Skip to content

Migration Style Guide

When writing migrations for GitLab, you have to take into account that these are run by hundreds of thousands of organizations of all sizes, some with many years of data in their database.

In addition, having to take a server offline for an upgrade small or big is a big burden for most organizations. For this reason, it is important that your migrations are written carefully, can be applied online, and adhere to the style guide below.

Migrations are not allowed to require GitLab installations to be taken offline ever. Migrations always must be written in such a way to avoid downtime. In the past we had a process for defining migrations that allowed for downtime by setting a DOWNTIME constant. You may see this when looking at older migrations. This process was in place for 4 years without ever being used and as such we've learned we can always figure out how to write a migration differently to avoid downtime.

When writing your migrations, also consider that databases might have stale data or inconsistencies and guard for that. Try to make as few assumptions as possible about the state of the database.

Don't depend on GitLab-specific code since it can change in future versions. If needed copy-paste GitLab code into the migration to make it forward compatible.

Choose an appropriate migration type

The first step before adding a new migration should be to decide which type is most appropriate.

There are currently three kinds of migrations you can create, depending on the kind of work it needs to perform and how long it takes to complete:

  1. Regular schema migrations. These are traditional Rails migrations in db/migrate that run before new application code is deployed (for GitLab.com before Canary is deployed). This means that they should be relatively fast, no more than a few minutes, so as not to unnecessarily delay a deployment.

    One exception is a migration that takes longer but is absolutely critical for the application to operate correctly. For example, you might have indices that enforce unique tuples, or that are needed for query performance in critical parts of the application. In cases where the migration would be unacceptably slow, however, a better option might be to guard the feature with a feature flag and perform a post-deployment migration instead. The feature can then be turned on after the migration finishes.

    Migrations used to add new models are also part of these regular schema migrations. The only differences are the Rails command used to generate the migrations and the additional generated files, one for the model and one for the model's spec.

  2. Post-deployment migrations. These are Rails migrations in db/post_migrate and are run independently from the GitLab.com deployments. Pending post migrations are executed on a daily basis at the discretion of release manager through the post-deploy migration pipeline. These migrations can be used for schema changes that aren't critical for the application to operate, or data migrations that take at most a few minutes. Common examples for schema changes that should run post-deploy include:

    • Clean-ups, like removing unused columns.
    • Adding non-critical indices on high-traffic tables.
    • Adding non-critical indices that take a long time to create.

    These migrations should not be used for schema changes that are critical for the application to operate. Making such schema changes in a post-deployment migration have caused issues in the past, for example this issue. Changes that should always be a regular schema migration and not be executed in a post-deployment migration include:

    • Creating a new table, example: create_table.
    • Adding a new column to an existing table, example: add_column.

    NOTE: Post-deployment migration is often abbreviated as PDM.

  3. Batched background migrations. These aren't regular Rails migrations, but application code that is executed via Sidekiq jobs, although a post-deployment migration is used to schedule them. Use them only for data migrations that exceed the timing guidelines for post-deploy migrations. Batched background migrations should not change the schema.

Use the following diagram to guide your decision, but keep in mind that it is just a tool, and the final outcome will always be dependent on the specific changes being made:

graph LR
    A{Schema<br/>changed?}
    A -->|Yes| C{Critical to<br/>speed or<br/>behavior?}
    A -->|No| D{Is it fast?}

    C -->|Yes| H{Is it fast?}
    C -->|No| F[Post-deploy migration]

    H -->|Yes| E[Regular migration]
    H -->|No| I[Post-deploy migration<br/>+ feature flag]

    D -->|Yes| F[Post-deploy migration]
    D -->|No| G[Background migration]

Also refer to Migration type to use when choosing which migration type to use when adding a database index.

How long a migration should take

In general, all migrations for a single deploy shouldn't take longer than 1 hour for GitLab.com. The following guidelines are not hard rules, they were estimated to keep migration duration to a minimum.

NOTE: Keep in mind that all durations should be measured against GitLab.com.

NOTE: The result of a database migration pipeline includes the timing information for migrations.

Migration Type Recommended Duration Notes
Regular migrations <= 3 minutes A valid exception are changes without which application functionality or performance would be severely degraded and which cannot be delayed.
Post-deployment migrations <= 10 minutes A valid exception are schema changes, since they must not happen in background migrations.
Background migrations > 10 minutes Since these are suitable for larger tables, it's not possible to set a precise timing guideline, however, any single query must stay below 1 second execution time with cold caches.

Decide which database to target

GitLab connects to two different Postgres databases: main and ci. This split can affect migrations as they may run on either or both of these databases.

Read Migrations for Multiple databases to understand if or how a migration you add should account for this.

Create a regular schema migration

To create a migration you can use the following Rails generator:

bundle exec rails g migration migration_name_here

This generates the migration file in db/migrate.

Regular schema migrations to add new models

To create a new model you can use the following Rails generator:

bundle exec rails g model model_name_here

This will generate:

  • the migration file in db/migrate
  • the model file in app/models
  • the spec file in spec/models

Schema Changes

Changes to the schema should be committed to db/structure.sql. This file is automatically generated by Rails when you run bundle exec rails db:migrate, so you typically should not edit this file by hand. If your migration is adding a column to a table, that column is added at the bottom. Do not reorder columns manually for existing tables as this causes confusion to other people using db/structure.sql generated by Rails.

NOTE: Creating an index asynchronously requires two merge requests. When done, commit the schema change in the merge request that adds the index with add_concurrent_index.

When your local database in your GDK is diverging from the schema from main it might be hard to cleanly commit the schema changes to Git. In that case you can use the scripts/regenerate-schema script to regenerate a clean db/structure.sql for the migrations you're adding. This script applies all migrations found in db/migrate or db/post_migrate, so if there are any migrations you don't want to commit to the schema, rename or remove them. If your branch is not targeting the default Git branch, you can set the TARGET environment variable.

# Regenerate schema against `main`
scripts/regenerate-schema

# Regenerate schema against `12-9-stable-ee`
TARGET=12-9-stable-ee scripts/regenerate-schema

The scripts/regenerate-schema script can create additional differences. If this happens, use a manual procedure where <migration ID> is the DATETIME part of the migration file.

