Serialized handling of requests

[liaden]

Currently running into a deadlock problem where micro service A is hitting micro service B which is hitting micro service A again. The deadlock causes a timeout.

With a controller like

class StatusController < ApplicationController
  def index
    sleep 25
  end
end

and the following:

time curl machine.dev &
time curl machine.dev &

we end up seeing request 1 takes 25.x seconds and request 2 takes 52 seconds to complete.

We have tried the following .praxconfig:

export PRAX_APP_THREADS=16
export PRAX_THREADS=32
source /usr/local/share/chruby/chruby.sh
source /usr/local/share/chruby/auto.sh

[ysbaddaden]

Prax isn't battle tested for this use case: many services calling each others. You'll eventually run out of workers, and deadlock.

Yet, is your sleep example called many times through the services, or just twice by curl?

[liaden]

In my example, sleep is called just twice by curl.

It also seems like there is a difference in the behavior of how prax runs my different rails apps if I have the application launched with rubinius instead of MRI. I am not entirely sure that changing PRAX_APP_THREADS or PRAX_THREADS is changing anything in how the system behave.

[ysbaddaden]

I never checked against Rails apps, but the same test with a raw Rack application didn't wait (as far as I remember).

I recall Rails before 4.0 (?) added a mutex around requests, to force requests to not run in parallel.

[liaden]

I am unsure of what Rails and Rack is doing, but when running on Macs with powify, the problem is not seen.