Resque Pool

Build Status Depfu Gem

Resque pool is a daemon for managing a pool of resque workers. With a simple config file, it manages your workers for you, starting up the appropriate number of workers for each worker type.

Benefits

  • Less config - With a simple YAML file, you can start up a pool daemon, and it will monitor your workers for you.
  • Less memory - If you are using ruby 2.0+ (with copy-on-write safe garbage collection), this should save you a lot of memory when you are managing many workers.
  • Faster startup - when you start many workers at once, they would normally compete for CPU as they load their environments. Resque-pool can load your application once, then rapidly fork the workers after setup. If a worker crashes or is killed, a new worker will start up and take its place right away.

Upgrading?

See Changelog.md in case there are important or helpful changes.

How to use

YAML file config

Create a config/resque-pool.yml (or resque-pool.yml) with your worker counts. The YAML file supports both using root level defaults as well as environment specific overrides (RACK_ENV, RAILS_ENV, and RESQUE_ENV environment variables can be used to determine environment). For example in config/resque-pool.yml:

foo: 1
bar: 2
"foo,bar,baz": 1

production:
  "foo,bar,baz": 4

Rake task config

Require the rake tasks (resque/pool/tasks) in your Rakefile, load your application environment, configure Resque as necessary, and configure resque:pool:setup to disconnect all open files and sockets in the pool manager and reconnect in the workers. For example, with rails you should put the following into lib/tasks/resque.rake:

require 'resque/pool/tasks'

# this task will get called before resque:pool:setup
# and preload the rails environment in the pool manager
task "resque:setup" => :environment do
  # generic worker setup, e.g. Hoptoad for failed jobs
end

task "resque:pool:setup" do
  # close any sockets or files in pool manager
  ActiveRecord::Base.connection.disconnect!
  # and re-open them in the resque worker parent
  Resque::Pool.after_prefork do |job|
    ActiveRecord::Base.establish_connection
  end
end

For normal work with fresh resque and resque-scheduler gems add next lines in lib/rake/resque.rake

task "resque:pool:setup" do
  Resque::Pool.after_prefork do |job|
    Resque.redis.client.reconnect
  end
end

Start the pool manager

Then you can start the queues via:

resque-pool --daemon --environment production

This will start up seven worker processes, one exclusively for the foo queue, two exclusively for the bar queue, and four workers looking at all queues in priority. With the config above, this is similar to if you ran the following:

rake resque:work RAILS_ENV=production QUEUES=foo &
rake resque:work RAILS_ENV=production QUEUES=bar &
rake resque:work RAILS_ENV=production QUEUES=bar &
rake resque:work RAILS_ENV=production QUEUES=foo,bar,baz &
rake resque:work RAILS_ENV=production QUEUES=foo,bar,baz &
rake resque:work RAILS_ENV=production QUEUES=foo,bar,baz &
rake resque:work RAILS_ENV=production QUEUES=foo,bar,baz &

The pool manager will stay around monitoring the resque worker parents, giving three levels: a single pool manager, many worker parents, and one worker child per worker (when the actual job is being processed). For example, ps -ef f | grep [r]esque (in Linux) might return something like the following:

resque    13858     1  0 13:44 ?        S      0:02 resque-pool-manager: managing [13867, 13875, 13871, 13872, 13868, 13870, 13876]
resque    13867 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Waiting for foo
resque    13868 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Waiting for bar
resque    13870 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Waiting for bar
resque    13871 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Waiting for foo,bar,baz
resque    13872 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Forked 7481 at 1280343254
resque     7481 13872  0 14:54 ?        S      0:00      \_ resque-1.9.9: Processing foo since 1280343254
resque    13875 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Waiting for foo,bar,baz
resque    13876 13858  0 13:44 ?        S      0:00  \_ resque-1.9.9: Forked 7485 at 1280343255
resque     7485 13876  0 14:54 ?        S      0:00      \_ resque-1.9.9: Processing bar since 1280343254

Running as a daemon will default to placing the pidfile and logfiles in the conventional rails locations, although you can configure that. See resque-pool --help for more options.