# Rebase against master
git rebase master

# Rollback changes
VERSION=<migration ID> bundle exec rails db:rollback:main

# Checkout db/structure.sql from master
git checkout origin/master db/structure.sql

# Migrate changes
VERSION=<migration ID> bundle exec rails db:migrate:main

After a table has been created, it should be added to the database dictionary, following the steps mentioned in the database dictionary guide.

Migration checksum file

When a migration is first executed, a new migration checksum file is created in db/schema_migrations containing a SHA256 generated from the migration's timestamp. The name of this new file is the same as the timestamp portion of the migration filename, for example db/schema_migrations/20241021120146. The content of this file is the SHA256 of the timestamp portion, for example:

$ echo -n "20241021120146" | sha256sum
7a3e382a6e5564bfa7004bca1a357a910b151e7399c6466113daf01526d97470  -

The SHA256 adds unique content to the file so Git rename detection sees them as separate files.

This migration checksum file indicates that the migration executed successfully and the result recorded in db/structure.sql. The presence of this file prevents the same migration from being executed twice, and therefore, it's necessary to include this file in the merge request that adds the new migration.

See Development change: Database schema version handling outside of structure.sql for more details about the db/schema_migrations directory.

Keeping the migration checksum file up-to-date

  • when a new migration is created, run rake db:migrate to execute the migration and generate the corresponding db/schema_migration/<timestamp> checksum file, and add this file into version control.
  • if the migration is deleted, remove the corresponding db/schema_migration/<timestamp> checksum file.
  • if the timestamp portion of the migration is changed, remove the corresponding db/schema_migration/<timestamp> checksum file and run rake db:migrate to generate a new one, and add this file into version control.
  • if the content of the migration is changed, no changes are required to the db/schema_migration/<timestamp> checksum file.

Avoiding downtime

The document "Avoiding downtime in migrations" specifies various database operations, such as:

and explains how to perform them without requiring downtime.

Reversibility

Your migration must be reversible. This is very important, as it should be possible to downgrade in case of a vulnerability or bugs.

Note: On GitLab production environments, if a problem occurs, a roll-forward strategy is used instead of rolling back migrations using db:rollback. On self-managed instances we advise users to restore the backup which was created before the upgrade process started. The down method is used primarily in the development environment, for example, when a developer wants to ensure their local copy of structure.sql file and database are in a consistent state when switching between commits or branches.

In your migration, add a comment describing how the reversibility of the migration was tested.

Some migrations cannot be reversed. For example, some data migrations can't be reversed because we lose information about the state of the database before the migration. You should still create a down method with a comment, explaining why the changes performed by the up method can't be reversed, so that the migration itself can be reversed, even if the changes performed during the migration can't be reversed:

def down
  # no-op

  # comment explaining why changes performed by `up` cannot be reversed.
end

Migrations like this are inherently risky and additional actions are required when preparing the migration for review.

Atomicity and transaction

By default, migrations are a single transaction: it's opened at the beginning of the migration, and committed after all steps are processed.

Running migrations in a single transaction makes sure that if one of the steps fails, none of the steps are executed, leaving the database in a valid state. Therefore, either:

  • Put all migrations in one single-transaction migration.
  • If necessary, put most actions in one migration and create a separate migration for the steps that cannot be done in a single transaction.

For example, if you create an empty table and need to build an index for it, you should use a regular single-transaction migration and the default rails schema statement: add_index. This operation is a blocking operation, but it doesn't cause problems because the table is not yet used, and therefore it does not have any records yet.

NOTE: Subtransactions are disallowed in general. Use multiple, separate transactions if needed as described in Heavy operations in a single transaction.

Heavy operations in a single transaction

When using a single-transaction migration, a transaction holds a database connection for the duration of the migration, so you must make sure the actions in the migration do not take too much time. In general, transactions must execute quickly. To that end, observe the maximum query time limit for each query run in the migration.

If your single-transaction migration takes long to finish, you have several options. In all cases, remember to select the appropriate migration type depending on how long a migration takes

  • Split the migration into multiple single-transaction migrations.

  • Use multiple transactions by using disable_ddl_transaction!.

  • Keep using a single-transaction migration after adjusting statement and lock timeout settings. If your heavy workload must use the guarantees of a transaction, you should check your migration can execute without hitting the timeout limits. The same advice applies to both single-transaction migrations and individual transactions.

    • Statement timeout: the statement timeout is configured to be 15s for GitLab.com's production database but creating an index often takes more than 15 seconds. When you use the existing helpers including add_concurrent_index, they automatically turn off the statement timeout as needed. In rare cases, you might need to set the timeout limit yourself by using disable_statement_timeout.

NOTE: To run migrations, we directly connect to the primary database, bypassing PgBouncer to control settings like statement_timeout and lock_wait_timeout.

Temporarily turn off the statement timeout limit

The migration helper disable_statement_timeout enables you to temporarily set the statement timeout to 0 per transaction or per connection.

  • You use the per-connection option when your statement does not support running inside an explicit transaction, like CREATE INDEX CONCURRENTLY.

  • If your statement does support an explicit transaction block, like ALTER TABLE ... VALIDATE CONSTRAINT, the per-transaction option should be used.

Using disable_statement_timeout is rarely needed, because the most migration helpers already use them internally when needed. For example, creating an index usually takes more than 15 seconds, which is the default statement timeout configured for GitLab.com's production database. The helper add_concurrent_index creates an index inside the block passed to disable_statement_timeout to disable the statement timeout per connection.

If you are writing raw SQL statements in a migration, you may need to manually use disable_statement_timeout. Consult the database reviewers and maintainers when you do.

Disable transaction-wrapped migration

You can opt out of running your migration as a single transaction by using disable_ddl_transaction!, an ActiveRecord method. The method might be called in other database systems, with different results. At GitLab we exclusively use PostgreSQL. You should always read disable_ddl_transaction! as meaning:

"Do not execute this migration in a single PostgreSQL transaction. I'll open PostgreSQL transaction(s) only when and if I need them."

NOTE: Even if you don't use an explicit PostgreSQL transaction .transaction (or BEGIN; COMMIT;), every SQL statement is still executed as a transaction. See the PostgreSQL documentation on transactions.

