project_id" = recursive_keyset_cte.array_cte_id_array[POSITION]
                                                                 AND "issues"."issue_type" = recursive_keyset_cte.array_cte_value_array[POSITION]
                                                                 AND recursive_keyset_cte.issues_created_at_array[POSITION] IS NOT NULL
                                                                 AND "issues"."created_at" = recursive_keyset_cte.issues_created_at_array[POSITION]
                                                                 AND "issues"."id" > recursive_keyset_cte.issues_id_array[POSITION]
                                                               ORDER BY "issues"."created_at" ASC, "issues"."id" ASC)) issues
                                                      ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
                                                      LIMIT 1) record ON TRUE
                                                   LIMIT 1) AS next_cursor_values)) SELECT (records).*
   FROM "recursive_keyset_cte" AS "issues"
   WHERE (COUNT <> 0)) issues
LIMIT 20

NOTE: To make the query efficient, the following columns need to be covered with an index: project_id, issue_type, created_at, and id.

Using calculated `ORDER BY` expression

The following example orders epic records by the duration between the creation time and closed time. It is calculated with the following formula:

SELECT EXTRACT('epoch' FROM epics.closed_at - epics.created_at) FROM epics

The query above returns the duration in seconds (double precision) between the two timestamp columns in seconds. To order the records by this expression, you must reference it in the ORDER BY clause:

SELECT EXTRACT('epoch' FROM epics.closed_at - epics.created_at)
FROM epics
ORDER BY EXTRACT('epoch' FROM epics.closed_at - epics.created_at) DESC

To make this ordering efficient on the group-level with the in-operator optimization, use a custom ORDER BY configuration. Since the duration is not a distinct value (no unique index present), you must add a tie-breaker column (id).

The following example shows the final ORDER BY clause:

ORDER BY extract('epoch' FROM epics.closed_at - epics.created_at) DESC, epics.id DESC

Snippet for loading records ordered by the calculated duration:

arel_table =  Epic.arel_table
order = Gitlab::Pagination::Keyset::Order.build([
  Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
    attribute_name: 'duration_in_seconds',
    order_expression: Arel.sql('EXTRACT(EPOCH FROM epics.closed_at - epics.created_at)').desc,
    sql_type: 'double precision' # important for calculated SQL expressions
  ),
  Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
    attribute_name: 'id',
    order_expression: arel_table[:id].desc
  )
])

records = Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder.new(
  scope: Epic.where.not(closed_at: nil).reorder(order), # filter out NULL values
  array_scope: Group.find(9970).self_and_descendants.select(:id),
  array_mapping_scope: -> (id_expression) { Epic.where(Epic.arel_table[:group_id].eq(id_expression)) }
).execute.limit(20)

puts records.pluck(:duration_in_seconds, :id) # other columns are not available

Building the query requires quite a bit of configuration. For the order configuration you can find more information within the complex order configuration section for keyset paginated database queries.

The query requires a specialized database index:

CREATE INDEX index_epics_on_duration ON epics USING btree (group_id, EXTRACT(EPOCH FROM epics.closed_at - epics.created_at) DESC, id DESC) WHERE (closed_at IS NOT NULL);

Notice that the finder_query parameter is not used. The query only returns the ORDER BY columns which are the duration_in_seconds (calculated column) and the id columns. This is a limitation of the feature, defining the finder_query with calculated ORDER BY expressions is not supported. To get the complete database records, an extra query can be invoked by the returned id column:

records_by_id = records.index_by(&:id)
complete_records = Epic.where(id: records_by_id.keys).index_by(&:id)

# Printing the complete records according to the `ORDER BY` clause
records_by_id.each do |id, _|
  puts complete_records[id].attributes
end

Ordering by `JOIN` columns

Ordering records by mixed columns where one or more columns are coming from JOIN tables works with limitations. It requires extra configuration via Common Table Expression (CTE). The trick is to use a non-materialized CTE to act as a "fake" table which exposes all required columns.

NOTE: The query performance might not improve compared to the standard IN query. Always check the query plan.

