SQL Query Guidelines

This document describes various guidelines to follow when writing SQL queries, either using ActiveRecord/Arel or raw SQL queries.

Using LIKE Statements

The most common way to search for data is using the LIKE statement. For example, to get all issues with a title starting with “Draft:” you’d write the following query:

SELECT *
FROM issues
WHERE title LIKE 'Draft:%';

On PostgreSQL the LIKE statement is case-sensitive. To perform a case-insensitive LIKE you have to use ILIKE instead.

To handle this automatically you should use LIKE queries using Arel instead of raw SQL fragments, as Arel automatically uses ILIKE on PostgreSQL.

Issue.where('title LIKE ?', 'Draft:%')

You’d write this instead:

Issue.where(Issue.arel_table[:title].matches('Draft:%'))

Here matches generates the correct LIKE / ILIKE statement depending on the database being used.

If you need to chain multiple OR conditions you can also do this using Arel:

table = Issue.arel_table

Issue.where(table[:title].matches('Draft:%').or(table[:foo].matches('Draft:%')))

On PostgreSQL, this produces:

SELECT *
FROM issues
WHERE (title ILIKE 'Draft:%' OR foo ILIKE 'Draft:%')

LIKE & Indexes

PostgreSQL does not use any indexes when using LIKE / ILIKE with a wildcard at the start. For example, this does not use any indexes:

SELECT *
FROM issues
WHERE title ILIKE '%Draft:%';

Because the value for ILIKE starts with a wildcard the database is not able to use an index as it doesn’t know where to start scanning the indexes.

Luckily, PostgreSQL does provide a solution: trigram Generalized Inverted Index (GIN) indexes. These indexes can be created as follows:

CREATE INDEX [CONCURRENTLY] index_name_here
ON table_name
USING GIN(column_name gin_trgm_ops);

The key here is the GIN(column_name gin_trgm_ops) part. This creates a GIN index with the operator class set to gin_trgm_ops. These indexes can be used by ILIKE / LIKE and can lead to greatly improved performance. One downside of these indexes is that they can easily get quite large (depending on the amount of data indexed).

To keep naming of these indexes consistent, use the following naming pattern:

index_TABLE_on_COLUMN_trigram

For example, a GIN/trigram index for issues.title would be called index_issues_on_title_trigram.

Due to these indexes taking quite some time to be built they should be built concurrently. This can be done by using CREATE INDEX CONCURRENTLY instead of just CREATE INDEX. Concurrent indexes can not be created inside a transaction. Transactions for migrations can be disabled using the following pattern:

class MigrationName < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!
end

For example:

class AddUsersLowerUsernameEmailIndexes < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!

  def up
    execute 'CREATE INDEX CONCURRENTLY index_on_users_lower_username ON users (LOWER(username));'
    execute 'CREATE INDEX CONCURRENTLY index_on_users_lower_email ON users (LOWER(email));'
  end

  def down
    remove_index :users, :index_on_users_lower_username
    remove_index :users, :index_on_users_lower_email
  end
end

Reliably referencing database columns

ActiveRecord by default returns all columns from the queried database table. In some cases the returned rows might need to be customized, for example:

  • Specify only a few columns to reduce the amount of data returned from the database.
  • Include columns from JOIN relations.
  • Perform calculations (SUM, COUNT).

In this example we specify the columns, but not their tables:

  • path from the projects table
  • user_id from the merge_requests table

The query:

# bad, avoid
Project.select("path, user_id").joins(:merge_requests) # SELECT path, user_id FROM "projects" ...

Later on, a new feature adds an extra column to the projects table: user_id. During deployment there might be a short time window where the database migration is already executed, but the new version of the application code is not deployed yet. When the query mentioned above executes during this period, the query fails with the following error message: PG::AmbiguousColumn: ERROR: column reference "user_id" is ambiguous

The problem is caused by the way the attributes are selected from the database. The user_id column is present in both the users and merge_requests tables. The query planner cannot decide which table to use when looking up the user_id column.