NOTE: In GitLab, we've sometimes referred to the migrations that used disable_ddl_transaction! as non-transactional migrations. It just meant the migrations were not executed as single transactions.

When should you use disable_ddl_transaction!? In most cases, the existing RuboCop rules or migration helpers can detect if you should be using disable_ddl_transaction!. Skip disable_ddl_transaction! if you are unsure whether to use it or not in your migration, and let the RuboCop rules and database reviews guide you.

Use disable_ddl_transaction! when PostgreSQL requires an operation to be executed outside an explicit transaction.

  • The most prominent example of such operation is the command CREATE INDEX CONCURRENTLY. PostgreSQL allows the blocking version (CREATE INDEX) to be run inside a transaction. Unlike CREATE INDEX, CREATE INDEX CONCURRENTLY must be performed outside a transaction. Therefore, even though a migration may run just one statement CREATE INDEX CONCURRENTLY, you should disable disable_ddl_transaction!. It's also the reason why the use of the helper add_concurrent_index requires disable_ddl_transaction! CREATE INDEX CONCURRENTLY is more of the exception than the rule.

Use disable_ddl_transaction! when you need to run multiple transactions in a migration for any reason. Most of the time you would be using multiple transactions to avoid running one slow transaction.

  • For example, when you insert, update, or delete (DML) a large amount of data, you should perform them in batches. Should you need to group operations for each batch, you can explicitly open a transaction block when processing a batch. Consider using a batched background migration for any reasonably large workload.

Use disable_ddl_transaction! when migration helpers require them. Various migration helpers need to run with disable_ddl_transaction! because they require a precise control on when and how to open transactions.

  • A foreign key can be added inside a transaction, unlike CREATE INDEX CONCURRENTLY. However, PostgreSQL does not provide an option similar to CREATE INDEX CONCURRENTLY. The helper add_concurrent_foreign_key instead opens its own transactions to lock the source and target table in a manner that minimizes locking while adding and validating the foreign key.
  • As advised earlier, skip disable_ddl_transaction! if you are unsure and see if any RuboCop check is violated.

Use disable_ddl_transaction! when your migration does not actually touch PostgreSQL databases or does touch multiple PostgreSQL databases.

  • For example, your migration might target a Redis server. As a rule, you cannot interact with an external service inside a PostgreSQL transaction.
  • A transaction is used for a single database connection. If your migrations are targeting multiple databases, such as both ci and main database, follow Migrations for multiple databases.

Naming conventions

Names for database objects (such as tables, indexes, and views) must be lowercase. Lowercase names ensure that queries with unquoted names don't cause errors.

We keep column names consistent with ActiveRecord's schema conventions.

Custom index and constraint names should follow the constraint naming convention guidelines.

Truncate long index names

PostgreSQL limits the length of identifiers, like column or index names. Column names are not usually a problem, but index names tend to be longer. Some methods for shortening a name that's too long:

  • Prefix it with i_ instead of index_.
  • Skip redundant prefixes. For example, index_vulnerability_findings_remediations_on_vulnerability_remediation_id becomes index_vulnerability_findings_remediations_on_remediation_id.
  • Instead of columns, specify the purpose of the index, such as index_users_for_unconfirmation_notification.

Migration timestamp age

The timestamp portion of a migration filename determines the order in which migrations are run. It's important to maintain a rough correlation between:

  1. When a migration is added to the GitLab codebase.
  2. The timestamp of the migration itself.

A new migration's timestamp should never be before the previous required upgrade stop. Migrations are occasionally squashed, and if a migration is added whose timestamp falls before the previous required stop, a problem like what happened in issue 408304 can occur.

For example, if we are currently developing against GitLab 16.0, the previous required stop is 15.11. 15.11 was released on April 23rd, 2023. Therefore, the minimum acceptable timestamp would be 20230424000000.

Best practice

While the above should be considered a hard rule, it is a best practice to try to keep migration timestamps to within three weeks of the date it is anticipated that the migration will be merged upstream, regardless of how much time has elapsed since the last required stop.

To update a migration timestamp:

  1. Migrate down the migration for the ci and main databases:

    rake db:migrate:down:main VERSION=<timestamp>
rake db:migrate:down:ci VERSION=<timestamp>

  2. Delete the migration file.

  3. Recreate the migration following the migration style guide.

Migration helpers and versioning

Various helper methods are available for many common patterns in database migrations. Those helpers can be found in Gitlab::Database::MigrationHelpers and related modules.

In order to allow changing a helper's behavior over time, we implement a versioning scheme for migration helpers. This allows us to maintain the behavior of a helper for already existing migrations but change the behavior for any new migrations.

For that purpose, all database migrations should inherit from Gitlab::Database::Migration, which is a "versioned" class. For new migrations, the latest version should be used (which can be looked up in Gitlab::Database::Migration::MIGRATION_CLASSES) to use the latest version of migration helpers.

In this example, we use version 2.1 of the migration class:

class TestMigration < Gitlab::Database::Migration[2.1]
  def change
  end
end

Do not include Gitlab::Database::MigrationHelpers directly into a migration. Instead, use the latest version of Gitlab::Database::Migration, which exposes the latest version of migration helpers automatically.

Retry mechanism when acquiring database locks

When changing the database schema, we use helper methods to invoke DDL (Data Definition Language) statements. In some cases, these DDL statements require a specific database lock.

Example:

def change
  remove_column :users, :full_name, :string
end

Executing this migration requires an exclusive lock on the users table. When the table is concurrently accessed and modified by other processes, acquiring the lock may take a while. The lock request is waiting in a queue and it may also block other queries on the users table once it has been enqueued.

More information about PostgreSQL locks: Explicit Locking

For stability reasons, GitLab.com has a short statement_timeout set. When the migration is invoked, any database query has a fixed time to execute. In a worst-case scenario, the request sits in the lock queue, blocking other queries for the duration of the configured statement timeout, then failing with canceling statement due to statement timeout error.

This problem could cause failed application upgrade processes and even application stability issues, since the table may be inaccessible for a short period of time.

To increase the reliability and stability of database migrations, the GitLab codebase offers a method to retry the operations with different lock_timeout settings and wait time between the attempts. Multiple shorter attempts to acquire the necessary lock allow the database to process other statements.

