Active Record Replica
Redirect ActiveRecord (Rails) reads to replica databases while ensuring all writes go to the primary database.
This is a slight modification of Rocket Job's original library, simply renaming it from
In order to more clearly distinguish the library from
active_record_slave, we also incremented the major version – it is, however, functionally equivalent.
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.
Production Ready. Actively used in large production environments.
- 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,,b,, Replica: Role Load (3.0ms) SELECT `roles`.* FROM `roles` WHERE `roles`.`name` = 'manager' LIMIT 1 03-13-12 05:56:22 pm,,b,, 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,,b,, Role Load (2.0ms) SELECT `roles`.* FROM `roles` WHERE `roles`.`name` = 'manager' LIMIT 1 03-13-12 06:02:09 pm,,b,, 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
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"
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
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
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
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:
Using this observation
active_record_replica only needs to intercept calls to the known select apis:
Calls to the above methods are redirected to the replica active record model
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:
- Any SQL calls directly to
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
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.
~~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 ~~
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 ~~
Run bundler to install:
Or, without Bundler:
gem install active_record_replica
To enable replica reads for any environment just add a replica: entry to database.yml along with all the usual ActiveRecord database configuration options.
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:
Then use this method and supply block to read from the replica database:
ActiveRecordReplica.read_from_replica do User.count end
See .travis.yml for the list of tested Ruby platforms
This project uses Semantic Versioning.
Fork repository in Github.
Checkout your forked repository:
git clone https://github.com/your_github_username/active_record_replica.git cd active_record_replica
Create branch for your contribution:
git co -b your_new_branch_name
Make code changes.
Ensure tests pass.
Push to your fork origin.
git push origin
Submit PR from the branch on your fork in Github.
Reid Morrison :: @reidmorrison