When writing a customized SELECT statement, it’s better to explicitly specify the columns with the table name.

Good (prefer)

Project.select(:path, 'merge_requests.user_id').joins(:merge_requests)

# SELECT "projects"."path", merge_requests.user_id as user_id FROM "projects" ...
Project.select(:path, :'merge_requests.user_id').joins(:merge_requests)

# SELECT "projects"."path", "merge_requests"."id" as user_id FROM "projects" ...

Example using Arel (arel_table):

Project.select(:path, MergeRequest.arel_table[:user_id]).joins(:merge_requests)

# SELECT "projects"."path", "merge_requests"."user_id" FROM "projects" ...

When writing raw SQL query:

SELECT projects.path, merge_requests.user_id FROM "projects"...

When the raw SQL query is parameterized (needs escaping):

include ActiveRecord::ConnectionAdapters::Quoting

"""
SELECT
  #{quote_table_name('projects')}.#{quote_column_name('path')},
  #{quote_table_name('merge_requests')}.#{quote_column_name('user_id')}
FROM ...
"""

Bad (avoid)

Project.select('id, path, user_id').joins(:merge_requests).to_sql

# SELECT id, path, user_id FROM "projects" ...
Project.select("path", "user_id").joins(:merge_requests)
# SELECT "projects"."path", "user_id" FROM "projects" ...

# or

Project.select(:path, :user_id).joins(:merge_requests)
# SELECT "projects"."path", "user_id" FROM "projects" ...

When a column list is given, ActiveRecord tries to match the arguments against the columns defined in the projects table and prepend the table name automatically. In this case, the id column is not a problem, but the user_id column could return unexpected data:

Project.select(:id, :user_id).joins(:merge_requests)

# Before deployment (user_id is taken from the merge_requests table):
# SELECT "projects"."id", "user_id" FROM "projects" ...

# After deployment (user_id is taken from the projects table):
# SELECT "projects"."id", "projects"."user_id" FROM "projects" ...

Plucking IDs

Never use ActiveRecord’s pluck to pluck a set of values into memory only to use them as an argument for another query. For example, this executes an extra unnecessary database query and load a lot of unnecessary data into memory:

projects = Project.all.pluck(:id)

MergeRequest.where(source_project_id: projects)

Instead you can just use sub-queries which perform far better:

MergeRequest.where(source_project_id: Project.all.select(:id))

The only time you should use pluck is when you actually need to operate on the values in Ruby itself (for example, writing them to a file). In almost all other cases you should ask yourself “Can I not just use a sub-query?”.

In line with our CodeReuse/ActiveRecord cop, you should only use forms like pluck(:id) or pluck(:user_id) within model code. In the former case, you can use the ApplicationRecord-provided .pluck_primary_key helper method instead. In the latter, you should add a small helper method to the relevant model.

If you have strong reasons to use pluck, it could make sense to limit the number of records plucked. MAX_PLUCK defaults to 1_000 in ApplicationRecord.

Inherit from ApplicationRecord

Most models in the GitLab codebase should inherit from ApplicationRecord or Ci::ApplicationRecord rather than from ActiveRecord::Base. This allows helper methods to be easily added.

An exception to this rule exists for models created in database migrations. As these should be isolated from application code, they should continue to subclass from MigrationRecord which is available only in migration context.

Use UNIONs

UNIONs aren’t very commonly used in most Rails applications but they’re very powerful and useful. Queries tend to use a lot of JOINs to get related data or data based on certain criteria, but JOIN performance can quickly deteriorate as the data involved grows.

For example, if you want to get a list of projects where the name contains a value or the name of the namespace contains a value most people would write the following query:

SELECT *
FROM projects
JOIN namespaces ON namespaces.id = projects.namespace_id
WHERE projects.name ILIKE '%gitlab%'
OR namespaces.name ILIKE '%gitlab%';

Using a large database this query can easily take around 800 milliseconds to run. Using a UNION we’d write the following instead:

SELECT projects.*
FROM projects
WHERE projects.name ILIKE '%gitlab%'

UNION

SELECT projects.*
FROM projects
JOIN namespaces ON namespaces.id = projects.namespace_id
WHERE namespaces.name ILIKE '%gitlab%';

This query in turn only takes around 15 milliseconds to complete while returning the exact same records.