Lock retries are controlled by two different helpers:

  1. enable_lock_retries!: enabled by default for all transactional migrations.
  2. with_lock_retries: enabled manually for a block within non-transactional migrations.

Transactional migrations

Regular migrations execute the full migration in a transaction. lock-retry mechanism is enabled by default (unless disable_ddl_transaction!).

This leads to the lock timeout being controlled for the migration. Also, it can lead to retrying the full migration if the lock could not be granted within the timeout.

Occasionally a migration may need to acquire multiple locks on different objects. To prevent catalog bloat, ask for all those locks explicitly before performing any DDL. A better strategy is to split the migration, so that we only need to acquire one lock at the time.

Multiple changes on the same table

With the lock-retry methodology enabled, all operations wrap into a single transaction. When you have the lock, you should do as much as possible inside the transaction rather than trying to get another lock later. Be careful about running long database statements within the block. The acquired locks are kept until the transaction (block) finishes and depending on the lock type, it might block other database operations.

def up
  add_column :users, :full_name, :string
  add_column :users, :bio, :string
end

def down
  remove_column :users, :full_name
  remove_column :users, :bio
end

Changing default value for a column

Changing column defaults can cause application downtime if a multi-release process is not followed. See avoiding downtime in migrations for changing column defaults for details.

def up
  change_column_default :merge_requests, :lock_version, from: nil, to: 0
end

def down
  change_column_default :merge_requests, :lock_version, from: 0, to: nil
end

Creating a new table when we have two foreign keys

Only one foreign key should be created per transaction. This is because the addition of a foreign key constraint requires a SHARE ROW EXCLUSIVE lock on the referenced table, and locking multiple tables in the same transaction should be avoided.

For this, we need three migrations:

  1. Creating the table without foreign keys (with the indices).
  2. Add foreign key to the first table.
  3. Add foreign key to the second table.

Creating the table:

def up
  create_table :imports do |t|
    t.bigint :project_id, null: false
    t.bigint :user_id, null: false
    t.string :jid, limit: 255

    t.index :project_id
    t.index :user_id
  end
end

def down
  drop_table :imports
end

Adding foreign key to projects:

We can use the add_concurrent_foreign_key method in this case, as this helper method has the lock retries built into it.

disable_ddl_transaction!

def up
  add_concurrent_foreign_key :imports, :projects, column: :project_id, on_delete: :cascade
end

def down
  with_lock_retries do
    remove_foreign_key :imports, column: :project_id
  end
end

Adding foreign key to users:

disable_ddl_transaction!

def up
  add_concurrent_foreign_key :imports, :users, column: :user_id, on_delete: :cascade
end

def down
  with_lock_retries do
    remove_foreign_key :imports, column: :user_id
  end
end

Usage with non-transactional migrations

Only when we disable transactional migrations using disable_ddl_transaction!, we can use the with_lock_retries helper to guard an individual sequence of steps. It opens a transaction to execute the given block.

A custom RuboCop rule ensures that only allowed methods can be placed within the lock retries block.

disable_ddl_transaction!

def up
  with_lock_retries do
    add_column(:users, :name, :text, if_not_exists: true)
  end

  add_text_limit :users, :name, 255 # Includes constraint validation (full table scan)
end

The RuboCop rule generally allows standard Rails migration methods, listed below. This example causes a RuboCop offense:

disable_ddl_transaction!

def up
  with_lock_retries do
    add_concurrent_index :users, :name
  end
end

When to use the helper method

You can only use the with_lock_retries helper method when the execution is not already inside an open transaction (using PostgreSQL subtransactions is discouraged). It can be used with standard Rails migration helper methods. Calling more than one migration helper is not a problem if they're executed on the same table.

Using the with_lock_retries helper method is advised when a database migration involves one of the high-traffic tables.

Example changes:

  • add_foreign_key / remove_foreign_key
  • add_column / remove_column
  • change_column_default
  • create_table / drop_table

The with_lock_retries method cannot be used within the change method, you must manually define the up and down methods to make the migration reversible.

How the helper method works

  1. Iterate 50 times.
  2. For each iteration, set a pre-configured lock_timeout.
  3. Try to execute the given block. (remove_column).
  4. If LockWaitTimeout error is raised, sleep for the pre-configured sleep_time and retry the block.
  5. If no error is raised, the current iteration has successfully executed the block.

For more information check the Gitlab::Database::WithLockRetries class. The with_lock_retries helper method is implemented in the Gitlab::Database::MigrationHelpers module.

In a worst-case scenario, the method:

  • Executes the block for a maximum of 50 times over 40 minutes.
    • Most of the time is spent in a pre-configured sleep period after each iteration.
  • After the 50th retry, the block is executed without lock_timeout, just like a standard migration invocation.
  • If a lock cannot be acquired, the migration fails with statement timeout error.

The migration might fail if there is a very long running transaction (40+ minutes) accessing the users table.

Lock-retry methodology at the SQL level

In this section, we provide a simplified SQL example that demonstrates the use of lock_timeout. You can follow along by running the given snippets in multiple psql sessions.

When altering a table to add a column, AccessExclusiveLock, which conflicts with most lock types, is required on the table. If the target table is a very busy one, the transaction adding the column may fail to acquire AccessExclusiveLock in a timely fashion.

Suppose a transaction is attempting to insert a row into a table:

-- Transaction 1
BEGIN;
INSERT INTO my_notes (id) VALUES (1);

At this point Transaction 1 acquired RowExclusiveLock on my_notes. Transaction 1 could still execute more statements prior to committing or aborting. There could be other similar, concurrent transactions that touch my_notes.

Suppose a transactional migration is attempting to add a column to the table without using any lock retry helper:

-- Transaction 2
BEGIN;
ALTER TABLE my_notes ADD COLUMN title text;

Transaction 2 is now blocked because it cannot acquire AccessExclusiveLock on my_notes table as Transaction 1 is still executing and holding the RowExclusiveLock on my_notes.

