Active Record Replica

Redirect ActiveRecord (Rails) reads to replica databases while ensuring all writes go to the primary database.

Status

This is a slight modification of Rocket Job's original library, simply renaming it from active_record_slave to active_record_replica.

In order to more clearly distinguish the library from active_record_slave, we also incremented the major version – it is, however, functionally equivalent.

Introduction

active_record_replica redirects all database reads to replica instances while ensuring that all writes go to the primary database. active_record_replica ensures that any reads that are performed within a database transaction are by default directed to the primary database to ensure data consistency.

Status

Production Ready. Actively used in large production environments.

Features

• Redirecting reads to a single replica database.
• Works with any database driver that works with ActiveRecord.
• Supports all Rails 3, 4, or 5 read apis.
  • Including dynamic finders, AREL, and ActiveRecord::Base.select.
  • NOTE: In Rails 3 and 4, QueryCache is only enabled for BaseConnection by default. In Rails 5, it's enabled for all connections. (PR)
• Transaction aware
  • Detects when a query is inside of a transaction and sends those reads to the primary by default.
  • Can be configured to send reads in a transaction to replica databases.
• Lightweight footprint.
• No overhead whatsoever when a replica is not configured.
• Negligible overhead when redirecting reads to the replica.
• Connection Pools to both databases are retained and maintained independently by ActiveRecord.
• The primary and replica databases do not have to be of the same type.
  • For example Oracle could be the primary with MySQL as the replica database.
• Debug logs include a prefix of Replica: to indicate which SQL statements are going to the replica database.

Example showing Replica redirected read

# Read from the replica database
r = Role.where(name: 'manager').first
r.description = 'Manager'

# Save changes back to the primary database
r.save!

Log file output:

03-13-12 05:56:05 pm,[2608],b[0],[0],  Replica: Role Load (3.0ms)  SELECT `roles`.* FROM `roles` WHERE `roles`.`name` = 'manager' LIMIT 1
03-13-12 05:56:22 pm,[2608],b[0],[0],  AREL (12.0ms)  UPDATE `roles` SET `description` = 'Manager' WHERE `roles`.`id` = 5

Example showing how reads within a transaction go to the primary

Role.transaction do
  r = Role.where(name: 'manager').first
  r.description = 'Manager'
  r.save!
end

Log file output:

03-13-12 06:02:09 pm,[2608],b[0],[0],  Role Load (2.0ms)  SELECT `roles`.* FROM `roles` WHERE `roles`.`name` = 'manager' LIMIT 1
03-13-12 06:02:09 pm,[2608],b[0],[0],  AREL (2.0ms)  UPDATE `roles` SET `description` = 'Manager' WHERE `roles`.`id` = 4

Forcing a read against the primary

Sometimes it is necessary to read from the primary:

ActiveRecordReplica.read_from_primary do
  r = Role.where(name: 'manager').first
end

Usage Notes

delete_all

Delete all executes against the primary database since it is only a delete:

D, [2012-11-06T19:47:29.125932 #89772] DEBUG -- :   SQL (1.0ms)  DELETE FROM "users"

destroy_all

First performs a read against the replica database and then deletes the corresponding data from the primary

D, [2012-11-06T19:43:26.890674 #89002] DEBUG -- :   Replica: User Load (0.1ms)  SELECT "users".* FROM "users"
D, [2012-11-06T19:43:26.890972 #89002] DEBUG -- :    (0.0ms)  begin transaction
D, [2012-11-06T19:43:26.891667 #89002] DEBUG -- :   SQL (0.4ms)  DELETE FROM "users" WHERE "users"."id" = ?  [["id", 3]]
D, [2012-11-06T19:43:26.892697 #89002] DEBUG -- :    (0.9ms)  commit transaction

Transactions

By default ActiveRecordReplica detects when a call is inside a transaction and will send all reads to the primary when a transaction is active.

It is now possible to send reads to database replicas and ignore whether currently inside a transaction:

In file config/application.rb:

# Read from replica even when in an active transaction
config.active_record_replica.ignore_transactions = true

It is important to identify any code in the application that depends on being able to read any changes already part of the transaction, but not yet committed and wrap those reads with ActiveRecordReplica.read_from_primary

Inquiry.transaction do
  # Create a new inquiry
  Inquiry.create

  # The above inquiry is not visible yet if already in a Rails transaction.
  # Use `read_from_primary` to ensure it is included in the count below:
  ActiveRecordReplica.read_from_primary do
    count = Inquiry.count
  end

end

Note

active_record_replica is a very simple layer that inserts itself into the call chain whenever a replica is configured. By observation we noticed that all reads are made to a select set of methods and all writes are made directly to one method: execute.

