Avoiding downtime in migrations

When working with a database certain operations may require downtime. Since we cannot have downtime in migrations we need to use a set of steps to get the same end result without downtime. This guide describes various operations that may appear to need downtime, their impact, and how to perform them without requiring downtime.

Dropping Columns

Removing columns is tricky because running GitLab processes may still be using the columns. To work around this safely, you will need three steps in three releases:

  1. Ignoring the column (release M)
  2. Dropping the column (release M+1)
  3. Removing the ignore rule (release M+2)

The reason we spread this out across three releases is that dropping a column is a destructive operation that can't be rolled back easily.

Following this procedure helps us to make sure there are no deployments to GitLab.com and upgrade processes for self-managed installations that lump together any of these steps.

Step 1: Ignoring the column (release M)

The first step is to ignore the column in the application code. This is necessary because Rails caches the columns and re-uses this cache in various places. This can be done by defining the columns to ignore. For example, to ignore updated_at in the User model you'd use the following:

class User < ApplicationRecord
  include IgnorableColumns
  ignore_column :updated_at, remove_with: '12.7', remove_after: '2020-01-22'

Multiple columns can be ignored, too:

ignore_columns %i[updated_at created_at], remove_with: '12.7', remove_after: '2020-01-22'

If the model exists in CE and EE, the column has to be ignored in the CE model. If the model only exists in EE, then it has to be added there.

We require indication of when it is safe to remove the column ignore with:

  • remove_with: set to a GitLab release typically two releases (M+2) after adding the column ignore.
  • remove_after: set to a date after which we consider it safe to remove the column ignore, typically after the M+1 release date, during the M+2 development cycle.

This information allows us to reason better about column ignores and makes sure we don't remove column ignores too early for both regular releases and deployments to GitLab.com. For example, this avoids a situation where we deploy a bulk of changes that include both changes to ignore the column and subsequently remove the column ignore (which would result in a downtime).

In this example, the change to ignore the column went into release 12.5.

Step 2: Dropping the column (release M+1)

Continuing our example, dropping the column goes into a post-deployment migration in release 12.6:

 remove_column :user, :updated_at

Step 3: Removing the ignore rule (release M+2)

With the next release, in this example 12.7, we set up another merge request to remove the ignore rule. This removes the ignore_column line and - if not needed anymore - also the inclusion of IgnoreableColumns.

This should only get merged with the release indicated with remove_with and once the remove_after date has passed.

Renaming Columns

Renaming columns the normal way requires downtime as an application may continue using the old column name during/after a database migration. To rename a column without requiring downtime we need two migrations: a regular migration, and a post-deployment migration. Both these migration can go in the same release.

Step 1: Add The Regular Migration

First we need to create the regular migration. This migration should use Gitlab::Database::MigrationHelpers#rename_column_concurrently to perform the renaming. For example

# A regular migration in db/migrate
class RenameUsersUpdatedAtToUpdatedAtTimestamp < Gitlab::Database::Migration[1.0]

  def up
    rename_column_concurrently :users, :updated_at, :updated_at_timestamp

  def down
    undo_rename_column_concurrently :users, :updated_at, :updated_at_timestamp

This will take care of renaming the column, ensuring data stays in sync, and copying over indexes and foreign keys.

If a column contains one or more indexes that don't contain the name of the original column, the previously described procedure will fail. In that case, you'll first need to rename these indexes.

Step 2: Add A Post-Deployment Migration

The renaming procedure requires some cleaning up in a post-deployment migration. We can perform this cleanup using Gitlab::Database::MigrationHelpers#cleanup_concurrent_column_rename:

# A post-deployment migration in db/post_migrate
class CleanupUsersUpdatedAtRename < Gitlab::Database::Migration[1.0]

  def up
    cleanup_concurrent_column_rename :users, :updated_at, :updated_at_timestamp

  def down
    undo_cleanup_concurrent_column_rename :users, :updated_at, :updated_at_timestamp

If you're renaming a large table, please carefully consider the state when the first migration has run but the second cleanup migration hasn't been run yet. With Canary it is possible that the system runs in this state for a significant amount of time.

Changing Column Constraints

Adding or removing a NOT NULL clause (or another constraint) can typically be done without requiring downtime. However, this does require that any application changes are deployed first. Thus, changing the constraints of a column should happen in a post-deployment migration.

Avoid using change_column as it produces an inefficient query because it re-defines the whole column type.

You can check the following guides for each specific use case:

Changing Column Types

Changing the type of a column can be done using Gitlab::Database::MigrationHelpers#change_column_type_concurrently. This method works similarly to rename_column_concurrently. For example, let's say we want to change the type of users.username from string to text.

