How to add index to a big table of your Rails app

… and check why 5600+ Rails engineers read also this

How to add index to a big table of your Rails app

When your application is successful, some of the tables can grow pretty big — I’m looking at you users table. If you’re curious enough, you periodically check how your database performs. If any slow query pops up in the metrics, there’s a great chance that some index is missing.

State of the DB engines

While most modern database engines can create indexes in an asynchronous, non–blocking manner, it’s good to get familiar with all the exceptions from this rule. I highly recommend reading the documentation of PostgreSQL, MySQL. I’m not sure about SQLite here as the documentation doesn’t clearly state that. However, my quick chit–chat with LLM may give you some insights.

What’s the problem then?

As you already know CREATE INDEX statement will be handled asynchronously by the database if appropriate algorithm is used. This means that no reads, writes and update will be blocked.

Typically, for the Rails application, you’ll run a migration via Ruby process during deployment, using all the goodies from ActiveRecord::Migration class and its surroundings.

Let’s imagine the database schema like that:

ActiveRecord::Schema[7.1].define(version: 2024_06_13_121701) do
  create_table "users", force: :cascade do |t|
    t.string "email"
    t.boolean "active"
    t.datetime "created_at", null: false
    t.datetime "updated_at", null: false
  end
end

Imagine you want to quickly find all those users which didn’t activate the account:

User.where(active: false)

If you have enough users, speaking of dozens or hundreds of millions, doing full table scan could simply kill the database performance. Full table scan happens when database has no index to use and need to check every row whether it meets the criteria.

I will stick with PostreSQL in the examples if not stated otherwise.

Obvious solution

Let’s add the index then:

➜  trololo git:(master) bin/rails g migration AddIndexOnActiveUsers
      invoke  active_record
      create    db/migrate/20240613121751_add_index_on_active_users.rb

Implementation:

# PostgreSQL
class AddIndexOnActiveUsers < ActiveRecord::Migration[7.1]
  disable_ddl_transaction!

  def change
    add_index :users, :active, algorithm: :concurrently
  end
end

# MySQL
class AddIndexOnActiveUsers < ActiveRecord::Migration[7.1]
  def change
    add_index :users, :active, algorithm: :inplace
  end
end

Let’s run it:

➜  trololo git:(master) bin/rails db:migrate
== 20240613121751 AddIndexOnActiveUsers: migrating ============================
-- add_index(:users, :active)
   -> 0.0013s
== 20240613121751 AddIndexOnActiveUsers: migrated (0.0013s) ===================

Ok, it was quite fast, right? That’s correct for your dev machine, not necessarily for the production setup. Now magnify this time 0.0013s by 6 or 7 orders of magnitude if your table is big enough:

➜  trololo git:(master) bin/rails db:migrate
== 20240613121751 AddIndexOnActiveUsers: migrating ============================
-- add_index(:users, :active)
   -> 13000.4928s
== 20240613121751 AddIndexOnActiveUsers: migrated (0.0013s) ===================

I’ll do the maths for you: 1300.4928s means 21 minutes 40.49 seconds. But it can be even longer — don’t ask me how I found about this.

While the process of migration will end up eventually, the migration blocking any other deployments to your application during this time may be unacceptable for various reasons:

  • you cannot release any other change to production until migration completes
  • something other goes wrong and you need to rapidly deploy a hotfix, but you can’t since the deployment is blocked by long running migration™
  • process manager on a deployment machine may expect output within, e.g. 5 minutes. Long running migration will get killed in such scenario.

What to do then?

Simply skip the migration body for RAILS_ENV=production:

class AddIndexOnActiveUsers < ActiveRecord::Migration[7.1]
  disable_ddl_transaction!

  def change
    add_index :users, :active, algorithm: :concurrently, if_not_exists: true unless Rails.env.production?
  end
end

The migration will be executed on production environment, but will have no effect on your database. schema.rb or structure.sql (depending on what you use) will be aligned. Sufficient entry will also appear in schema_migrations table. All the developers will have index added on their local databases, test environment will be aligned too.