A more pernicious effect is blocking the transactions that would normally not conflict with Transaction 1 because Transaction 2 is queueing to acquire AccessExclusiveLock. In a normal situation, if another transaction attempted to read from and write to the same table my_notes at the same time as Transaction 1, the transaction would go through since the locks needed for reading and writing would not conflict with RowExclusiveLock held by Transaction 1. However, when the request to acquire AccessExclusiveLock is queued, the subsequent requests for conflicting locks on the table would block although they could be executed concurrently alongside Transaction 1.

If we used with_lock_retries, Transaction 2 would instead quickly timeout after failing to acquire the lock within the specified time period and allow other transactions to proceed:

-- Transaction 2 (version with lock timeout)
BEGIN;
SET LOCAL lock_timeout to '100ms'; -- added by the lock retry helper.
ALTER TABLE my_notes ADD COLUMN title text;

The lock retry helper would repeatedly try the same transaction at different time intervals until it succeeded.

SET LOCAL scopes the parameter (lock_timeout) change to the transaction.

Removing indexes

If the table is not empty when removing an index, make sure to use the method remove_concurrent_index instead of the regular remove_index method. The remove_concurrent_index method drops indexes concurrently, so no locking is required, and there is no need for downtime. To use this method, you must disable single-transaction mode by calling the method disable_ddl_transaction! in the body of your migration class like so:

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

  INDEX_NAME = 'index_name'

  def up
    remove_concurrent_index :table_name, :column_name, name: INDEX_NAME
  end
end

You can verify that the index is not being used with Grafana:

sum by (type)(rate(pg_stat_user_indexes_idx_scan{env="gprd", indexrelname="INSERT INDEX NAME HERE"}[30d]))

It is not necessary to check if the index exists prior to removing it, however it is required to specify the name of the index that is being removed. This can be done either by passing the name as an option to the appropriate form of remove_index or remove_concurrent_index, or by using the remove_concurrent_index_by_name method. Explicitly specifying the name is important to ensure the correct index is removed.

For a small table (such as an empty one or one with less than 1,000 records), it is recommended to use remove_index in a single-transaction migration, combining it with other operations that don't require disable_ddl_transaction!.

Disabling an index

Disabling an index is not a safe operation.

Adding indexes

Before adding an index, consider if one is necessary. The Adding Database indexes guide contains more details to help you decide if an index is necessary and provides best practices for adding indexes.

Testing for existence of indexes

If a migration requires conditional logic based on the absence or presence of an index, you must test for existence of that index using its name. This helps avoids problems with how Rails compares index definitions, which can lead to unexpected results.

For more details, review the Adding Database Indexes guide.

Adding foreign-key constraints

When adding a foreign-key constraint to either an existing or a new column also remember to add an index on the column.

This is required for all foreign-keys, for example, to support efficient cascading deleting: when a lot of rows in a table get deleted, the referenced records need to be deleted too. The database has to look for corresponding records in the referenced table. Without an index, this results in a sequential scan on the table, which can take a long time.

Here's an example where we add a new column with a foreign key constraint. Note it includes index: true to create an index for it.

class Migration < Gitlab::Database::Migration[2.1]

  def change
    add_reference :model, :other_model, index: true, foreign_key: { on_delete: :cascade }
  end
end

When adding a foreign-key constraint to an existing column in a non-empty table, we have to employ add_concurrent_foreign_key and add_concurrent_index instead of add_reference.

If you have a new or empty table that doesn't reference a high-traffic table, we recommend that you use add_reference in a single-transaction migration. You can combine it with other operations that don't require disable_ddl_transaction!.

You can read more about adding foreign key constraints to an existing column.

NOT NULL constraints

See the style guide on NOT NULL constraints for more information.

Adding Columns With Default Values

With PostgreSQL 11 being the minimum version in GitLab, adding columns with default values has become much easier and the standard add_column helper should be used in all cases.

Before PostgreSQL 11, adding a column with a default was problematic as it would have caused a full table rewrite.

Removing the column default for non-nullable columns

If you have added a non-nullable column, and used the default value to populate existing data, you need to keep that default value around until at least after the application code is updated. You cannot remove the default value in the same migration, as the migrations run before the model code is updated and models will have an old schema cache, meaning they won't know about this column and won't be able to set it. In this case it's recommended to:

  1. Add the column with default value in a standard migration.
  2. Remove the default in a post-deployment migration.

The post-deployment migration happens after the application restarts, ensuring the new column has been discovered.

Changing the column default

One might think that changing a default column with change_column_default is an expensive and disruptive operation for larger tables, but in reality it's not.

Take the following migration as an example:

class DefaultRequestAccessGroups < Gitlab::Database::Migration[2.1]
  def change
    change_column_default(:namespaces, :request_access_enabled, from: false, to: true)
  end
end

Migration above changes the default column value of one of our largest tables: namespaces. This can be translated to:

ALTER TABLE namespaces
ALTER COLUMN request_access_enabled
SET DEFAULT false

In this particular case, the default value exists and we're just changing the metadata for request_access_enabled column, which does not imply a rewrite of all the existing records in the namespaces table. Only when creating a new column with a default, all the records are going be rewritten.

NOTE: A faster ALTER TABLE ADD COLUMN with a non-null default was introduced on PostgreSQL 11.0, removing the need of rewriting the table when a new column with a default value is added.

For the reasons mentioned above, it's safe to use change_column_default in a single-transaction migration without requiring disable_ddl_transaction!.

Updating an existing column

To update an existing column to a particular value, you can use update_column_in_batches. This splits the updates into batches, so we don't update too many rows at in a single statement.

This updates the column foo in the projects table to 10, where some_column is 'hello':

update_column_in_batches(:projects, :foo, 10) do |table, query|
  query.where(table[:some_column].eq('hello'))
end

If a computed update is needed, the value can be wrapped in Arel.sql, so Arel treats it as an SQL literal. It's also a required deprecation for Rails 6.

The below example is the same as the one above, but the value is set to the product of the bar and baz columns:

update_value = Arel.sql('bar * baz')

update_column_in_batches(:projects, :foo, update_value) do |table, query|
  query.where(table[:some_column].eq('hello'))
end

In the case of update_column_in_batches, it may be acceptable to run on a large table, as long as it is only updating a small subset of the rows in the table, but do not ignore that without validating on the GitLab.com staging environment - or asking someone else to do so for you - beforehand.