Example: order issues by projects.name, issues.id within the group hierarchy

The first step is to create a CTE, where all required columns are collected in the SELECT clause.

cte_query = Issue
  .select('issues.id AS id', 'issues.project_id AS project_id', 'projects.name AS projects_name')
  .joins(:project)

cte = Gitlab::SQL::CTE.new(:issue_with_projects, cte_query, materialized: false)

Custom order object configuration:

order = Gitlab::Pagination::Keyset::Order.build([
          Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
            attribute_name: 'projects_name',
            order_expression: Issue.arel_table[:projects_name].asc,
            sql_type: 'character varying',
            nullable: :nulls_last
          ),
          Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
            attribute_name: :id,
            order_expression: Issue.arel_table[:id].asc
          )
        ])

Generate the query:

scope = cte.apply_to(Issue.where({}).reorder(order))

opts = {
  scope: scope,
  array_scope: Project.where(namespace_id: top_level_group.self_and_descendants.select(:id)).select(:id),
  array_mapping_scope: -> (id_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)) }
}

records = Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder
  .new(**opts)
  .execute
  .limit(20)

Batch iteration

Batch iteration over the records is possible via the keyset Iterator class.

scope = Issue.order(:created_at, :id)
array_scope = Group.find(9970).all_projects.select(:id)
array_mapping_scope = -> (id_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)) }
finder_query = -> (created_at_expression, id_expression) { Issue.where(Issue.arel_table[:id].eq(id_expression)) }

opts = {
  in_operator_optimization_options: {
    array_scope: array_scope,
    array_mapping_scope: array_mapping_scope,
    finder_query: finder_query
  }
}

Gitlab::Pagination::Keyset::Iterator.new(scope: scope, **opts).each_batch(of: 100) do |records|
  puts records.select(:id).map { |r| [r.id] }
end

NOTE: The query loads complete database rows from the disk. This may cause increased I/O and slower database queries. Depending on the use case, the primary key is often only needed for the batch query to invoke additional statements. For example, UPDATE or DELETE. The id column is included in the ORDER BY columns (created_at and id) and is already loaded. In this case, you can omit the finder_query parameter.

Example for loading the ORDER BY columns only:

scope = Issue.order(:created_at, :id)
array_scope = Group.find(9970).all_projects.select(:id)
array_mapping_scope = -> (id_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)) }

opts = {
  in_operator_optimization_options: {
    array_scope: array_scope,
    array_mapping_scope: array_mapping_scope
  }
}

Gitlab::Pagination::Keyset::Iterator.new(scope: scope, **opts).each_batch(of: 100) do |records|
  puts records.select(:id).map { |r| [r.id] } # only id and created_at are available
end

Keyset pagination

The optimization works out of the box with GraphQL and the keyset_paginate helper method. Read more about keyset pagination.

array_scope = Group.find(9970).all_projects.select(:id)
array_mapping_scope = -> (id_expression) { Issue.where(Issue.arel_table[:project_id].eq(id_expression)) }
finder_query = -> (created_at_expression, id_expression) { Issue.where(Issue.arel_table[:id].eq(id_expression)) }

opts = {
  in_operator_optimization_options: {
    array_scope: array_scope,
    array_mapping_scope: array_mapping_scope,
    finder_query: finder_query
  }
}

issues = Issue
  .order(:created_at, :id)
  .keyset_paginate(per_page: 20, keyset_order_options: opts)
  .records

Offset pagination with Kaminari

The ActiveRecord scope produced by the InOperatorOptimization class can be used in offset-paginated queries.

Gitlab::Pagination::Keyset::InOperatorOptimization::QueryBuilder
  .new(...)
  .execute
  .page(1)
  .per(20)
  .without_count

Generalized `IN` optimization technique

Let's dive into how QueryBuilder builds the optimized query to fetch the twenty oldest created issues from the group gitlab-org using the generalized IN optimization technique.

Array CTE

As the first step, we use a Common Table Expression (CTE) for collecting the projects.id values. This is done by wrapping the incoming array_scope ActiveRecord relation parameter with a CTE.

