dnahodil/zero_downtime_migrations

Zero downtime migrations with ActiveRecord and PostgreSQL

zero_downtime_migrations

Zero downtime migrations with ActiveRecord and PostgreSQL.

Catch problematic migrations at development/test time! Heavily inspired by these similar projects:

• https://github.com/ankane/strong_migrations
• https://github.com/foobarfighter/safe-migrations

Installation

Simply add this gem to the project Gemfile.

gem "zero_downtime_migrations"

Usage

This gem will automatically raise exceptions when potential database locking migrations are detected.

It checks for common things like:

• Adding a column with a default
• Adding a non-concurrent index
• Mixing data changes with index or schema migrations
• Performing data or schema migrations with the DDL transaction disabled
• Using each instead of find_each to loop thru ActiveRecord objects

These exceptions display clear instructions of how to perform the same operation the "zero downtime way".

Validations

Adding a column with a default

Bad

This can take a long time with significant database size or traffic and lock your table!

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published, :boolean, default: true
  end
end

Good

First let’s add the column without a default. When we add a column with a default it has to lock the table while it performs an UPDATE for ALL rows to set this new default.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published, :boolean
  end
end

Then we’ll set the new column default in a separate migration. Note that this does not update any existing data! This only sets the default for newly inserted rows going forward.

class SetPublishedDefaultOnPosts < ActiveRecord::Migration[5.0]
  def change
    change_column_default :posts, :published, true
  end
end

Finally we’ll backport the default value for existing data in batches. This should be done in its own migration as well. Updating in batches allows us to lock 1000 rows at a time (or whatever batch size we prefer).

class BackportPublishedDefaultOnPosts < ActiveRecord::Migration[5.0]
  def change
    Post.select(:id).find_in_batches.with_index do |batch, index|
      puts "Processing batch #{index}\r"
      Post.where(id: batch).update_all(published: true)
    end
  end
end

Adding an index concurrently

Bad

This action can lock your database table while indexing existing data!

class IndexUsersOnEmail < ActiveRecord::Migration[5.0]
  def change
    add_index :users, :email
  end
end

Good

Instead, let's add the index concurrently in its own migration with the DDL transaction disabled.

This allows PostgreSQL to build the index without locking in a way that prevent concurrent inserts, updates, or deletes on the table. Standard indexes lock out writes (but not reads) on the table.

class IndexUsersOnEmail < ActiveRecord::Migration[5.0]
  disable_ddl_transaction!

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

Mixing data/index/schema migrations

Bad

Performing migrations that change the schema, update data, or add indexes within one big transaction is unsafe!

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published, :boolean
    Post.update_all(published: true)
    add_index :posts, :published
  end
end

Good

Instead, let's split apart these types of migrations into separate files.

• Introduce schema changes with methods like create_table or add_column in one file. These should be run within a DDL transaction so that they can be rolled back if there are any issues.
• Update data with methods like update_all or save in another file. Data migrations tend to be much more error prone than changing the schema or adding indexes.
• Add indexes concurrently within their own file as well. Indexes should be created without the DDL transaction enabled to avoid table locking.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published, :boolean
  end
end

class BackportPublishedOnPosts < ActiveRecord::Migration[5.0]
  def change
    Post.update_all(published: true)
  end
end

class IndexPublishedOnPosts < ActiveRecord::Migration[5.0]
  disable_ddl_transaction!

  def change
    add_index :posts, :published, algorithm: :concurrently
  end
end

Disabling the DDL transaction

Bad

The DDL transaction should only be disabled for migrations that add indexes. All other types of migrations should keep the DDL transaction enabled so that changes can be rolled back if any unexpected errors occur.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  disable_ddl_transaction!

  def change
    add_column :posts, :published, :boolean
  end
end

class UpdatePublishedOnPosts < ActiveRecord::Migration[5.0]
  disable_ddl_transaction!

  def change
    Post.update_all(published: true)
  end
end

Good

Any other data or schema changes must live in their own migration files with the DDL transaction enabled just in case they make changes that need to be rolled back.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published, :boolean
  end
end

class UpdatePublishedOnPosts < ActiveRecord::Migration[5.0]
  def change
    Post.update_all(published: true)
  end
end

Looping thru ActiveRecord::Base objects

Bad

This might accidentally load tens or hundreds of thousands of records into memory all at the same time!

class BackportPublishedDefaultOnPosts < ActiveRecord::Migration[5.0]
  def change
    Post.all.each do |post|
      post.update_attribute(published: true)
    end
  end
end

Good

Let's use the find_each method to fetch records in batches instead.

class BackportPublishedDefaultOnPosts < ActiveRecord::Migration[5.0]
  def change
    Post.all.find_each do |post|
      post.update_attribute(published: true)
    end
  end
end

Renaming a column

Bad

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    rename_column :posts, :published, :published_new_name
  end
end

Good

• Introduce schema changes with methods like create_table or add_column
• Copy data with methods like update_all
• Introduce code to write to both columns
• Deploy
• Add code to stop writing to original column
• Deploy
• Drop the column in last separate deploy

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    add_column :posts, :published_DEPRECATED, :boolean  #published is original column
    Post.update_all("\"published_DEPRECATED\"=published") #need to quote capitals in postgres
  end
end

class CopyPublishedDataOnPosts < ActiveRecord::Migration[5.0]
  def change
    remove_column :posts, :published
  end
end

TODO

• Changing a column type
• Removing a column
• Renaming a table

Disabling "zero downtime migration" enforcements

We can disable any of these "zero downtime migration" enforcements by wrapping them in a safety_assured block.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  def change
    safety_assured do
      add_column :posts, :published, :boolean, default: true
    end
  end
end

We can also mark an entire migration as safe by using the safety_assured helper method.

class AddPublishedToPosts < ActiveRecord::Migration[5.0]
  safety_assured

  def change
    add_column :posts, :published, :boolean
    Post.where("created_at >= ?", 1.day.ago).update_all(published: true)
  end
end

Enforcements can be globally disabled by setting ENV["SAFETY_ASSURED"] when running migrations.

SAFETY_ASSURED=1 bundle exec rake db:migrate --trace

These enforcements are automatically disabled by default for the following scenarios:

• The database schema is being loaded with rake db:schema:load instead of db:migrate
• The current migration is a reverse (down) migration
• The current migration is named RollupMigrations

Testing

bundle exec rspec

Contributing

• Fork the project.
• Make your feature addition or bug fix.
• Add tests for it. This is important so we don't break it in a future version unintentionally.
• Commit, do not mess with the version or history.
• Open a pull request. Bonus points for topic branches.

Authors

• Sean Huber

License

MIT - Copyright © 2016 LendingHome