Step 1: Create A Regular Migration

A regular migration is used to create a new column with a temporary name along with setting up some triggers to keep data in sync. Such a migration would look as follows:

# A regular migration in db/migrate
class ChangeUsersUsernameStringToText < Gitlab::Database::Migration[1.0]

  def up
    change_column_type_concurrently :users, :username, :text

  def down
    undo_change_column_type_concurrently :users, :username

Step 2: Create A Post Deployment Migration

Next we need to clean up our changes using a post-deployment migration:

# A post-deployment migration in db/post_migrate
class ChangeUsersUsernameStringToTextCleanup < Gitlab::Database::Migration[1.0]

  def up
    cleanup_concurrent_column_type_change :users, :username

  def down
    undo_cleanup_concurrent_column_type_change :users, :username, :string

And that's it, we're done!

Casting data to a new type

Some type changes require casting data to a new type. For example when changing from text to jsonb. In this case, use the type_cast_function option. Make sure there is no bad data and the cast will always succeed. You can also provide a custom function that handles casting errors.

Example migration:

  def up
    change_column_type_concurrently :users, :settings, :jsonb, type_cast_function: 'jsonb'

Changing The Schema For Large Tables

While change_column_type_concurrently and rename_column_concurrently can be used for changing the schema of a table without downtime, it doesn't work very well for large tables. Because all of the work happens in sequence the migration can take a very long time to complete, preventing a deployment from proceeding. They can also produce a lot of pressure on the database due to it rapidly updating many rows in sequence.

To reduce database pressure you should instead use a background migration when migrating a column in a large table (for example, issues). This will spread the work / load over a longer time period, without slowing down deployments.

For more information, see the documentation on cleaning up background migrations.

Adding Indexes

Adding indexes does not require downtime when add_concurrent_index is used.

See also Migration Style Guide for more information.

Dropping Indexes

Dropping an index does not require downtime.

Adding Tables

This operation is safe as there's no code using the table just yet.

Dropping Tables

Dropping tables can be done safely using a post-deployment migration, but only if the application no longer uses the table.

Renaming Tables

Renaming tables requires downtime as an application may continue using the old table name during/after a database migration.

If the table and the ActiveRecord model is not in use yet, removing the old table and creating a new one is the preferred way to "rename" the table.

Renaming a table is possible without downtime by following our multi-release rename table process.

Adding Foreign Keys

Adding foreign keys usually works in 3 steps:

  1. Start a transaction
  2. Run ALTER TABLE to add the constraint(s)
  3. Check all existing data

Because ALTER TABLE typically acquires an exclusive lock until the end of a transaction this means this approach would require downtime.

GitLab allows you to work around this by using Gitlab::Database::MigrationHelpers#add_concurrent_foreign_key. This method ensures that no downtime is needed.

Removing Foreign Keys

This operation does not require downtime.

Migrating integer primary keys to bigint

To prevent the overflow risk for some tables with integer primary key (PK), we have to migrate their PK to bigint. The process to do this without downtime and causing too much load on the database is described below.

Initialize the conversion and start migrating existing data (release N)

To start the process, add a regular migration to create the new bigint columns. Use the provided initialize_conversion_of_integer_to_bigint helper. The helper also creates a database trigger to keep in sync both columns for any new records (see an example):

class InitializeConversionOfCiStagesToBigint < ActiveRecord::Migration[6.1]
  include Gitlab::Database::MigrationHelpers

  TABLE = :ci_stages
  COLUMNS = %i(id)

  def up
    initialize_conversion_of_integer_to_bigint(TABLE, COLUMNS)

  def down
    revert_initialize_conversion_of_integer_to_bigint(TABLE, COLUMNS)

Ignore the new bigint columns:

module Ci
  class Stage < Ci::ApplicationRecord
    include IgnorableColumns
    ignore_column :id_convert_to_bigint, remove_with: '14.2', remove_after: '2021-08-22'

To migrate existing data, we introduced new type of batched background migrations. Unlike the classic background migrations, built on top of Sidekiq, batched background migrations don't have to enqueue and schedule all the background jobs at the beginning. They also have other advantages, like automatic tuning of the batch size, better progress visibility, and collecting metrics. To start the process, use the provided backfill_conversion_of_integer_to_bigint helper (example):

class BackfillCiStagesForBigintConversion < ActiveRecord::Migration[6.1]
  include Gitlab::Database::MigrationHelpers

  TABLE = :ci_stages
  COLUMNS = %i(id)

  def up
    backfill_conversion_of_integer_to_bigint(TABLE, COLUMNS)

  def down
    revert_backfill_conversion_of_integer_to_bigint(TABLE, COLUMNS)

Monitor the background migration

Check how the migration is performing while it's running. Multiple ways to do this are described below.

High-level status of batched background migrations

See how to check the status of batched background migrations.