WITH array_cte AS MATERIALIZED (
  SELECT "projects"."id"
   FROM "projects"
   WHERE "projects"."namespace_id" IN
       (SELECT traversal_ids[array_length(traversal_ids, 1)] AS id
        FROM "namespaces"
        WHERE (traversal_ids @> ('{9970}')))
)

This query produces the following result set with only one column (projects.id):

ID
9
2
5
10

Array mapping

For each project (that is, each record storing a project ID in array_cte), we fetch the cursor value identifying the first issue respecting the ORDER BY clause.

As an example, let's pick the first record ID=9 from array_cte. The following query should fetch the cursor value (created_at, id) identifying the first issue record respecting the ORDER BY clause for the project with ID=9:

SELECT "issues"."created_at", "issues"."id"
FROM "issues"."project_id"=9
ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
LIMIT 1;

We use LATERAL JOIN to loop over the records in the array_cte and find the cursor value for each project. The query would be built using the array_mapping_scope lambda function.

SELECT ARRAY_AGG("array_cte"."id") AS array_cte_id_array,
  ARRAY_AGG("issues"."created_at") AS issues_created_at_array,
  ARRAY_AGG("issues"."id") AS issues_id_array
FROM (
  SELECT "array_cte"."id" FROM array_cte
) array_cte
LEFT JOIN LATERAL
(
  SELECT "issues"."created_at", "issues"."id"
  FROM "issues"
  WHERE "issues"."project_id" = "array_cte"."id"
  ORDER BY "issues"."created_at" ASC, "issues"."id" ASC
  LIMIT 1
) issues ON TRUE

Since we have an index on project_id, created_at, and id, index-only scans should quickly locate all the cursor values.

This is how the query could be translated to Ruby:

created_at_values = []
id_values = []
project_ids.map do |project_id|
  created_at, id = Issue.select(:created_at, :id).where(project_id: project_id).order(:created_at, :id).limit(1).first # N+1 but it's fast
  created_at_values << created_at
  id_values << id
end

This is what the result set would look like:

`project_ids`	`created_at_values`	`id_values`
2	2020-01-10	5
5	2020-01-05	4
10	2020-01-15	7
9	2020-01-05	3

The table shows the cursor values (created_at, id) of the first record for each project respecting the ORDER BY clause.

At this point, we have the initial data. To start collecting the actual records from the database, we use a recursive CTE query where each recursion locates one row until the LIMIT is reached or no more data can be found.

Here's an outline of the steps we take in the recursive CTE query (expressing the steps in SQL is non-trivial but is explained next):

Sort the initial resultset according to the ORDER BY clause.
Pick the top cursor to fetch the record, this is our first record. In the example, this cursor would be (2020-01-05, 3) for project_id=9.
We can use (2020-01-05, 3) to fetch the next issue respecting the ORDER BY clause project_id=9 filter. This produces an updated resultset.

project_ids created_at_values id_values

2 2020-01-10 5

5 2020-01-05 4

10 2020-01-15 7

9 2020-01-06 6
Repeat 1 to 3 with the updated resultset until we have fetched N=20 records.

Initializing the recursive CTE query

For the initial recursive query, we need to produce exactly one row, we call this the initializer query (initializer_query).

Use ARRAY_AGG function to compact the initial result set into a single row and use the row as the initial value for the recursive CTE query:

Example initializer row:

`records`	`project_ids`	`created_at_values`	`id_values`	`Count`	`Position`
`NULL::issues`	`[9, 2, 5, 10]`	`[...]`	`[...]`	`0`	`NULL`

The records column contains our sorted database records, and the initializer query sets the first value to NULL, which is filtered out later.
The count column tracks the number of records found. We use this column to filter out the initializer row from the result set.

Recursive portion of the CTE query

The result row is produced with the following steps:

Order the keyset arrays.
Find the next cursor.
Produce a new row.

Order the keyset arrays

Order the keyset arrays according to the original ORDER BY clause with LIMIT 1 using the UNNEST [] WITH ORDINALITY table function. The function locates the "lowest" keyset cursor values and gives us the array position. These cursor values are used to locate the record.

NOTE: At this point, we haven't read anything from the database tables, because we relied on fast index-only scans.

