Magical Unicorn Configuration for Heroku

config.ru

# add something like this to config.ru
# see https://github.com/kzk/unicorn-worker-killer

if Integer(ENV['UNICORN_KILLER'] || 0) != 0
  require 'unicorn/worker_killer'
  # Max memory size (RSS) per worker
  use Unicorn::WorkerKiller::Oom, (350*(1024**2)), (400*(1024**2)), 30, true
end

unicorn.rb

# sample config:
# http://unicorn.bogomips.org/examples/unicorn.conf.rb

# may want Out-of-band GC
# http://unicorn.bogomips.org/Unicorn/OobGC.html

# and unicorn killer to avoid swapping
# https://github.com/kzk/unicorn-worker-killer

# lots of good comments here:
# https://blog.heroku.com/archives/2013/2/27/unicorn_rails

worker_processes Integer(ENV['UNICORN_WORKERS'] || 2)

# Load your app into the master before forking
# workers for super-fast worker spawn times
preload_app true

# Immediately restart any workers that haven't responded
# Heroku will hangup after 30 seconds
# https://devcenter.heroku.com/articles/rails-unicorn#timeouts
timeout Integer(ENV['UNICORN_TIMEOUT'] || 15)

# https://devcenter.heroku.com/articles/forked-pg-connections
before_fork do |server, worker|
  Signal.trap 'TERM' do
    puts 'Unicorn master intercepting TERM and sending myself QUIT instead'
    Process.kill 'QUIT', Process.pid
  end

  defined?(ActiveRecord::Base) and
    ActiveRecord::Base.connection.disconnect!
end

after_fork do |server, worker|
  Signal.trap 'TERM' do
    puts 'Unicorn worker intercepting TERM and doing nothing. Wait for master to sent QUIT'
  end

  defined?(ActiveRecord::Base) and
    ActiveRecord::Base.establish_connection

  # Dalli does not need anything special in Unicorn/Passenger since 2.0.4. It will
  # detect sockets shared with child processes and gracefully reopen the socket.
  # https://github.com/mperham/dalli/issues/208

  # Redis/Resque and other external services may need special handling:
  # https://devcenter.heroku.com/articles/rails-unicorn
end

# If you are running unicorn on multiple machines, lowering this number can help
# your load balancer detect when a machine is overloaded and give requests to a
# different machine. (default 1024)
# http://unicorn.bogomips.org/Unicorn/Configurator.html#method-i-listen
# http://manpages.ubuntu.com/manpages/gutsy/man2/listen.2.html (max 128)
listen ENV['PORT'], :backlog => Integer(ENV['UNICORN_BACKLOG'] || 15)

Why do you use 350 to 400 limit on Heroku? On Heroku there is a 512Mb limit per dyno. How many unicorn workers do you have?

Awesome gist!

@e3matheus, That is the memory limit of a single unicorn worker. You typically have multiple workers per Dyno. So if you have 2 workers running at 250MB and and 380MB, you would want to kill the 380MB one only.