SIGNALS

The pool manager responds to the following signals:

  • HUP - reset config loader (reload the config file), reload logfiles, restart all workers.
  • QUIT - gracefully shut down workers (via QUIT) and shutdown the manager after all workers are done.
  • INT - gracefully shut down workers (via QUIT) and immediately shutdown manager
  • TERM - immediately shut down workers (via INT) and immediately shutdown manager (configurable via command line options)
  • WINCH - (only when running as a daemon) send QUIT to each worker, but keep manager running (send HUP to reload config and restart workers)
  • USR1/USR2/CONT - pass the signal on to all worker parents (see Resque docs).

Use HUP to help logrotate run smoothly and to change the number of workers per worker type. Signals can be sent via the kill command, e.g. kill -HUP $master_pid

If the environment variable TERM_CHILD is set, QUIT and TERM will respond as defined by Resque 1.22 and above. See http://hone.herokuapp.com/resque/2012/08/21/resque-signals.html for details, overriding any command-line configuration for TERM. Setting TERM_CHILD tells us you know what you're doing.

Custom Configuration Loader

If the static YAML file configuration approach does not meet your needs, you can specify a custom configuration loader.

Set the config_loader class variable on Resque::Pool to an object that responds to #call (which can simply be a lambda/Proc). The class attribute needs to be set before starting the pool. This is usually accomplished in the resque:pool:setup rake task, as described above.

For example, if you wanted to vary the number of worker processes based on a value stored in Redis, you could do something like:

task "resque:pool:setup" do
  Resque::Pool.config_loader = lambda do |env|
    worker_count = Redis.current.get("pool_workers_#{env}").to_i
    {"queueA,queueB" => worker_count }
  end
end

In order to query ActiveRecord, be sure establish the database connection:

task "resque:pool:setup" do                                           
  Resque::Pool.config_loader = lambda do |env|
    ActiveRecord::Base.establish_connection
    JobQueue.pluck(:name, :workers).to_h
  end
end

The configuration loader's #call method will be invoked about once a second. This allows the configuration to constantly change, allowing you to scale the number of workers up or down based on different conditions. If the response is generally static, the loader may want to cache the value it returns. It can optionally implement a #reset! method, which will be invoked when the HUP signal is received, allowing the loader to flush its cache, or perform any other re-initialization.

You can reduce the frequency that your configuration loader is called by wrapping it with Resque::Pool::ConfigLoaders::Throttled and specifying a time (in seconds) to cache the previous value:

task "resque:pool:setup" do
  redis_loader = lambda do |env|
    worker_count = Redis.current.get("pool_workers_#{env}").to_i
    { "queueA,queueB" => worker_count }
  end

  # calls through to redis_loader at most once per 10 seconds
  Resque::Pool.config_loader = Resque::Pool::ConfigLoaders::Throttled.new(
    redis_loader, period: 10
  )
end

See the spec for more details.

Zero-downtime code deploys

In a production environment you will likely want to manage the daemon using a process supervisor like runit or god or an init system like systemd or upstart. Example configurations for some of these are included in the examples directory. With these systems, reload typically sends a HUP signal, which will reload the configuration but not application code. The simplest way to make workers pick up new code after a deploy is to stop and start the daemon. This will result in a period where new jobs are not being processed.

To avoid this downtime, leave the old pool running and start a new pool with resque-pool --hot-swap.

The --hot-swap flag will turn off pidfiles (so multiple pools can run at once), enable a lock file (so your init system knows when the pool is running), and shut down other pools after the workers have started for this pool. These behaviors can also be configured separately (see resque-pool --help). The upstart configs in the examples directory demonstrate how to supervise a daemonized pool with hot swap.

Please be aware that this approach uses more memory than a simple restart, since two copies of the application code are loaded at once. TODO: #139

Other Features

You can also start a pool manager via rake resque:pool or from a plain old ruby script by calling Resque::Pool.run.

Workers will watch the pool manager, and gracefully shutdown (after completing their current job) if the manager process disappears before them.

You can specify an alternate config file by setting the RESQUE_POOL_CONFIG or with the --config command line option.

See the examples directory for example chef cookbook and god config. In the chef cookbook, you can also find example init.d and muninrc templates (all very out of date, pull requests welcome).

TODO

See the TODO list at github issues.

Contributors

See list of contributors on github or in the changelog