Removing a foreign key constraint

When removing a foreign key constraint, we need to acquire a lock on both tables that are related to the foreign key. For tables with heavy write patterns, it's a good idea to use with_lock_retries, otherwise you might fail to acquire a lock in time. You might also run into deadlocks when acquiring a lock, because ordinarily the application writes in parent,child order. However, removing a foreign key acquires the lock in child,parent order. To resolve this, you can explicitly acquire the lock in parent,child, for example:

disable_ddl_transaction!

def up
  with_lock_retries do
    execute('lock table ci_pipelines, ci_builds in access exclusive mode')

    remove_foreign_key :ci_builds, to_table: :ci_pipelines, column: :pipeline_id, on_delete: :cascade, name: 'the_fk_name'
  end
end

def down
  add_concurrent_foreign_key :ci_builds, :ci_pipelines, column: :pipeline_id, on_delete: :cascade, name: 'the_fk_name'
end

Dropping a database table

NOTE: After a table has been dropped, it should be added to the database dictionary, following the steps in the database dictionary guide.

Dropping a database table is uncommon, and the drop_table method provided by Rails is generally considered safe. Before dropping the table, consider the following:

If your table has foreign keys on a high-traffic table (like projects), then the DROP TABLE statement is likely to stall concurrent traffic until it fails with statement timeout error.

Table has no records (feature was never in use) and no foreign keys:

  • Use the drop_table method in your migration.
def change
  drop_table :my_table
end

Table has records but no foreign keys:

  • Remove the application code related to the table, such as models, controllers and services.
  • In a post-deployment migration, use drop_table.

This can all be in a single migration if you're sure the code is not used. If you want to reduce risk slightly, consider putting the migrations into a second merge request after the application changes are merged. This approach provides an opportunity to roll back.

def up
  drop_table :my_table
end

def down
  # create_table ...
end

Table has foreign keys:

  • Remove the application code related to the table, such as models, controllers, and services.
  • In a post-deployment migration, remove the foreign keys using the with_lock_retries helper method. In another subsequent post-deployment migration, use drop_table.

This can all be in a single migration if you're sure the code is not used. If you want to reduce risk slightly, consider putting the migrations into a second merge request after the application changes are merged. This approach provides an opportunity to roll back.

Removing the foreign key on the projects table using a non-transactional migration:

# first migration file
class RemovingForeignKeyMigrationClass < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!

  def up
    with_lock_retries do
      remove_foreign_key :my_table, :projects
    end
  end

  def down
    add_concurrent_foreign_key :my_table, :projects, column: COLUMN_NAME
  end
end

Dropping the table:

# second migration file
class DroppingTableMigrationClass < Gitlab::Database::Migration[2.1]
  def up
    drop_table :my_table
  end

  def down
    # create_table with the same schema but without the removed foreign key ...
  end
end

Dropping a sequence

Dropping a sequence is uncommon, but you can use the drop_sequence method provided by the database team.

Under the hood, it works like this:

Remove a sequence:

  • Remove the default value if the sequence is actually used.
  • Execute DROP SEQUENCE.

Re-add a sequence:

  • Create the sequence, with the possibility of specifying the current value.
  • Change the default value of the column.

A Rails migration example:

class DropSequenceTest < Gitlab::Database::Migration[2.1]
  def up
    drop_sequence(:ci_pipelines_config, :pipeline_id, :ci_pipelines_config_pipeline_id_seq)
  end

  def down
    default_value = Ci::Pipeline.maximum(:id) + 10_000

    add_sequence(:ci_pipelines_config, :pipeline_id, :ci_pipelines_config_pipeline_id_seq, default_value)
  end
end

NOTE: add_sequence should be avoided for columns with foreign keys. Adding sequence to these columns is only allowed in the down method (restore previous schema state).

Truncate a table

Truncating a table is uncommon, but you can use the truncate_tables! method provided by the database team.

Under the hood, it works like this:

  • Finds the gitlab_schema for the tables to be truncated.
  • If the gitlab_schema for the tables is included in the connection's gitlab_schemas, it then executes the TRUNCATE statement.
  • If the gitlab_schema for the tables is not included in the connection's gitlab_schemas, it does nothing.

Swapping primary key

Swapping the primary key is required to partition a table as the partition key must be included in the primary key.

You can use the swap_primary_key method provided by the database team.

Under the hood, it works like this:

  • Drop the primary key constraint.
  • Add the primary key using the index defined beforehand.
class SwapPrimaryKey < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!

  TABLE_NAME = :table_name
  PRIMARY_KEY = :table_name_pkey
  OLD_INDEX_NAME = :old_index_name
  NEW_INDEX_NAME = :new_index_name

  def up
    swap_primary_key(TABLE_NAME, PRIMARY_KEY, NEW_INDEX_NAME)
  end

  def down
    add_concurrent_index(TABLE_NAME, :id, unique: true, name: OLD_INDEX_NAME)
    add_concurrent_index(TABLE_NAME, [:id, :partition_id], unique: true, name: NEW_INDEX_NAME)

    unswap_primary_key(TABLE_NAME, PRIMARY_KEY, OLD_INDEX_NAME)
  end
end

NOTE: Make sure to introduce the new index beforehand in a separate migration in order to swap the primary key.

Integer column type

By default, an integer column can hold up to a 4-byte (32-bit) number. That is a max value of 2,147,483,647. Be aware of this when creating a column that holds file sizes in byte units. If you are tracking file size in bytes, this restricts the maximum file size to just over 2GB.

To allow an integer column to hold up to an 8-byte (64-bit) number, explicitly set the limit to 8-bytes. This allows the column to hold a value up to 9,223,372,036,854,775,807.

Rails migration example:

add_column(:projects, :foo, :integer, default: 10, limit: 8)

Strings and the Text data type

See the text data type style guide for more information.

Timestamp column type

By default, Rails uses the timestamp data type that stores timestamp data without time zone information. The timestamp data type is used by calling either the add_timestamps or the timestamps method.

Also, Rails converts the :datetime data type to the timestamp one.