This doesn’t mean you should start using UNIONs everywhere, but it’s something to keep in mind when using lots of JOINs in a query and filtering out records based on the joined data.

GitLab comes with a Gitlab::SQL::Union class that can be used to build a UNION of multiple ActiveRecord::Relation objects. You can use this class as follows:

union = Gitlab::SQL::Union.new([projects, more_projects, ...])

Project.from("(#{union.to_sql}) projects")

The FromUnion model concern provides a more convenient method to produce the same result as above:

class Project
  include FromUnion
  ...
end

Project.from_union(projects, more_projects, ...)

UNION is common through the codebase, but it’s also possible to use the other SQL set operators of EXCEPT and INTERSECT:

class Project
  include FromIntersect
  include FromExcept
  ...
end

intersected = Project.from_intersect(all_projects, project_set_1, project_set_2)
excepted = Project.from_except(all_projects, project_set_1, project_set_2)

Uneven columns in the UNION sub-queries

When the UNION query has uneven columns in the SELECT clauses, the database returns an error. Consider the following UNION query:

SELECT id FROM users WHERE id = 1
UNION
SELECT id, name FROM users WHERE id = 2
end

The query results in the following error message:

each UNION query must have the same number of columns

This problem is apparent and it can be easily fixed during development. One edge-case is when UNION queries are combined with explicit column listing where the list comes from the ActiveRecord schema cache.

Example (bad, avoid it):

scope1 = User.select(User.column_names).where(id: [1, 2, 3]) # selects the columns explicitly
scope2 = User.where(id: [10, 11, 12]) # uses SELECT users.*

User.connection.execute(Gitlab::SQL::Union.new([scope1, scope2]).to_sql)

When this code is deployed, it doesn’t cause problems immediately. When another developer adds a new database column to the users table, this query breaks in production and can cause downtime. The second query (SELECT users.*) includes the newly added column; however, the first query does not. The column_names method returns stale values (the new column is missing), because the values are cached within the ActiveRecord schema cache. These values are usually populated when the application boots up.

At this point, the only fix would be a full application restart so that the schema cache gets updated. Since GitLab 16.1, the schema cache will be automatically reset so that subsequent queries will succeed. This reset can be disabled by disabling the ops feature flag reset_column_information_on_statement_invalid.

The problem can be avoided if we always use SELECT users.* or we always explicitly define the columns.

Using SELECT users.*:

# Bad, avoid it
scope1 = User.select(User.column_names).where(id: [1, 2, 3])
scope2 = User.where(id: [10, 11, 12])

# Good, both queries generate SELECT users.*
scope1 = User.where(id: [1, 2, 3])
scope2 = User.where(id: [10, 11, 12])

User.connection.execute(Gitlab::SQL::Union.new([scope1, scope2]).to_sql)

Explicit column list definition:

# Good, the SELECT columns are consistent
columns = User.cached_column_list # The helper returns fully qualified (table.column) column names (Arel)
scope1 = User.select(*columns).where(id: [1, 2, 3]) # selects the columns explicitly
scope2 = User.select(*columns).where(id: [10, 11, 12]) # uses SELECT users.*

User.connection.execute(Gitlab::SQL::Union.new([scope1, scope2]).to_sql)

Ordering by Creation Date

When ordering records based on the time they were created, you can order by the id column instead of ordering by created_at. Because IDs are always unique and incremented in the order that rows are created, doing so produces the exact same results. This also means there’s no need to add an index on created_at to ensure consistent performance as id is already indexed by default.

Use WHERE EXISTS instead of WHERE IN

While WHERE IN and WHERE EXISTS can be used to produce the same data it is recommended to use WHERE EXISTS whenever possible. While in many cases PostgreSQL can optimize WHERE IN quite well there are also many cases where WHERE EXISTS performs (much) better.

