Skip to content

a multi-threaded, SQL-driven ActiveJob backend used at Betterment to process millions of background jobs per day

License

Notifications You must be signed in to change notification settings

Betterment/delayed

Repository files navigation

Delayed

Gem Version CI

Delayed is a multi-threaded, SQL-driven ActiveJob backend used at Betterment to process millions of background jobs per day.

It supports postgres, mysql, and sqlite, and is designed to be:

  • Reliable, with co-transactional job enqueues and guaranteed, at-least-once execution
  • Scalable, with an optimized pickup query and concurrent job execution
  • Resilient, with built-in retry mechanisms, exponential back-off, and failed job preservation
  • Maintainable, with robust instrumentation, continuous monitoring, and priority-based alerting

For an overview of how Betterment uses delayed to build resilience into distributed systems, read the announcement blog post, and/or check out the talk ✨Can I break this?✨ given at RailsConf 2021!

Why Delayed?

The delayed gem is a targeted fork of both delayed_job and delayed_job_active_record, combining them into a single library. It is designed for applications with the kinds of operational needs seen at Betterment, and includes numerous features extracted from Betterment's codebases, such as:

  • Multithreaded job execution via concurrent-ruby
  • A highly optimized, SKIP LOCKED-based pickup query (on postgres)
  • Built-in instrumentation and continuous monitoring via a new monitor process
  • Named priority ranges, defaulting to :interactive, :user_visible, :eventual, and :reporting
  • Priority-based alerting threshholds for job age, run time, and attempts
  • An experimental autoscaling metric, for use by a horizontal autoscaler (we use Kubernetes)
  • A custom adapter that extends ActiveJob with Delayed-specific behaviors

This gem benefits immensely from the many years of development, maintenance, and community support that have gone into delayed_job, and many of the core DJ APIs (like .delay) are still available in delayed as undocumented features. Over time, these APIs may be removed as this gem focuses itself around ActiveJob-based usage, but the aim will be to provide bidirectional migration paths where possible.

Table of Contents

Getting Started

This gem is designed to work with Rails 5.2+ and Ruby 2.6+ on postgres 9.5+ or mysql 5.6+

Installation

Add the following to your Gemfile:

gem 'delayed'

Then run bundle install.

Before you can enqueue and run jobs, you will need a jobs table. You can create this table by running the following command:

rails generate delayed:migration
rails db:migrate

Then, to use this background job processor with ActiveJob, add the following to your application config:

config.active_job.queue_adapter = :delayed

See the Rails guide for more details.

Basic Usage

Running a worker process

In order for any jobs to execute, you must first start a worker process, which will work off jobs:

rake delayed:work

By default, a worker process will pick up 2 jobs at a time (ordered by priority) and run each in a separate thread. To change the number of jobs picked up (and, in turn, increase the size of the thread pool), use the MAX_CLAIMS environment variable:

MAX_CLAIMS=5 rake delayed:work

Work off specific queues by setting the QUEUE or QUEUES environment variable:

QUEUE=tracking rake delayed:work
QUEUES=mailers,tasks rake delayed:work

You can stop the worker with CTRL-C or by sending a SIGTERM signal to the process. The worker will attempt to complete outstanding jobs and gracefully shutdown. Some platforms (like Heroku) will send a SIGKILL after a designated timeout, which will immediately terminate the process and may result in long-running jobs remaining locked until Delayed::Worker.max_run_time has elapsed. (By default this is 20 minutes.)

Enqueuing Jobs

The recommended usage of this gem is via ActiveJob. You can define a job like so:

def MyJob < ApplicationJob
  def perform(any: 'arguments')
    # do something here
  end
end

Then, enqueue the job with perform_later:

MyJob.perform_later(arguments: 'go here')

Jobs will be enqueued to the delayed_jobs table, which can be accessed via the Delayed::Job ActiveRecord model using standard ActiveRecord query methods (.find, .where, etc).

To override specific columns or parameters of the job, use set:

MyJob.set(priority: 11).perform_later(some_more: 'arguments')
MyJob.set(queue: 'video_encoding').perform_later(video)
MyJob.set(wait: 3.hours).perform_later
MyJob.set(wait_until: 1.day.from_now).perform_later

Priority ranges are mapped to configurable shorthand names:

MyJob.set(priority: :interactive).perform_later
MyJob.set(priority: :user_visible).perform_later
MyJob.set(priority: :eventual).perform_later
MyJob.set(priority: :reporting).perform_later

Delayed::Job.last.priority.user_visible? # => false
Delayed::Priority.new(99).reporting? # => true
Delayed::Priority.new(11).to_i # => 11
Delayed::Priority.new(3).to_s # => 'interactive'

To change the default priority names, or to adjust other aspects of job execution, see the Configuration section below.