`project_ids`	`created_at_values`	`id_values`
2	2020-01-10	5
5	2020-01-05	4
10	2020-01-15	7
9	2020-01-05	3

The first row is the 4th one (position = 4), because it has the lowest created_at and id values. The UNNEST function also exposes the position using an extra column (note: PostgreSQL uses 1-based index).

Demonstration of the UNNEST [] WITH ORDINALITY table function:

SELECT position FROM unnest('{2020-01-10, 2020-01-05, 2020-01-15, 2020-01-05}'::timestamp[], '{5, 4, 7, 3}'::int[])
  WITH ORDINALITY AS t(created_at, id, position) ORDER BY created_at ASC, id ASC LIMIT 1;

Result:

position
----------
         4
(1 row)

Find the next cursor

Now, let's find the next cursor values (next_cursor_values_query) for the project with id = 9. To do that, we build a keyset pagination SQL query. Find the next row after created_at = 2020-01-05 and id = 3. Because we order by two database columns, there can be two cases:

There are rows with created_at = 2020-01-05 and id > 3.
There are rows with created_at > 2020-01-05.

Generating this query is done by the generic keyset pagination library. After the query is done, we have a temporary table with the next cursor values:

`created_at`	ID
2020-01-06	6

Produce a new row

As the final step, we need to produce a new row by manipulating the initializer row (data_collector_query method). Two things happen here:

Read the full row from the DB and return it in the records columns. (result_collector_columns method)
Replace the cursor values at the current position with the results from the keyset query.

Reading the full row from the database is only one index scan by the primary key. We use the ActiveRecord query passed as the finder_query:

(SELECT "issues".* FROM issues WHERE id = id_values[position] LIMIT 1)

By adding parentheses, the result row can be put into the records column.

Replacing the cursor values at position can be done via standard PostgreSQL array operators:

-- created_at_values column value
created_at_values[:position-1]||next_cursor_values.created_at||created_at_values[position+1:]

-- id_values column value
id_values[:position-1]||next_cursor_values.id||id_values[position+1:]

The Ruby equivalent would be the following:

id_values[0..(position - 1)] + [next_cursor_values.id] + id_values[(position + 1)..-1]

After this, the recursion starts again by finding the next lowest cursor value.

Finalizing the query

For producing the final issues rows, we wrap the query with another SELECT statement:

SELECT "issues".*
FROM (
  SELECT (records).* -- similar to ruby splat operator
  FROM recursive_keyset_cte
  WHERE recursive_keyset_cte.count <> 0 -- filter out the initializer row
) AS issues

Performance comparison

Assuming that we have the correct database index in place, we can compare the query performance by looking at the number of database rows accessed by the query.

Number of groups: 100
Number of projects: 500
Number of issues (in the group hierarchy): 50 000

Standard IN query:

Query	Entries read from index	Rows read from the table	Rows sorted in memory
group hierarchy subquery	100	0	0
project lookup query	500	0	0
issue lookup query	50 000	20	50 000

Optimized IN query:

Query	Entries read from index	Rows read from the table	Rows sorted in memory
group hierarchy subquery	100	0	0
project lookup query	500	0	0
issue lookup query	519	20	10 000

The group and project queries are not using sorting, the necessary columns are read from database indexes. These values are accessed frequently so it's very likely that most of the data is in the PostgreSQL's buffer cache.

The optimized IN query reads maximum 519 entries (cursor values) from the index:

500 index-only scans for populating the arrays for each project. The cursor values of the first record is here.
Maximum 19 additional index-only scans for the consecutive records.

The optimized IN query sorts the array (cursor values per project array) 20 times, which means we sort 20 x 500 rows. However, this might be a less memory-intensive task than sorting 10 000 rows at once.

Performance comparison for the gitlab-org group:

Query	Number of 8K Buffers involved	Uncached execution time	Cached execution time
`IN` query	240833	1.2s	660ms
Optimized `IN` query	9783	450ms	22ms

NOTE: Before taking measurements, the group lookup query was executed separately to make the group data available in the buffer cache. Since it's a frequently called query, it hits many shared buffers during the query execution in the production environment.

Efficient IN operator queries