Using this observation active_record_replica only needs to intercept calls to the known select apis:

• select_all
• select_one
• select_rows
• select_value
• select_values

Calls to the above methods are redirected to the replica active record model ActiveRecordReplica::Replica. This model is 100% managed by the regular Active Record mechanisms such as connection pools etc.

This lightweight approach ensures that all calls to the above API's are redirected to the replica without impacting:

• Transactions
• Writes
• Any SQL calls directly to execute

One of the limitations with this approach is that any code that performs a query by calling execute direct will not be redirected to the replica instance. In this case replace the use of execute with one of the the above select methods.

Note when using dependent: destroy

When performing in-memory only model assignments Active Record will create a transaction against the primary even though the transaction may never be used.

Even though the transaction is unused it sends the following messages to the primary database: ~~ SET autocommit=0 commit SET autocommit=1 ~~

This will impact the primary database if sufficient calls are made, such as in batch workers.

For Example:

~~ruby class Parent < ActiveRecord::Base has_one :child, dependent: :destroy end

class Child < ActiveRecord::Base belongs_to :parent end

The following code will create an unused transaction against the primary, even when reads are going to replicas:

parent = Parent.new parent.child = Child.new ~~

If the dependent: :destroy is removed it no longer creates a transaction, but it also means dependents are not destroyed when a parent is destroyed.

For this scenario when we are 100% confident no writes are being performed the following can be performed to ignore any attempt Active Record makes at creating the transaction:

~~ruby ActiveRecordReplica.skip_transactions do parent = Parent.new parent.child = Child.new end ~~

To help identify any code within a block that is creating transactions, wrap the code with ActiveRecordReplica.block_transactions to make it raise an exception anytime a transaction is attempted:

~~ruby ActiveRecordReplica.block_transactions do parent = Parent.new parent.child = Child.new end ~~

Install

Add to Gemfile

gem 'active_record_replica'

Run bundler to install:

bundle

Or, without Bundler:

gem install active_record_replica

Configuration

To enable replica reads for any environment just add a replica: entry to database.yml along with all the usual ActiveRecord database configuration options.

For Example:

production:
  database: production
  username: username
  password: password
  encoding: utf8
  adapter:  mysql
  host:     primary1
  pool:     50
  replica:
    database: production
    username: username
    password: password
    encoding: utf8
    adapter:  mysql
    host:     replica1
    pool:     50

Sometimes it is useful to turn on replica reads per host, for example to activate replica reads only on the linux host 'batch':

production:
  database: production
  username: username
  password: password
  encoding: utf8
  adapter:  mysql
  host:     primary1
  pool:     50
<% if `hostname`.strip == 'batch' %>
  replica:
    database: production
    username: username
    password: password
    encoding: utf8
    adapter:  mysql
    host:     replica1
    pool:     50
<% end %>

If there are multiple replicas, it is possible to randomly select a replica on startup to balance the load across the replicas:

production:
  database: production
  username: username
  password: password
  encoding: utf8
  adapter:  mysql
  host:     primary1
  pool:     50
  replica:
    database: production
    username: username
    password: password
    encoding: utf8
    adapter:  mysql
    host:     <%= %w(replica1 replica2 replica3).sample %>
    pool:     50

Replicas can also be assigned to specific hosts by using the hostname:

production:
  database: production
  username: username
  password: password
  encoding: utf8
  adapter:  mysql
  host:     primary1
  pool:     50
  replica:
    database: production
    username: username
    password: password
    encoding: utf8
    adapter:  mysql
    host:     <%= `hostname`.strip == 'app1' ? 'replica1' : 'replica2' %>
    pool:     50

Set primary as default for Read

The default behavior can also set to read/write operations against primary database.

Create an initializer file config/initializer/active_record_replica.rb to force read from primary:

    ActiveRecordReplica.read_from_primary!

Then use this method and supply block to read from the replica database:

ActiveRecordReplica.read_from_replica do
   User.count
end

Dependencies

See .travis.yml for the list of tested Ruby platforms

Versioning

This project uses Semantic Versioning.

Contributing

1. Fork repository in Github.

2. Checkout your forked repository:

   git clone https://github.com/your_github_username/active_record_replica.git
   cd active_record_replica

3. Create branch for your contribution:

   git co -b your_new_branch_name

4. Make code changes.

5. Ensure tests pass.

6. Push to your fork origin.

   git push origin

7. Submit PR from the branch on your fork in Github.

Author

Reid Morrison :: @reidmorrison