Other ActiveJob Features

All other ActiveJob features should work out of the box, such as the queue_as and queue_with_priority class-level directives:

class MyJob < ApplicationJob
  queue_as 'some_other_queue'
  queue_with_priority 42

  # ...
end

ActiveJob also supports the following lifecycle hooks:

Read more about ActiveJob usage on the Active Job Basics documentation page.

Operational Considerations

Delayed has been shaped around Betterment's day-to-day operational needs. In order to benefit from these design decisions, there are a few things you'll want to keep in mind.

Co-transactionality

The :delayed job backend is designed for co-transactional job enqueues. This means that you can safely enqueue jobs inside of ACID-compliant business operations, like so:

def save
  ActiveRecord::Base.transaction do
    user.lock!

    if user.update(email: new_email)
      EmailChangeJob.perform_later(user, new_email, old_email)

      true
    else
      false
    end
  end
end

If the transaction rolls back, the enqueued job will also roll back, ensuring that the entire operation is all-or-nothing. A job will never become visible to a worker until the transaction commits.

Important: the above assumes that the connection used by the transaction is the one provided by ActiveRecord::Base. (Support for enqueuing jobs via other database connections is possible, but is not yet exposed as a configuration.)

At-Least-Once Delivery

Each job is guaranteed to run at least once, but under certain conditions may run more than once. As such, you'll want to ensure that your jobs are idempotent, meaning they can be safely repeated, regardless of the outcome of any prior attempts.

When Jobs Fail

Unlike other job queue backends, delayed will not delete failing jobs by default. These are jobs that have reached their max_attempts (25 by default), and they will remain in the queue until you manually intervene.

The general idea is that you should treat these as operational issues (like an error on your bugtracker), and you should aim to resolve the issue by making the job succeed. This might involve shipping a bugfix, making a data change, or updating the job's implementation to handle certain corner cases more gracefully (perhaps by no-opping). When you're ready to re-run, you may clear the failed_at column and reset attempts to 0:

Delayed::Job.find(failing_job_id).update!(failed_at: nil, attempts: 0, run_at: Time.zone.now)

Monitoring Jobs & Workers

Delayed will emit ActiveSupport::Notifications at various points during job and worker lifecycles, and can also be configured for continuious monitoring. You are strongly encouraged to tie these up to your preferred application monitoring solution by calling ActiveSupport::Notification.subscribe in an initializer.

Lifecycle Hooks

The following events will be emitted automatically by workers as jobs are reserved and performed:

  • delayed.job.run - an event measuring the duration of a job's execution
  • delayed.job.error - an event indicating that a job has errored and may be retried (no duration attached)
  • delayed.job.failure - an event indicating that a job has permanently failed (no duration attached)
  • delayed.job.enqueue - an event measuring the time it takes to enqueue a job
  • delayed.worker.reserve_jobs - an event measuring the duration of the job "pickup query"

The "run", "error", "failure" and "enqueue" events will include a :job argument in the event's payload, providing access to the job instance.

ActiveSupport::Notifications.subscribe('delayed.job.run') do |*args|
  event = ActiveSupport::Notifications::Event.new(*args)

  # Emit the event via your preferred metrics/instrumentation provider:
  tags = event.payload.except(:job).map { |k,v| "#{k.to_s[0..64]}:#{v.to_s[0..255]}" }
  StatsD.distribution(event.name, event.duration, tags: tags)
end

ActiveSupport::Notifications.subscribe(/delayed\.job\.(error|failure)/) do |*args|
  # ...
  Statsd.increment(...)
end

ActiveSupport::Notifications.subscribe('delayed.job.enqueue') do |*args|
  # ...
  StatsD.distribution(...)
end

ActiveSupport::Notifications.subscribe('delayed.worker.reserve_jobs') do |*args|
  # ...
  StatsD.distribution(...)
end

Priority-based Alerting Threshholds

By default, jobs support "alerting threshholds" that allow them to warn if they come within range of max_run_time or max_attempts (without exceeding them), or if they spend too long waiting in the queue (i.e. their "age").

