AcidicJob

Gem Version main workflow

Idempotent operations for Rails apps (for ActiveJob or Sidekiq)

At the conceptual heart of basically any software are "operations"—the discrete actions the software performs. Rails provides a powerful abstraction layer for building operations in the form of ActiveJob, or we Rubyists can use the tried and true power of pure Sidekiq. With either we can easily trigger from other Ruby code throughout our Rails application (controller actions, model methods, model callbacks, etc.); we can run operations both synchronously (blocking execution and then returning its response to the caller) and asychronously (non-blocking and the caller doesn't know its response); and we can also retry a specific operation if needed seamlessly.

However, in order to ensure that our operational jobs are robust, we need to ensure that they are properly idempotent and transactional. As stated in the GitLab Sidekiq Style Guide:

As a general rule, a worker can be considered idempotent if:

  • It can safely run multiple times with the same arguments.
  • Application side-effects are expected to happen only once (or side-effects of a second run do not have an effect).

This is, of course, far easier said than done. Thus, AcidicJob.

AcidicJob provides a framework to help you make your operational jobs atomic ⚛️, consistent 🤖, isolated 🕴🏼, and durable ⛰️. Its conceptual framework is directly inspired by a truly wonderful loosely collected series of articles written by Brandur Leach, which together lay out core techniques and principles required to make an HTTP API properly ACIDic:

  1. https://brandur.org/acid
  2. https://brandur.org/http-transactions
  3. https://brandur.org/job-drain
  4. https://brandur.org/idempotency-keys

AcidicJob brings these techniques and principles into the world of a standard Rails application.

Installation

Add this line to your application's Gemfile:

gem 'acidic_job'

And then execute:

$ bundle install

Or simply execute to install the gem yourself:

$ bundle add acidic_job

Then, use the following command to copy over the AcidicJob::Run migration file.

rails generate acidic_job:install

Usage

AcidicJob is a concern that you include into your base ApplicationJob.

class ApplicationJob < ActiveJob::Base
  include AcidicJob
end

This is useful because the module needs to be mixed into any and all jobs that you want to either make acidic or enqueue acidicly.

It provides a suite of functionality that empowers you to create complex, robust, and acidic jobs.

TL;DR

Key Features

  • Transactional Steps — break your job into a series of steps, each of which will be run within an acidic database transaction, allowing retries to jump back to the last "recovery point".
  • Persisted Attributes — when retrying jobs at later steps, we need to ensure that data created in previous steps is still available to later steps on retry.
  • Transactionally Staged Jobs — enqueue additional jobs within the acidic transaction safely
  • Custom Idempotency Keys — use something other than the job ID for the idempotency key of the job run
  • Sidekiq Callbacks — bring ActiveJob-like callbacks into your pure Sidekiq Workers
  • Sidekiq Batches — leverage the power of Sidekiq Pro's batch functionality without the hassle

Transactional Steps

The first and foundational feature acidic_job provides is the with_acidity method, which takes a block of transactional step methods (defined via the step) method:

class RideCreateJob < ActiveJob::Base
  include AcidicJob

  def perform(user_id, ride_params)
    user = User.find(user_id)

    with_acidity providing: { user: user, params: ride_params, ride: nil } do
      step :create_ride_and_audit_record
      step :create_stripe_charge
      step :send_receipt
    end
  end

  def create_ride_and_audit_record
    # ...
  end

  def create_stripe_charge
    # ...
  end

  def send_receipt
    # ...
  end
end

with_acidity takes only the providing: named parameter and a block where you define the steps of this operation. step simply takes the name of a method available in the job. That's all!

Now, each execution of this job will find or create an AcidicJob::Run record, which we leverage to wrap every step in a database transaction. Moreover, this database record allows acidic_job to ensure that if your job fails on step 3, when it retries, it will simply jump right back to trying to execute the method defined for the 3rd step, and won't even execute the first two step methods. This means your step methods only need to be idempotent on failure, not on success, since they will never be run again if they succeed.

Persisted Attributes

Any objects passed to the providing option on the with_acidity method are not just made available to each of your step methods, they are made available across retries. This means that you can set an attribute in step 1, access it in step 2, have step 2 fail, have the job retry, jump directly back to step 2 on retry, and have that object still accessible. This is done by serializing all objects to a field on the AcidicJob::Run and manually providing getters and setters that sync with the database record.

class RideCreateJob < ActiveJob::Base
  include AcidicJob

  def perform(ride_params)
    with_acidity providing: { ride: nil } do
      step :create_ride_and_audit_record
      step :create_stripe_charge
      step :send_receipt
    end
  end

  def create_ride_and_audit_record
    self.ride = Ride.create!
  end

  def create_stripe_charge
    Stripe::Charge.create(amount: 20_00, customer: self.ride.user)
  end

  # ...
end

Note: This does mean that you are restricted to objects that can be serialized by ActiveRecord, thus no Procs, for example.

