Puma Worker Killer

What

If you have a memory leak in your code, finding and plugging it can be a herculean effort. Instead what if you just killed your processes when they got to be too large? The Puma Worker Killer does just that. Similar to Unicorn Worker Killer but for the Puma web server.

Puma worker killer can only function if you have enabled cluster mode or hybrid mode (threads + worker cluster). If you are only using threads (and not workers) then puma worker killer cannot help keep your memory in control.

BTW restarting your processes to control memory is like putting a bandaid on a gunshot wound, try figuring out the reason you're seeing so much memory bloat derailed benchmarks can help.

Install

In your Gemfile add:

gem 'puma_worker_killer'

Then run $ bundle install

Now supports Heroku log-runtime-metrics

This gem can now get memory usage from your Heroku logs.

First, enable log-runtime-metrics on your heroku app:

$ heroku labs:enable log-runtime-metrics --app YOUR_APP_NAME

Next, get your app's Heroku auth token:

$ heroku auth:token # for a one-year token
# or...
$ heroku authorizations:create # for a long-lived token

Then, set these variables in your Heroku config.

$ heroku config:set API_TOKEN_HEROKU='the-token-you-just-got' APP_NAME_HEROKU='name-of-your-app'

If you set these variables in your development environment as well, you will see memory usage from the production app in your development logs.

Please note that occasionally the HTTP request for Heroku logs won't return lines with memory information. In these instances, you will see reaper status log entries like: PumaWorkerKiller: Consuming 0.0 mb with master and 3 workers. It is normal to see this occasionally. If every reaper status log entry looks like that, though, your app is not correctly receiving logs.

Finish with the configuration options below.

Turn on Rolling Restarts

A rolling restart will kill each of your workers on a rolling basis. You set the frequency which it conducts the restart. This is a simple way to keep memory down as Ruby web programs generally increase memory usage over time. If you're using Heroku it is difficult to measure RAM from inside of a container accurately, so if you don't use the feature above it is recommended to use this feature or use a log-drain-based dyno killer. You can enable rolling restarts by running:

# config/puma.rb

before_fork do
  require 'puma_worker_killer'

  PumaWorkerKiller.enable_rolling_restart # Default is every 6 hours
end

or you can pass in the restart frequency:

PumaWorkerKiller.enable_rolling_restart(12 * 3600) # 12 hours in seconds

Make sure if you do this to not accidentally call PumaWorkerKiller.start as well.

Enable Worker Killing

If you're not running on a containerized platform you can try to detect the amount of memory you're using and only kill Puma workers when you're over that limit. It may allow you to go for longer periods of time without killing a worker however it is more error prone than rolling restarts. To enable measurement based worker killing put this in your config/puma.rb:

# config/puma.rb

before_fork do
  require 'puma_worker_killer'

  PumaWorkerKiller.start
end

That's it. Now on a regular basis the size of all Puma and all of it's forked processes will be evaluated and if they're over the RAM threshold will be killed. Don't worry Puma will notice a process is missing and spawn a fresh copy with a much smaller RAM footprint ASAP.

Troubleshooting

When you boot your program locally you should see debug output:

[77773] Puma starting in cluster mode...
[77773] * Version 3.1.0 (ruby 2.3.1-p112), codename: El Niño Winter Wonderland
[77773] * Min threads: 0, max threads: 16
[77773] * Environment: development
[77773] * Process workers: 2
[77773] * Phased restart available
[77773] * Listening on tcp://0.0.0.0:9292
[77773] Use Ctrl-C to stop
[77773] PumaWorkerKiller: Consuming 54.34765625 mb with master and 2 workers.

If you don't see any PumaWorkerKiller output, make sure that you are running with multiple workers. PWK only functions if you have workers enabled, you should see something like this when Puma boots:

[77773] * Process workers: 2

If you've configured PWK's frequency try reducing it to a very low value

Configure

Before calling start you can configure PumaWorkerKiller. You can do so using a configure block or calling methods directly:

PumaWorkerKiller.config do |config|
  config.ram           = 1024 # mb
  config.frequency     = 5    # seconds
  config.percent_usage = 0.98
  config.rolling_restart_frequency = 12 * 3600 # 12 hours in seconds, or 12.hours if using Rails
  config.reaper_status_logs = true # setting this to false will not log lines like:
  # PumaWorkerKiller: Consuming 54.34765625 mb with master and 2 workers.

  config.pre_term = -> (worker) { puts "Worker #{worker.inspect} being killed" }

  # For Heroku log-runtime-metrics
  config.heroku_api_token = ENV['API_TOKEN_HEROKU'] # required
  config.heroku_app_name = ENV['APP_NAME_HEROKU'] # required
  config.reaper_status_logs = !Rails.env.production? # recommended to avoid redundancy with log-runtime-metrics memory data.
end
PumaWorkerKiller.start

pre_term

config.pre_term will be called just prior to worker termination with the worker that is about to be terminated. This may be useful to use in keeping track of metrics, time of day workers are restarted, etc.

By default Puma Worker Killer will emit a log when a worker is being killed

PumaWorkerKiller: Out of memory. 5 workers consuming total: 500 mb out of max: 450 mb. Sending TERM to pid 23 consuming 53 mb.

or

PumaWorkerKiller: Rolling Restart. 5 workers consuming total: 650mb mb. Sending TERM to pid 34.

However you may want to collect more data, such as sending an event to an error collection service like rollbar or airbrake. The pre_term lambda gets called before any worker is killed by PWK for any reason.

on_calculation

config.on_calculation will be called every time Puma Worker Killer calculates memory usage (config.frequency). This may be useful for monitoring your total puma application memory usage, which can be contrasted with other application monitoring solutions.

This callback lambda is given a single value for the amount of memory used.

Attention

If you start puma as a daemon, to add puma worker killer config into puma config file, rather than into initializers: Sample like this: (in config/puma.rb file):

before_fork do
  PumaWorkerKiller.config do |config|
    config.ram           = 1024 # mb
    config.frequency     = 5    # seconds
    config.percent_usage = 0.98
    config.rolling_restart_frequency = 12 * 3600 # 12 hours in seconds, or 12.hours if using Rails
  end
  PumaWorkerKiller.start
end

It is important that you tell your code how much RAM is available on your system. The default is 512 mb (the same size as a Heroku 1x dyno). You can change this value like this:

PumaWorkerKiller.ram = 1024 # mb

By default it is assumed that you do not want to hit 100% utilization, that is if your code is actually using 512 mb out of 512 mb it would be bad (this is dangerously close to swapping memory and slowing down your programs). So by default processes will be killed when they are at 99 % utilization of the value specified in PumaWorkerKiller.ram. You can change that value to 98 % like this:

PumaWorkerKiller.percent_usage = 0.98

You may want to tune the worker killer to run more or less often. You can adjust frequency:

PumaWorkerKiller.frequency = 20 # seconds

You may want to periodically restart all of your workers rather than simply killing your largest. To do that set:

PumaWorkerKiller.rolling_restart_frequency = 12 * 3600 # 12 hours in seconds, or 12.hours if using Rails

By default PumaWorkerKiller will perform a rolling restart of all your worker processes every 6 hours. To disable, set to false.

You may want to hide the following log lines: PumaWorkerKiller: Consuming 54.34765625 mb with master and 2 workers.. To do that set:

PumaWorkerKiller.reaper_status_logs = false

Note: It is true by default.

License

MIT

Feedback

Open up an issue or ping me on twitter @schneems.