The threshholds are fully configurable, and default to the following values:

Delayed::Priority.alerts = {
  interactive: { age: 1.minute, run_time: 30.seconds, attempts: 3 },
  user_visible: { age: 3.minutes, run_time: 90.seconds, attempts: 5 },
  eventual: { age: 1.5.hours, run_time: 5.minutes, attempts: 8 },
  reporting: { age: 4.hours, run_time: 10.minutes, attempts: 8 },
}

These may also be configured on a per-job basis:

class MyVeryHighThroughputJob < ApplicationJob
  # ...

  def alert_run_time
    5.seconds # must execute in under 5 seconds
  end

  def alert_attempts
    1 # will begin alerting after 1 attempt
  end
end

If a job completes but was uncomfortably close to timing-out, it may make sense to emit an alert:

ActiveSupport::Notifications.subscribe('delayed.job.run') do |_name, _start, _finish, _id, payload|
  job = payload[:job]
  TeamAlerter.alert!("Job with ID #{job.id} took #{job.run_time} seconds to run") if job.run_time_alert?
end

Similarly, if a job is erroring repeatedly, you may choose to emit some form of notification before it reaches its full attempt count:

ActiveSupport::Notifications.subscribe('delayed.job.error') do |_name, _start, _finish, _id, payload|
  job = payload[:job]
  TeamAlerter.alert!("Job with ID #{job.id} has made #{job.attempts} attempts") if job.attempts_alert?
end

The last threshhold (job.age_alert?) refers to the time spent in the queue, and may be best monitored in aggregate (covered in the next section!), as it generally describes the ability of workers to pick up jobs fast enough.

Continuous Monitoring

To continuously monitor the state of your job queues, you may run a single "monitor" process alongside your workers. (Only one instance of this process is needed, as it will emit aggregate metrics.)

rake delayed:monitor

The monitor process accepts the same queue configurations as the worker process, and can be used to monitor the same sets of queues as the workers:

QUEUE=tracking rake delayed:monitor
QUEUES=mailers,tasks rake delayed:monitor

The following events will be emitted, grouped by priority name (e.g. "interactive") and queue name, and the metric's ":value" will be available in the event's payload. This means that there will be one value per unique combination of queue & priority, and totals must be computed via downstream aggregation (e.g. as a StatsD "gauge" metric).

  • delayed.job.count - the total number of jobs
  • delayed.job.future_count - jobs where run_at is in the future
  • delayed.job.working_count - jobs that are currently being worked off (excludes failed jobs)
  • delayed.job.workable_count - jobs that are waiting to be worked off
  • delayed.job.erroring_count - jobs where attempts > 0
  • delayed.job.failed_count - jobs where failed_at is not nil
  • delayed.job.max_lock_age - the age of the oldest locked_at value (excludes failed jobs)
  • delayed.job.max_age - the age of the oldest run_at value (excludes failed jobs)

An additional experimental metric is available, intended for use with application autoscaling:

  • delayed.job.alert_age_percent - the percent to which the oldest job has reached the "age alert" threshold. (See the Alerting Threshholds section above.)

All of these events may be subscribed to via a single regular expression (again, in your application config or in an initializer):

ActiveSupport::Notifications.subscribe(/delayed\.job\..*_(count|age|percent)/) do |*args|
  event = ActiveSupport::Notifications::Event.new(*args)
  value = event.payload.delete(:value)

  # Emit the event via your preferred metrics/instrumentation provider:
  tags = event.payload.map { |k,v| "#{k.to_s[0..64]}:#{v.to_s[0..255]}" }
  StatsD.gauge(event.name, value, sample_rate: 1.0, tags: tags)
end

Additionally, the monitor process with emit a delayed.monitor.run event with a duration attached, so that you can monitor the time it takes to emit these aggregate metrics.

ActiveSupport::Notifications.subscribe('delayed.monitor.run') do |*args|
  # ...
  StatsD.distribution(...)
end

Configuration

Delayed is highly configurable, but ships with opinionated defaults. If you need to change any default behaviors, you can do so in an initializer (e.g. config/initializers/delayed.rb).

By default, workers will claim 5 jobs at a time (run in concurrent threads). If no jobs are found, workers will sleep for 5 seconds.

# The max number of jobs a worker may lock at a time (also the size of the thread pool):
Delayed::Worker.max_claims = 5

# The number of jobs to which a worker may "read ahead" when locking jobs (mysql only!):
Delayed::Worker.read_ahead = 5

# If a worker finds no jobs, it will sleep this number of seconds in between attempts:
Delayed::Worker.sleep_delay = 5

# Until version 1.0, the worker will not sleep at all between attemps if it finds jobs.
# This can be configured by setting the minimum reserve interval:
Delayed::Worker.min_reserve_interval = 0.5 # seconds

If a job fails, it will be rerun up to 25 times (with an exponential back-off). Jobs will also time-out after 20 minutes.