But Hey, where’s my index on production?! you might ask. And that’s a pretty valid question. What you’ll need is a way to run a rails runner or rake task on production. Depending on the capabilities you have, you might choose to:

  • Run it within bin/rails console:
# PostgreSQL
ActiveRecord::Migration.add_index :users, :active, algorithm: :concurrently, if_not_exists: true

# MySQL
ActiveRecord::Migration.add_index :users, :active, algorithm: :inplace, if_not_exists: true
  • Do the same via bin/rails runner:
# PostgreSQL
bin/rails r "ActiveRecord::Migration.add_index :users, :active, algorithm: :concurrently, if_not_exists: true"

#MySQL
bin/rails r "ActiveRecord::Migration.add_index :users, :active, algorithm: :inplace, if_not_exists: true"
  • Last, but not least, implement a Rake task. It has the advantage that it has to be committed to the repository so you don’t lose the history what’s happened:
# PostgreSQL
namespace :indexes do
  task add_index_on_active_users: :environment do
    ActiveRecord::Migration.add_index :users, :active, algorithm: :concurrently, if_not_exists: true
  end
end

# MySQL
namespace :indexes do
  task add_index_on_active_users: :environment do
    ActiveRecord::Migration.add_index :users, :active, algorithm: :inplace, if_not_exists: true
  end
end

Execute it with bin/rails indexes:add_index_on_active_users.

For the last option it’s also easy to enhance it with logging to easily identify execution in Grafana, Datadog or any other tool you use for your logs.

namespace :indexes do
  task add_index_on_active_users: :environment do
    Rails.logger.info("task indexes:add_index_on_active_users started")

    ActiveRecord::Migration.add_index :users, :active, algorithm: :concurrently, if_not_exists: true

    Rails.logger.info("task indexes:add_index_on_active_users finished”)
  end
end

Tiny details

If you’re aware enough, you probably spotted if_not_exists: true flag. We like idempotence and that’s the reason. If anyone runs this task again, nothing will happen. If you prefer to see ActiveRecord::StatementInvalid instead, feel free to skip it.

As mentioned in the preface, to use appropriate algorithm for index creation algorithm: :concurrently for PostgreSQL and algorithm: inplace for MySQL has to be specified.

There’s another quirk for PostgreSQL:

Another difference is that a regular CREATE INDEX command can be performed within a transaction block, but CREATE INDEX CONCURRENTLY cannot.

Each Rails migration execution is wrapped within transaction. To disable this behavior, you need to use disable_ddl_transaction! method within your migration. Otherwise, following error will pop up:

bin/rails aborted!
StandardError: An error has occurred, this and all later migrations canceled: (StandardError)

PG::ActiveSqlTransaction: ERROR:  CREATE INDEX CONCURRENTLY cannot run inside a transaction block
/Users/fidel/code/fidel/trololo/db/migrate/20240613121751_add_index_on_active_users.rb:5:in `change'

Caused by:
ActiveRecord::StatementInvalid: PG::ActiveSqlTransaction: ERROR:  CREATE INDEX CONCURRENTLY cannot run inside a transaction block (ActiveRecord::StatementInvalid)
/Users/fidel/code/fidel/trololo/db/migrate/20240613121751_add_index_on_active_users.rb:5:in `change'

Caused by:
PG::ActiveSqlTransaction: ERROR:  CREATE INDEX CONCURRENTLY cannot run inside a transaction block (PG::ActiveSqlTransaction)
/Users/fidel/code/fidel/trololo/db/migrate/20240613121751_add_index_on_active_users.rb:5:in `change'
Tasks: TOP => db:migrate
(See full trace by running task with --trace)

However, it’s not a problem for our custom script or Rake task run on production environment.

You might also like