In Rails you have to use this by creating SQL fragments:

Project.where('EXISTS (?)', User.select(1).where('projects.creator_id = users.id AND users.foo = X'))

This would then produce a query along the lines of the following:

SELECT *
FROM projects
WHERE EXISTS (
    SELECT 1
    FROM users
    WHERE projects.creator_id = users.id
    AND users.foo = X
)

.find_or_create_by is not atomic

The inherent pattern with methods like .find_or_create_by and .first_or_create and others is that they are not atomic. This means, it first runs a SELECT, and if there are no results an INSERT is performed. With concurrent processes in mind, there is a race condition which may lead to trying to insert two similar records. This may not be desired, or may cause one of the queries to fail due to a constraint violation, for example.

Using transactions does not solve this problem.

To solve this we’ve added the ApplicationRecord.safe_find_or_create_by.

This method can be used the same way as find_or_create_by, but it wraps the call in a new transaction (or a subtransaction) and retries if it were to fail because of an ActiveRecord::RecordNotUnique error.

To be able to use this method, make sure the model you want to use this on inherits from ApplicationRecord.

In Rails 6 and later, there is a .create_or_find_by method. This method differs from our .safe_find_or_create_by methods because it performs the INSERT, and then performs the SELECT commands only if that call fails.

If the INSERT fails, it leaves a dead tuple around and increment the primary key sequence (if any), among other downsides.

We prefer .safe_find_or_create_by if the common path is that we have a single record which is reused after it has first been created. However, if the more common path is to create a new record, and we only want to avoid duplicate records to be inserted on edge cases (for example a job-retry), then .create_or_find_by can save us a SELECT.

Both methods use subtransactions internally if executed within the context of an existing transaction. This can significantly impact overall performance, especially if more than 64 live subtransactions are being used inside a single transaction.

Can I use .safe_find_or_create_by?

If your code is generally isolated (for example it’s executed in a worker only) and not wrapped with another transaction, then you can use .safe_find_or_create_by. However, there is no tooling to catch cases when someone else calls your code within a transaction. Using .safe_find_or_create_by will definitely carry some risks that cannot be eliminated completely at the moment.

Additionally, we have a RuboCop rule Performance/ActiveRecordSubtransactionMethods that prevents the usage of .safe_find_or_create_by. This rule can be disabled on a case by case basis via # rubocop:disable Performance/ActiveRecordSubtransactionMethods.

Alternative 1: UPSERT

The .upsert method can be an alternative solution when the table is backed by a unique index.

Simple usage of the .upsert method:

BuildTrace.upsert(
  {
    build_id: build_id,
    title: title
  },
  unique_by: :build_id
)

A few things to be careful about:

  • The sequence for the primary key will be incremented, even if the record was only updated.
  • The created record is not returned. The returning option only returns data when an INSERT happens (new record).
  • ActiveRecord validations are not executed.

An example of the .upsert method with validations and record loading:

params = {
  build_id: build_id,
  title: title
}

build_trace = BuildTrace.new(params)

unless build_trace.valid?
  raise 'notify the user here'
end

BuildTrace.upsert(params, unique_by: :build_id)

build_trace = BuildTrace.find_by!(build_id: build_id)

# do something with build_trace here

The code snippet above will not work well if there is a model-level uniqueness validation on the build_id column because we invoke the validation before calling .upsert.

To work around this, we have two options:

  • Remove the uniqueness validation from the ActiveRecord model.
  • Use the on keyword and implement context-specific validation.

Alternative 2: Check existence and rescue

When the chance of concurrently creating the same record is very low, we can use a simpler approach:

def my_create_method
  params = {
    build_id: build_id,
    title: title
  }

  build_trace = BuildTrace
    .where(build_id: params[:build_id])
    .first

  build_trace = BuildTrace.new(params) if build_trace.blank?

  build_trace.update!(params)

rescue ActiveRecord::RecordInvalid => invalid
  retry if invalid.record&.errors&.of_kind?(:build_id, :taken)
end

The method does the following:

  1. Look up the model by the unique column.
  2. If no record found, build a new one.
  3. Persist the record.

There is a short race condition between the lookup query and the persist query where another process could insert the record and cause an ActiveRecord::RecordInvalid exception.