Note: You will note the use of self.ride = ... in the code sample above. In order to call the attribute setter method that will sync with the database record, you must use this style. @ride = ... and/or ride = ... will both fail to sync the value with the database record.

Transactionally Staged Jobs

A standard problem when inside of database transactions is enqueuing other jobs. On the one hand, you could enqueue a job inside of a transaction that then rollbacks, which would leave that job to fail and retry and fail. On the other hand, you could enqueue a job that is picked up before the transaction commits, which would mean the records are not yet available to this job.

In order to mitigate against such issues without forcing you to use a database-backed job queue, acidic_job provides perform_acidicly and deliver_acidicly methods to "transactionally stage" enqueuing other jobs from within a step (whether another ActiveJob or a Sidekiq::Worker or an ActionMailer delivery). These methods will create a new AcidicJob::Run record, but inside of the database transaction of the step. Upon commit of that transaction, a model callback pushes the job to your actual job queue. Once the job has been successfully performed, the AcidicJob::Run record is deleted so that this table doesn't grow unbounded and unnecessarily.

class RideCreateJob < ActiveJob::Base
  include AcidicJob

  def perform(user_id, ride_params)
    user = User.find(user_id)

    with_acidity providing: { user: user, params: ride_params, ride: nil } do
      step :create_ride_and_audit_record
      step :create_stripe_charge
      step :send_receipt
    end
  end

  # ...

  def send_receipt
    RideMailer.with(user: @user, ride: @ride).confirm_charge.delivery_acidicly
  end
end

Custom Idempotency Keys

By default, AcidicJob uses the job identifier provided by the queueing system (ActiveJob or Sidekiq) as the idempotency key for the job run. The idempotency key is what is used to guarantee that no two runs of the same job occur. However, sometimes we need particular jobs to be idempotent based on some other criteria. In these cases, AcidicJob provides a collection of tools to allow you to ensure the idempotency of your jobs.

Firstly, you can configure your job class to explicitly use either the job identifier or the job arguments as the foundation for the idempotency key. A job class that calls the acidic_by_job_id class method (which is the default behavior) will simply make the job run's idempotency key the job's identifier:

class ExampleJob < ActiveJob::Base
  include AcidicJob
  acidic_by_job_id

  def perform
  end
end

Conversely, a job class can use the acidic_by_job_args method to configure that job class to use the arguments passed to the job as the foundation for the job run's idempotency key:

class ExampleJob < ActiveJob::Base
  include AcidicJob
  acidic_by_job_args

  def perform(arg_1, arg_2)
    # the idempotency key will be based on whatever the values of `arg_1` and `arg_2` are
  end
end

These options cover the two common situations, but sometimes our systems need finer-grained control. For example, our job might take some record as the job argument, but we need to use a combination of the record identifier and record status as the foundation for the idempotency key. In these cases you can pass a Proc to an acidic_by class method:

class ExampleJob < ActiveJob::Base
  include AcidicJob
  acidic_by ->(record:) { [record.id, record.status] }

  def perform(record:)
    # the idempotency key will be based on whatever the values of `record.id` and `record.status` are
  end
end

Note: The signature of the acidic_by proc needs to match the signature of the job's perform method.

Sidekiq Callbacks

In order to ensure that AcidicJob::Staged records are only destroyed once the related job has been successfully performed, whether it is an ActiveJob or a Sidekiq Worker, acidic_job also extends Sidekiq to support the ActiveJob callback interface.

This allows acidic_job to use an after_perform callback to delete the AcidicJob::Staged record, whether you are using the gem with ActiveJob or pure Sidekiq Workers. Of course, this means that you can add your own callbacks to any jobs or workers that include the AcidicJob module as well.