Example:

# timestamps
create_table :users do |t|
  t.timestamps
end

# add_timestamps
def up
  add_timestamps :users
end

# :datetime
def up
  add_column :users, :last_sign_in, :datetime
end

Instead of using these methods, one should use the following methods to store timestamps with time zones:

  • add_timestamps_with_timezone
  • timestamps_with_timezone
  • datetime_with_timezone

This ensures all timestamps have a time zone specified. This, in turn, means existing timestamps don't suddenly use a different time zone when the system's time zone changes. It also makes it very clear which time zone was used in the first place.

Storing JSON in database

The Rails 5 natively supports JSONB (binary JSON) column type. Example migration adding this column:

class AddOptionsToBuildMetadata < Gitlab::Database::Migration[2.1]
  def change
    add_column :ci_builds_metadata, :config_options, :jsonb
  end
end

By default hash keys will be strings. Optionally you can add a custom data type to provide different access to keys.

class BuildMetadata
  attribute :config_options, :ind_jsonb # for indifferent accesss or :sym_jsonb if you need symbols only as keys.
end

When using a JSONB column, use the JsonSchemaValidator to keep control of the data being inserted over time.

class BuildMetadata
  validates :config_options, json_schema: { filename: 'build_metadata_config_option' }
end

Additionally, you can expose the keys in a JSONB column as ActiveRecord attributes. Do this when you need complex validations, or ActiveRecord change tracking. This feature is provided by the jsonb_accessor gem, and does not replace JsonSchemaValidator.

module Organizations
  class OrganizationSetting < ApplicationRecord
    belongs_to :organization

    validates :settings, json_schema: { filename: "organization_settings" }

    jsonb_accessor :settings,
      restricted_visibility_levels: [:integer, { array: true }]

    validates_each :restricted_visibility_levels do |record, attr, value|
      value&.each do |level|
        unless Gitlab::VisibilityLevel.options.value?(level)
          record.errors.add(attr, format(_("'%{level}' is not a valid visibility level"), level: level))
        end
      end
    end
  end
end

You can now use restricted_visibility_levels as an ActiveRecord attribute:

> s = Organizations::OrganizationSetting.find(1)
=> #<Organizations::OrganizationSetting:0x0000000148d67628>
> s.settings
=> {"restricted_visibility_levels"=>[20]}
> s.restricted_visibility_levels
=> [20]
> s.restricted_visibility_levels = [0]
=> [0]
> s.changes
=> {"settings"=>[{"restricted_visibility_levels"=>[20]}, {"restricted_visibility_levels"=>[0]}], "restricted_visibility_levels"=>[[20], [0]]}

Encrypted attributes

Do not store attr_encrypted attributes as :text in the database; use :binary instead. This uses the bytea type in PostgreSQL and makes storage more efficient:

class AddSecretToSomething < Gitlab::Database::Migration[2.1]
  def change
    add_column :something, :encrypted_secret, :binary
    add_column :something, :encrypted_secret_iv, :binary
  end
end

When storing encrypted attributes in a binary column, we need to provide the encode: false and encode_iv: false options to attr_encrypted:

class Something < ApplicationRecord
  attr_encrypted :secret,
    mode: :per_attribute_iv,
    key: Settings.attr_encrypted_db_key_base_32,
    algorithm: 'aes-256-gcm',
    encode: false,
    encode_iv: false
end

Testing

See the Testing Rails migrations style guide.

Data migration

Prefer Arel and plain SQL over usual ActiveRecord syntax. In case of using plain SQL, you need to quote all input manually with quote_string helper.

Example with Arel:

users = Arel::Table.new(:users)
users.group(users[:user_id]).having(users[:id].count.gt(5))

#update other tables with these results

Example with plain SQL and quote_string helper:

select_all("SELECT name, COUNT(id) as cnt FROM tags GROUP BY name HAVING COUNT(id) > 1").each do |tag|
  tag_name = quote_string(tag["name"])
  duplicate_ids = select_all("SELECT id FROM tags WHERE name = '#{tag_name}'").map{|tag| tag["id"]}
  origin_tag_id = duplicate_ids.first
  duplicate_ids.delete origin_tag_id

  execute("UPDATE taggings SET tag_id = #{origin_tag_id} WHERE tag_id IN(#{duplicate_ids.join(",")})")
  execute("DELETE FROM tags WHERE id IN(#{duplicate_ids.join(",")})")
end

If you need more complex logic, you can define and use models local to a migration. For example:

class MyMigration < Gitlab::Database::Migration[2.1]
  class Project < MigrationRecord
    self.table_name = 'projects'
  end

  def up
    # Reset the column information of all the models that update the database
    # to ensure the Active Record's knowledge of the table structure is current
    Project.reset_column_information

    # ... ...
  end
end

When doing so be sure to explicitly set the model's table name, so it's not derived from the class name or namespace.

Be aware of the limitations when using models in migrations.

Modifying existing data

In most circumstances, prefer migrating data in batches when modifying data in the database.

We introduced a new helper each_batch_range which facilitates the process of iterating over a collection in a performant way. The default size of the batch is defined in the BATCH_SIZE constant.

See the following example to get an idea.

Purging data in batch:

include ::Gitlab::Database::DynamicModelHelpers

disable_ddl_transaction!

def up
  each_batch_range('ci_pending_builds', scope: ->(table) { table.ref_protected }, of: BATCH_SIZE) do |min, max|
    execute <<~SQL
      DELETE FROM ci_pending_builds
        USING ci_builds
        WHERE ci_builds.id = ci_pending_builds.build_id
          AND ci_builds.status != 'pending'
          AND ci_builds.type = 'Ci::Build'
          AND ci_pending_builds.id BETWEEN #{min} AND #{max}
    SQL
  end
end
  • The first argument is the table being modified: 'ci_pending_builds'.
  • The second argument calls a lambda which fetches the relevant dataset selected (the default is set to .all): scope: ->(table) { table.ref_protected }.
  • The third argument is the batch size (the default is set in the BATCH_SIZE constant): of: BATCH_SIZE.

Here is an example MR illustrating how to use our new helper.

Using application code in migrations (discouraged)

The use of application code (including models) in migrations is generally discouraged. This is because the migrations stick around for a long time and the application code it depends on may change and break the migration in future. In the past some background migrations needed to use application code in order to avoid copying hundreds of lines of code spread across multiple files into the migration. In these rare cases it's critical to ensure the migration has good tests so that anyone refactoring the code in future will learn if they break the migration. Using application code is also discouraged for batched background migrations , the model needs to be declared in the migration.