The code rescues this particular exception and retries the operation. For the second run, the record would be successfully located. For example check this block of code in PreventApprovalByAuthorService.

Monitor SQL queries in production

GitLab team members can monitor slow or canceled queries on GitLab.com using the PostgreSQL logs, which are indexed in Elasticsearch and searchable using Kibana.

See the runbook for more details.

When to use common table expressions

You can use common table expressions (CTEs) to create a temporary result set within a more complex query. You can also use a recursive CTE to reference the CTE’s result set within the query itself. The following example queries a chain of personal access tokens referencing each other in the previous_personal_access_token_id column.

WITH RECURSIVE "personal_access_tokens_cte" AS (
(
    SELECT
      "personal_access_tokens".*
    FROM
      "personal_access_tokens"
    WHERE
      "personal_access_tokens"."previous_personal_access_token_id" = 15)
  UNION (
    SELECT
      "personal_access_tokens".*
    FROM
      "personal_access_tokens",
      "personal_access_tokens_cte"
    WHERE
      "personal_access_tokens"."previous_personal_access_token_id" = "personal_access_tokens_cte"."id"))
SELECT
  "personal_access_tokens".*
FROM
  "personal_access_tokens_cte" AS "personal_access_tokens"

 id | previous_personal_access_token_id
----+-----------------------------------
 16 |                                15
 17 |                                16
 18 |                                17
 19 |                                18
 20 |                                19
 21 |                                20
(6 rows)

As CTEs are temporary result sets, you can use them within another SELECT statement. Using CTEs with UPDATE, or DELETE could lead to unexpected behavior:

Consider the following method:

def personal_access_token_chain(token)
  cte = Gitlab::SQL::RecursiveCTE.new(:personal_access_tokens_cte)
  personal_access_token_table = Arel::Table.new(:personal_access_tokens)

  cte << PersonalAccessToken
           .where(personal_access_token_table[:previous_personal_access_token_id].eq(token.id))
  cte << PersonalAccessToken
           .from([personal_access_token_table, cte.table])
           .where(personal_access_token_table[:previous_personal_access_token_id].eq(cte.table[:id]))
  PersonalAccessToken.with.recursive(cte.to_arel).from(cte.alias_to(personal_access_token_table))
end

It works as expected when it is used to query data:

> personal_access_token_chain(token)

WITH RECURSIVE "personal_access_tokens_cte" AS (
(
    SELECT
      "personal_access_tokens".*
    FROM
      "personal_access_tokens"
    WHERE
      "personal_access_tokens"."previous_personal_access_token_id" = 11)
  UNION (
    SELECT
      "personal_access_tokens".*
    FROM
      "personal_access_tokens",
      "personal_access_tokens_cte"
    WHERE
      "personal_access_tokens"."previous_personal_access_token_id" = "personal_access_tokens_cte"."id"))
SELECT
    "personal_access_tokens".*
FROM
    "personal_access_tokens_cte" AS "personal_access_tokens"

However, the CTE is dropped when used with #update_all. As a result, the method updates the entire table:

> personal_access_token_chain(token).update_all(revoked: true)

UPDATE
    "personal_access_tokens"
SET
    "revoked" = TRUE

To work around this behavior:

  1. Query the ids of the records:

    > token_ids = personal_access_token_chain(token).pluck_primary_key
    => [16, 17, 18, 19, 20, 21]
    
  2. Use this array to scope PersonalAccessTokens:

    PersonalAccessToken.where(id: token_ids).update_all(revoked: true)
    

Alternatively, combine these two steps:

PersonalAccessToken
  .where(id: personal_access_token_chain(token).pluck_primary_key)
  .update_all(revoked: true)
note
Avoid updating large volumes of unbounded data. If there are no application limits on the data, or you are unsure about the data volume, you should update the data in batches.