Using Puma on Heroku instead of Unicorn

A few years back I wrote about using Unicorn on Heroku along with Unicorn Worker Killer. Oyr need was to increase the number of web requests an app could handle on Heroku. Puma tested faster at the time but did not offer worker processes. You have to use threads and for an app that was never intended to be threaded running on Ruby MRI it was too much work to adjust. Puma has since added the concept of Worker Processes and, given the performance advantages found with Puma it made sense to install it. Still, I was used to some variability with Unicorn and wanted to replicate some of the same things. This article from Heroku offered most of the help I needed though I did reduce the default thread count from 5 to 1. Additionally the section on the use of the Rack Timeout gem was helpful though I wound up doing this so that I could adjust the timeout via environment variable:

 Rack::Timeout.timeout = 20 # seconds  

What I did not find was helpful instructions on queue length. Puma defaults to 1024 and the previously mentioned article from Heroku cautions strongly against altering that queue length. Still, from previous performance testing on Heroku sometimes a shorter queue length is warranted (DO YOUR OWN TESTING!). This blog post originally suggested altering the queue length but again he cautions against it on Heroku. If you find that a shorter queue length helps then it is a setting in the config/puma.rb file:

 backlog = Integer(ENV['PUMA_BACKLOG'] || 20)  
 bind "tcp://0.0.0.0:#{port}?backlog=#{backlog}"   

Good luck and Heroku on

Using Unicorn Worker Killer to help reduce queue backlog on Heroku with Unicorn

In light of the update from Heroku CTO Adam Wiggins and having been part of the efforts to test the larger Dynos with Unicorn on Heroku I felt it necessary to share Unicorn Worker Killer; a tool we have found indispensable.

Based on configurable thresholds for memory and number of requests received it will kill off a Unicorn Worker. For those of you on Heroku this is valuable in that it then returns all the requests back to the Heroku random routing. While not an exact correlation between too many requests and memory size it does help keep individual workers from becoming too overloaded.

Add this to your Gemfile:

gem 'unicorn-worker-killer'

The gem suggests that you configure your `config.ru` file for the thresholds. This can be cumbersome on Heroku if you need to test out different settings.

Thankfully you can also control the thresholds via environment variables. In config.ru do:

max_request_min =  ENV['MAX_REQUEST_MIN'].to_i || 3072
max_request_max =  ENV['MAX_REQUEST_MAX'].to_i || 4096

# Max requests per worker
use Unicorn::WorkerKiller::MaxRequests, max_request_min, max_request_max

oom_min = ((ENV['OOM_MIN'].to_i || 192) * (1024**2))
oom_max = ((ENV['OOM_MAX'].to_i || 256) * (1024**2))

# Max memory size (RSS) per worker
use Unicorn::WorkerKiller::Oom, oom_min, oom_max

The run this on the heroku command line:

heroku config:add OOM_MAX=256 memory_limit_min =192 MAX_REQUEST_MIN=3072 MAX_REQUEST_MAX=4096 -a unicon-ttm-sandbox

That will add in the variables needed to control the thresholds. The example shows the defaults though we found dropping the OOM_MAX to 216 worked best

Heroku with Unicorn backlog settings and performance

In light of this post from Rap Genius and subsequent blow up on Hacker News I decided to share what we did to play with the request queue on Unicorn & Heroku.

In the app/config/unicorn.rb we changed the backlog line to this:

:backlog => Integer(ENV['UNICORN_BACKLOG'] || 200)

And then we can alter the backlog as necessary via:

heroku config:set UNICORN_BACKLOG=25 -a <app_name>

We found 25 to be a sweet spot for two Unicorn Workers but it may be different with four (which we are experimenting with now). Again this is also relative to the app so your results may (and probably will) be different.

Update:

I neglected to mention that you have to move the port declaration from config.ru and put it in the unicorn.rb too. The full line should be something like this:

listen ENV['PORT'], :backlog => Integer(ENV['UNICORN_BACKLOG'] || 200)