Sidekiq 7.0 Capsules

Quick Jump

Sidekiq 7.0 Capsules
The Problem
The Solution

Sidekiq 7.0 contains the largest internal refactoring since Sidekiq 4.0. This refactoring is designed to improve deployment flexibility and allow new use cases.

The Problem

Before 7.0, Sidekiq used a large number of global methods on the Sidekiq module to access things like the Redis connection pool, the logger, and process configuration, e.g.

ruby

Sidekiq.logger.info "Hello world"
Sidekiq.redis {|c| c.sadd("some_set", "new_member") }
Sidekiq.configure_server {|config| config... }

The issue is that this pattern implies a global, mutable singleton. It does not work with Ractors. It does not allow multiple instances in one process. It does not allow embedding within another Ruby process (e.g. puma). Today the only supported Sidekiq deployment pattern is running bundle exec sidekiq.

The Solution

Sidekiq 7.0 aims to refactor Sidekiq internals to allow more flexibility in how Sidekiq can be used.

Sidekiq::Config

Before, all Sidekiq configuration went through the Sidekiq module and was stored in the top-level hash at Sidekiq.options. Now Sidekiq::CLI creates a Sidekiq::Config object which holds the global configuration at Sidekiq.default_configuration. This instance is now passed into Sidekiq.configure_{client,server} do |config|

Sidekiq::Capsule

Sidekiq::Capsule represents the set of resources necessary to process a set of queues. By default, Sidekiq::CLI creates one Sidekiq::Capsule instance and mutates it according to the command line parameters and the data in config/sidekiq.yml.

You create additional Capsules within your initializer, like so:

ruby

Sidekiq.configure_server do |config|
  config.capsule("single-threaded") do |cap|
    cap.concurrency = 1
    cap.queues = %w[single] # to specify weights: %w[important,10 mailers,7]
  end
end

Capsules can have their own customized middleware chains but by default will inherit the global middleware configuration. Each Capsule will have its own Redis connection pool sized to the configured concurrency.

Sidekiq::Launcher is the top-level component which takes a Sidekiq::Config and launches the tree of runtime components for each capsule. Once passed to Launcher, the global Config and each Capsule should be considered frozen and immutable.

Every internal component of Sidekiq takes a Sidekiq::Capsule instance and uses it. The Capsule holds previously "global" state like the connection pool, error handlers, lifecycle callbacks, etc.

There is still one iron-clad rule: a Sidekiq process only executes jobs from one Redis instance; all Capsules within a process must use the same Redis instance. If you want to process jobs from two separate Redis instances, you need to start two separate Sidekiq processes.

Use Cases

With Capsules, you can programmatically tune how a Sidekiq process handles specific queues. One Capsule can use 1 thread to process jobs within a thread_unsafe queue while another Capsule uses 10 threads to process default jobs.

ruby

# within your initializer
Sidekiq.configure_server do |config|
  config.capsule("unsafe") do |capsule|
    capsule.queues = %w(thread_unsafe)
    capsule.concurrency = 1
  end
end

The contents of config/sidekiq.yml configure the default capsule.

Redis Pools

Before 7.0, the Sidekiq process would create one redis pool sized to concurrency + 5. Now Sidekiq will create multiple Redis pools: an internal pool of ten connections available to Sidekiq components along with a pool of concurrency for the job processors within each Capsule.

For a Sidekiq process with a default Capsule and a single threaded Capsule, you should have three Redis pools of size 5, 10 and 1. Remember that connection pools are lazy so it won't create all those connections unless they are actively needed.

Sidekiq components and add-ons should avoid using Sidekiq.redis or Sidekiq.logger. Instead use the implicit redis or logger methods available on Sidekiq::Component, Sidekiq::Capsule or Sidekiq::{Client,Server}Middleware. Jobs may continue to use Sidekiq.redis, they will automagically use the capsule's pool of concurrency connections.

Sidekiq::Component

Sidekiq::Component is a module which provides helpful methods based on a config reader:

ruby

module Sidekiq::Component
  def config
    @config
  end

  def redis(&block)
    config.redis(&block)
  end

  def logger
    config.logger
  end

  def handle_exception(ex, ctx)
    # avoids calling `Sidekiq.error_handlers...`
    config.handle_exception(ex, ctx)
  end
end

class Sidekiq::Processor
  include Sidekiq::Component

  def initialize(capsule)
    @config = capsule
  end

  def ...
    # old
    Sidekiq.redis {|c| ... }
    Sidekiq.logger.info "Hello world!"

    # new
    redis {|c| ... }
    logger.info "Hello world!"
  rescue => ex
    handle_exception(ex, ...)
  end
end

Sidekiq::Capsule overrides Sidekiq::Config in order to provide Capsule-local resources; you'll see places within Sidekiq where Capsule acts like a Config. Code running within the Capsule will use resources local to that Capsule.

With this pattern, we greatly reduce the use of global APIs throughout Sidekiq internals. Where before we'd call Sidekiq.xyz, we instead provide similar functionality like config.xyz.