Usually you can avoid using application code (specifically models) in a migration by defining a class that inherits from MigrationRecord (see examples below).

If using are using a model (including defined in the migration), you should first clear the column cache using reset_column_information.

If using a model that leverages single table inheritance (STI), there are special considerations.

This avoids problems where a column that you are using was altered and cached in a previous migration.

Example: Add a column my_column to the users table

It is important not to leave out the User.reset_column_information command, to ensure that the old schema is dropped from the cache and ActiveRecord loads the updated schema information.

class AddAndSeedMyColumn < Gitlab::Database::Migration[2.1]
  class User < MigrationRecord
    self.table_name = 'users'
  end

  def up
    User.count # Any ActiveRecord calls on the model that caches the column information.

    add_column :users, :my_column, :integer, default: 1

    User.reset_column_information # The old schema is dropped from the cache.
    User.find_each do |user|
      user.my_column = 42 if some_condition # ActiveRecord sees the correct schema here.
      user.save!
    end
  end
end

The underlying table is modified and then accessed by using ActiveRecord.

This also needs to be used if the table is modified in a previous, different migration, if both migrations are run in the same db:migrate process.

This results in the following. Note the inclusion of my_column:

== 20200705232821 AddAndSeedMyColumn: migrating ==============================
D, [2020-07-06T00:37:12.483876 #130101] DEBUG -- :    (0.2ms)  BEGIN
D, [2020-07-06T00:37:12.521660 #130101] DEBUG -- :    (0.4ms)  SELECT COUNT(*) FROM "user"
-- add_column(:users, :my_column, :integer, {:default=>1})
D, [2020-07-06T00:37:12.523309 #130101] DEBUG -- :    (0.8ms)  ALTER TABLE "users" ADD "my_column" integer DEFAULT 1
   -> 0.0016s
D, [2020-07-06T00:37:12.650641 #130101] DEBUG -- :   AddAndSeedMyColumn::User Load (0.7ms)  SELECT "users".* FROM "users" ORDER BY "users"."id" ASC LIMIT $1  [["LIMIT", 1000]]
D, [2020-07-18T00:41:26.851769 #459802] DEBUG -- :   AddAndSeedMyColumn::User Update (1.1ms)  UPDATE "users" SET "my_column" = $1, "updated_at" = $2 WHERE "users"."id" = $3  [["my_column", 42], ["updated_at", "2020-07-17 23:41:26.849044"], ["id", 1]]
D, [2020-07-06T00:37:12.653648 #130101] DEBUG -- :   ↳ config/initializers/config_initializers_active_record_locking.rb:13:in `_update_row'
== 20200705232821 AddAndSeedMyColumn: migrated (0.1706s) =====================

If you skip clearing the schema cache (User.reset_column_information), the column is not used by ActiveRecord and the intended changes are not made, leading to the result below, where my_column is missing from the query.

== 20200705232821 AddAndSeedMyColumn: migrating ==============================
D, [2020-07-06T00:37:12.483876 #130101] DEBUG -- :    (0.2ms)  BEGIN
D, [2020-07-06T00:37:12.521660 #130101] DEBUG -- :    (0.4ms)  SELECT COUNT(*) FROM "user"
-- add_column(:users, :my_column, :integer, {:default=>1})
D, [2020-07-06T00:37:12.523309 #130101] DEBUG -- :    (0.8ms)  ALTER TABLE "users" ADD "my_column" integer DEFAULT 1
   -> 0.0016s
D, [2020-07-06T00:37:12.650641 #130101] DEBUG -- :   AddAndSeedMyColumn::User Load (0.7ms)  SELECT "users".* FROM "users" ORDER BY "users"."id" ASC LIMIT $1  [["LIMIT", 1000]]
D, [2020-07-06T00:37:12.653459 #130101] DEBUG -- :   AddAndSeedMyColumn::User Update (0.5ms)  UPDATE "users" SET "updated_at" = $1 WHERE "users"."id" = $2  [["updated_at", "2020-07-05 23:37:12.652297"], ["id", 1]]
D, [2020-07-06T00:37:12.653648 #130101] DEBUG -- :   ↳ config/initializers/config_initializers_active_record_locking.rb:13:in `_update_row'
== 20200705232821 AddAndSeedMyColumn: migrated (0.1706s) =====================

High traffic tables

Here's a list of current high-traffic tables.

Determining what tables are high-traffic can be difficult. Self-managed instances might use different features of GitLab with different usage patterns, thus making assumptions based on GitLab.com not enough.

To identify a high-traffic table for GitLab.com the following measures are considered. The metrics linked here are GitLab-internal only:

Any table which has some high read operation compared to current high-traffic tables might be a good candidate.

As a general rule, we discourage adding columns to high-traffic tables that are purely for analytics or reporting of GitLab.com. This can have negative performance impacts for all self-managed instances without providing direct feature value to them.

Milestone

Beginning in GitLab 16.6, all new migrations must specify a milestone, using the following syntax:

class AddFooToBar < Gitlab::Database::Migration[2.2]
  milestone '16.6'

  def change
    # Your migration here
  end
end

Adding the correct milestone to a migration enables us to logically partition migrations into their corresponding GitLab minor versions. This:

  • Simplifies the upgrade process.
  • Alleviates potential migration ordering issues that arise when we rely solely on the migration's timestamp for ordering.

Autovacuum wraparound protection

This is a special autovacuum run mode for PostgreSQL and it requires a ShareUpdateExclusiveLock on the table that it is vacuuming. For larger tables this could take hours and the lock can conflict with most DDL migrations that try to modify the table at the same time. Because the migrations will not be able to acquire the lock in time, they will fail and block the deployments.

The post-deploy migration (PDM) pipeline can check and halt its execution if it detects a wraparound prevention vacuum process on one of the tables. For this to happen we need to use the complete table name in the migration name. For example add_foreign_key_between_ci_builds_and_ci_job_artifacts will check for vacuum on ci_builds and ci_job_artifacts before executing the migrations.

If the migration doesn't have conflicting locks, the vacuum check can be skipped by not using the complete table name, for example create_async_index_on_job_artifacts.