Query the database

We can query the related database tables directly. Requires access to read-only replica. Example queries:

-- Get details for batched background migration for given table
SELECT * FROM batched_background_migrations WHERE table_name = 'namespaces'\gx

-- Get count of batched background migration jobs by status for given table
  batched_background_migrations.id, batched_background_migration_jobs.status, COUNT(*)
  JOIN batched_background_migration_jobs ON batched_background_migrations.id = batched_background_migration_jobs.batched_background_migration_id
  table_name = 'namespaces'
  batched_background_migrations.id, batched_background_migration_jobs.status;

-- Batched background migration progress for given table (based on estimated total number of tuples)
  LEAST(100 * sum(j.batch_size) / pg_class.reltuples, 100) AS percentage_complete
  batched_background_migrations m
  JOIN batched_background_migration_jobs j ON j.batched_background_migration_id = m.id
  JOIN pg_class ON pg_class.relname = m.table_name
  j.status = 3 AND m.table_name = 'namespaces'
GROUP BY m.id, pg_class.reltuples;

Sidekiq logs

We can also use the Sidekiq logs to monitor the worker that executes the batched background migrations:

  1. Sign in to Kibana with a @gitlab.com email address.
  2. Change the index pattern to pubsub-sidekiq-inf-gprd*.
  3. Add filter for json.queue: cronjob:database_batched_background_migration.

PostgreSQL slow queries log

Slow queries log keeps track of low queries that took above 1 second to execute. To see them for batched background migration:

  1. Sign in to Kibana with a @gitlab.com email address.
  2. Change the index pattern to pubsub-postgres-inf-gprd*.
  3. Add filter for json.endpoint_id.keyword: Database::BatchedBackgroundMigrationWorker.
  4. Optional. To see only updates, add a filter for json.command_tag.keyword: UPDATE.
  5. Optional. To see only failed statements, add a filter for json.error_severity.keyword: ERROR.
  6. Optional. Add a filter by table name.

Grafana dashboards

To monitor the health of the database, use these additional metrics:

  • PostgreSQL Tuple Statistics: if you see high rate of updates for the tables being actively converted, or increasing percentage of dead tuples for this table, it might mean that autovacuum cannot keep up.
  • PostgreSQL Overview: if you see high system usage or transactions per second (TPS) on the primary database server, it might mean that the migration is causing problems.

Prometheus metrics

Number of metrics for each batched background migration are published to Prometheus. These metrics can be searched for and visualized in Thanos (see an example).

Swap the columns (release N + 1)

After the background is completed and the new bigint columns are populated for all records, we can swap the columns. Swapping is done with post-deployment migration. The exact process depends on the table being converted, but in general it's done in the following steps:

  1. Using the provided ensure_batched_background_migration_is_finished helper, make sure the batched migration has finished (see an example). If the migration has not completed, the subsequent steps fail anyway. By checking in advance we aim to have more helpful error message.
  2. Create indexes using the bigint columns that match the existing indexes using the integer column (see an example).
  3. Create foreign keys (FK) using the bigint columns that match the existing FKs using the integer column. Do this both for FK referencing other tables, and FKs that reference the table that is being migrated (see an example).
  4. Inside a transaction, swap the columns:
    1. Lock the tables involved. To reduce the chance of hitting a deadlock, we recommended to do this in parent to child order (see an example).
    2. Rename the columns to swap names (see an example)
    3. Reset the trigger function (see an example).
    4. Swap the defaults (see an example).
    5. Swap the PK constraint (if any) (see an example).
    6. Remove old indexes and rename new ones (see an example).
    7. Remove old FKs (if still present) and rename new ones (see an example).

See example merge request, and migration.

Remove the trigger and old integer columns (release N + 2)

Using post-deployment migration and the provided cleanup_conversion_of_integer_to_bigint helper, drop the database trigger and the old integer columns (see an example).

Remove ignore rules (release N + 3)

In the next release after the columns were dropped, remove the ignore rules as we do not need them anymore (see an example).

Data migrations

Data migrations can be tricky. The usual approach to migrate data is to take a 3 step approach:

  1. Migrate the initial batch of data
  2. Deploy the application code
  3. Migrate any remaining data

Usually this works, but not always. For example, if a field's format is to be changed from JSON to something else we have a bit of a problem. If we were to change existing data before deploying application code we'll most likely run into errors. On the other hand, if we were to migrate after deploying the application code we could run into the same problems.

If you merely need to correct some invalid data, then a post-deployment migration is usually enough. If you need to change the format of data (for example, from JSON to something else) it's typically best to add a new column for the new data format, and have the application use that. In such a case the procedure would be:

  1. Add a new column in the new format
  2. Copy over existing data to this new column
  3. Deploy the application code
  4. In a post-deployment migration, copy over any remaining data

In general there is no one-size-fits-all solution, therefore it's best to discuss these kind of migrations in a merge request to make sure they are implemented in the best way possible.