Sidekiq Batches

One final feature for those of you using Sidekiq Pro: an integrated DSL for Sidekiq Batches. By simply adding the awaits option to your step declarations, you can attach any number of additional, asynchronous workers to your step. This is profoundly powerful, as it means that you can define a workflow where step 2 is started if and only if step 1 succeeds, but step 1 can have 3 different workers enqueued on 3 different queues, each running in parallel. Once all 3 workers succeed, acidic_job will move on to step 2. That's right, by leveraging the power of Sidekiq Batches, you can have workers that are executed in parallel, on separate queues, and asynchronously, but are still blocking—as a group—the next step in your workflow! This unlocks incredible power and flexibility for defining and structuring complex workflows and operations, and in my mind is the number one selling point for Sidekiq Pro.

In my opinion, any commercial software using Sidekiq should get Sidekiq Pro; it is absolutely worth the money. If, however, you are using acidic_job in a non-commercial application, you could use the open-source dropin replacement for this functionality: https://github.com/breamware/sidekiq-batch

class RideCreateJob < ActiveJob::Base
  include AcidicJob

  def perform(user_id, ride_params)
    user = User.find(user_id)

    with_acidity providing: { user: user, params: ride_params, ride: nil } do
      step :create_ride_and_audit_record, awaits: [SomeJob]
      step :create_stripe_charge, args: [1, 2, 3], kwargs: { some: 'thing' }
      step :send_receipt
    end
  end
end

Testing

When testing acidic jobs, you are likely to run into ActiveRecord::TransactionIsolationErrors:

ActiveRecord::TransactionIsolationError: cannot set transaction isolation in a nested transaction

This error is thrown because by default RSpec and most MiniTest test suites use database transactions to keep the test database clean between tests. The database transaction that is wrapping all of the code executed in your test is run at the standard isolation level, but acidic jobs then try to create another transaction run at a more conservative isolation level. You cannot have a nested transaction that runs at a different isolation level, thus, this error.

In order to avoid this error, you need to ensure firstly that your tests that run your acidic jobs are not using a database transaction and secondly that they use some different strategy to keep your test database clean. The DatabaseCleaner gem is a commonly used tool to manage different strategies for keeping your test database clean. As for which strategy to use, truncation and deletion are both safe, but their speed varies based on our app's table structure (see https://github.com/DatabaseCleaner/database_cleaner#what-strategy-is-fastest). Either is fine; use whichever is faster for your app.

In order to make this test setup simpler, AcidicJob provides a TestCase class that your MiniTest jobs tests can inherit from. It is simple; it inherits from ActiveJob::TestCase, sets use_transactional_tests to false, and ensures DatabaseCleaner is run for each of your tests. Moreover, it ensures that the system's original DatabaseCleaner configuration is maintained, options included, except that any transaction strategies for any ORMs are replaced with a deletion strategy. It does so by storing whatever the system DatabaseCleaner configuration is at the start of before_setup phase in an instance variable and then restores that configuration at the end of after_teardown phase. In between, it runs the configuration thru a pipeline that selectively replaces any transaction strategies with a corresponding deletion strategy, leaving any other configured strategies untouched.

For those of you using RSpec, you can require the acidic_job/rspec_configuration file, which will configure RSpec in the exact same way I have used in my RSpec projects to allow me to test acidic jobs with either the deletion strategy but still have all of my other tests use the fast transaction strategy:

require "database_cleaner/active_record"

# see https://github.com/DatabaseCleaner/database_cleaner#how-to-use
RSpec.configure do |config|
  config.use_transactional_fixtures = false

  config.before(:suite) do
    DatabaseCleaner.clean_with :truncation

    # Here we are defaulting to :transaction but swapping to deletion for some specs;
    # if your spec or its code-under-test uses
    # nested transactions then specify :transactional e.g.:
    #   describe "SomeWorker", :transactional do
    #
    DatabaseCleaner.strategy = :transaction

    config.before(:context, transactional: true) { DatabaseCleaner.strategy = :deletion }
    config.after(:context, transactional: true) { DatabaseCleaner.strategy = :transaction }
    config.before(:context, type: :system) { DatabaseCleaner.strategy = :deletion }
    config.after(:context, type: :system) { DatabaseCleaner.strategy = :transaction }
  end

  config.around(:each) do |example|
    DatabaseCleaner.cleaning do
      example.run
    end
  end
end

Development

After checking out the repo, run bin/setup to install dependencies. Then, run rake test to run the tests. You can also run bin/console for an interactive prompt that will allow you to experiment.

To install this gem onto your local machine, run bundle exec rake install. To release a new version, update the version number in version.rb, and then run bundle exec rake release, which will create a git tag for the version, push git commits and the created tag, and push the .gem file to rubygems.org.

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/[USERNAME]/acidic_job.