Job Iteration API

Meet Iteration, an extension for ActiveJob that makes your jobs interruptible and resumable, saving all progress that the job has made (aka checkpoint for jobs).

Background

Imagine the following job:

class SimpleJob < ApplicationJob
  def perform
    User.find_each do |user|
      user.notify_about_something
    end
  end
end

The job would run fairly quickly when you only have a hundred User records. But as the number of records grows, it will take longer for a job to iterate over all Users. Eventually, there will be millions of records to iterate and the job will end up taking hours or even days.

With frequent deploys and worker restarts, it would mean that a job will be either lost or restarted from the beginning. Some records (especially those in the beginning of the relation) will be processed more than once.

Cloud environments are also unpredictable, and there’s no way to guarantee that a single job will have reserved hardware to run for hours and days. What if AWS diagnosed the instance as unhealthy and will restart it in 5 minutes? What if a Kubernetes pod is getting evicted? Again, all job progress will be lost. At Shopify, we also use it to interrupt workloads safely when moving tenants between shards and move shards between regions.

Software that is designed for high availability must be resilient to interruptions that come from the infrastructure. That’s exactly what Iteration brings to ActiveJob. It’s been developed at Shopify to safely process long-running jobs, in Cloud, and has been working in production since May 2017.

We recommend that you watch one of our conference talks about the ideas and history behind Iteration API.

Guides

• Iteration: how it works
• Job argument semantics
• Best practices
• Writing custom enumerator
• Throttling

For more detailed documentation, see rubydoc.

Requirements

ActiveJob is the primary requirement for Iteration. While there’s nothing that prevents it, Iteration is not yet compatible with vanilla Sidekiq API.

API

Iteration job must respond to build_enumerator and each_iteration methods. build_enumerator must return Enumerator object that respects the cursor value.

Sidekiq adapter

Unless you are running on Heroku, we recommend you to tune Sidekiq’s timeout option from the default 8 seconds to 25-30 seconds, to allow the last each_iteration to complete and gracefully shutdown.

Resque adapter

There a few configuration assumptions that are required for Iteration to work with Resque. GRACEFUL_TERM must be enabled (giving the job ability to gracefully interrupt), and FORK_PER_JOB is recommended to be disabled (set to false).

FAQ

Why can’t I just iterate in #perform method and do whatever I want? You can, but then your job has to comply with a long list of requirements, such as the ones above. This creates leaky abstractions more easily, when instead we can expose a more powerful abstraction for developers–without exposing the underlying infrastructure.

What happens when my job is interrupted? A checkpoint will be persisted to Redis after the current each_iteration, and the job will be re-enqueued. Once it’s popped off the queue, the worker will work off from the next iteration.

What happens with retries? An interruption of a job does not count as a retry. If an exception occurs, the job will retry or be discarded as normal using Active Job configuration for the job. If the job retries, it processes the iteration that originally failed and progress will continue from there on if successful.

What happens if my iteration takes a long time? We recommend that a single each_iteration should take no longer than 30 seconds. In the future, this may raise an exception.

Why is it important that each_iteration takes less than 30 seconds? When the job worker is scheduled for restart or shutdown, it gets a notice to finish remaining unit of work. To guarantee that no progress is lost we need to make sure that each_iteration completes within a reasonable amount of time.

Why do I use have to use this ugly helper in build_enumerator? Why can’t you automatically infer it? This is how the first version of the API worked. We checked the type of object returned by build_enumerable, and whether it was ActiveRecord Relation or an Array, we used the matching adapter. This caused opaque type branching in Iteration internals and it didn’t allow developers to craft their own Enumerators and control the cursor value. We made a decision to always return Enumerator instance from build_enumerator. Now we provide explicit helpers to convert ActiveRecord Relation or an Array to Enumerator, and for more complex iteration flows developers can build their own Enumerator objects.

What is the difference between Enumerable and Enumerator? We recomend this post to learn more about Enumerators in Ruby.

My job has a complex flow. How do I write my own Enumerator? Iteration API takes care of persisting the cursor (that you may use to calculate an offset) and controlling the job state. The power of Enumerator object is that you can use the cursor in any way you want. One example is a cursorless job that pops records from a datastore until the job is interrupted:

class MyJob < ApplicationJob
  include JobIteration::Iteration

  def build_enumerator(cursor:)
    Enumerator.new do
      Redis.lpop("mylist") # or: Kafka.poll(timeout: 10.seconds)
    end
  end

  def each_iteration(element_from_redis)
    # ...
  end
end

Credits

This project would not be possible without these individuals (in alphabetical order):

• Daniella Niyonkuru
• Emil Stolarsky
• Florian Weingarten
• Guillaume Malette
• Hormoz Kheradmand
• Mohamed-Adam Chaieb
• Simon Eskildsen

Development

After checking out the repo, run bin/setup to install dependencies and create mysql database. Then, run bundle exec rake test to run the tests.

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 tags, and push the .gem file to rubygems.org.

Code of Conduct

Everyone interacting in the Job::Iteration project’s codebases, issue trackers, chat rooms and mailing lists is expected to follow the code of conduct.