# The max number of attempts jobs are given before they are permanently marked as failed:
Delayed::Worker.max_attempts = 25

# The max amount of time a job is allowed to run before it is stopped:
Delayed::Worker.max_run_time = 20.minutes

Individual jobs may specify their own max_attempts and max_run_time:

class MyJob < ApplicationJob
  def perform; end

  def max_run_time
    15.minutes # must be less than the global `max_run_time` default!
  end

  def max_attempts
    1
  end
end

By default, workers will work off all queues (including nil), and jobs will be enqueued to a 'default' queue.

# A list of queues to which all work is restricted. (e.g. `%w(queue1 queue2 queue3)`)
# If no queues are specified, then all queues will be worked off
Delayed::Worker.queues = []

# The default queue that jobs will be enqueued to, when no other queue is specified:
Delayed::Worker.default_queue_name = 'default'

Priority ranges are given names. These will default to "interactive" for 0-9, "user visible" for 10-19, "eventual" for 20-29, and "reporting" for 30+. The default priority for enqueued jobs is "user visible" (10), and workers will work off all priorities, unless otherwise configured.

# Default priority names, useful for enqueuing and for instrumentation/metrics.
Delayed::Priority.names = { interactive: 0, user_visible: 10, eventual: 20, reporting: 30 }

# The default priority for enqueued jobs, when no priority is specified.
# This aligns with the "user_visible" named priority.
Delayed::Worker.default_priority = 10

# A worker can also be told to work off specific priority ranges,
# if, say, you'd like a dedicated worker for high priority jobs:
Delayed::Worker.min_priority = nil
Delayed::Worker.max_priority = nil

Job priorities can specified by using the name of the desired range (i.e. :user_visible). By default, the value for a named priority will be the first value in that range. To set each priority's default value to the middle of its range (i.e. 15 for :user_visible), Delayed::Priority can be configured with:

Delayed::Priority.assign_at_midpoint = true

Logging verbosity is also configurable. The gem will attempt to default to Rails.logger with an "info" log level.

# Specify an alternate logger class:
Delayed.logger = Rails.logger

# Specify a default log level for all job lifecycle logging:
Delayed.default_log_level = 'info'

Migrating from other ActiveJob backends

For the most part, standard ActiveJob APIs should be fully compatible. However, when migrating from a Redis-backed queue (or some other queue that is not co-located with your ActiveRecord data), the Operational Considerations section of this README should be noted. You may wish to change the way that jobs are enqueued and executed in order to benefit from co-transactional / ACID guarantees.

To assist in migrating, you are encouraged to set queue_adapter on a per-job basis, so that you can move and monitor fewer job classes at a time:

class NewsletterJob < ApplicationJob
  self.queue_adapter = :sidekiq
end

class OrderPurchaseJob < ApplicationJob
  self.queue_adapter = :delayed
end

Migrating from DelayedJob

If you choose to use delayed in an app that was originally written against delayed_job, several non-ActiveJob APIs are still available. These include "plugins", lifecycle hooks, and the .delay and .handle_asynchronously methods. These APIs are intended to assist in migrating older codebases onto ActiveJob, and may eventually be removed or extracted into an optional gem.

For comprehensive information on the APIs and features that delayed has inherited from delayed_job and delayed_job_active_record, refer to DelayedJob's documentation.

When migrating from delayed_job, you may choose to manually apply its default configurations:

Delayed::Worker.max_run_time = 4.hours
Delayed::Worker.default_priority = 0
Delayed::Worker.default_queue_name = nil
Delayed::Worker.destroy_failed_jobs = true # WARNING: This will irreversably delete jobs.

Note that some configurations, like queue_attributes, exit_on_complete, backend, and raise_signal_exceptions have been removed entirely.

How to Contribute

We would love for you to contribute! Anything that benefits the majority of users—from a documentation fix to an entirely new feature—is encouraged.

Before diving in, check our issue tracker and consider creating a new issue to get early feedback on your proposed change.

Suggested Workflow

  • Fork the project and create a new branch for your contribution.
  • Write your contribution (and any applicable test coverage).
  • Make sure all tests pass (bundle exec rake).
  • Submit a pull request.

About

a multi-threaded, SQL-driven ActiveJob backend used at Betterment to process millions of background jobs per day

Resources

License

Stars

Watchers

Forks

Languages