Gunicorn frustrations

I like Gunicorn but as Circus grows I am getting frustrated of the lack of control I have over the web worker it manages.

Basically, Gunicorn does the same work than Circus on the processes it manages: it spawns them and manage their lives.

But Circus gives us much more control over the processes it runs. We can:

• get a continuous feedback on the CPU / Memory usage, per process
• add or remove more processes
• basically do any maintenance operation on a live stack

Also, features like the flapping control, or the Web Console coming in the 0.4 version, makes Circus much more appealing to use.

What I want to have is a simple Python process that gets HTTP requests and send back responses. From there I can drive my little workers in Circus as I would do for any other processes.

I can't do this with Gunicorn because it uses a prefork model and deal itself with its processes.

If I run a Gunicorn server with a single worker for instance, I can't add more workers once it's live.

The socket Gunicorn binds belong to its main process and you can't run another process on it.

edited - as Philip said in the comments, Gunicorn let you add/remove workers with the TTIN and TTOU signals, but Circus have much more control, see : http://circus.readthedocs.org/en/latest/commands/#circus-ctl-commands

ZeroMQ sockets are a bit different here, as you can connect as many processes as you want on a single socket, and have load-balancing for free.

I guess one solution would be to rewrite a WSGI server that interprets the HTTP request then pushes it to a ZMQ socket for a worker to pick it up.

But wait, it exists..

Mongrel2

Mongrel2 is exactly what I am looking for, because it basically forwards HTTP requests into a ZMQ socket, where I can connect as many workers as I want.

m2wsgi is a WSGI handler that can be used to get the ZMQ requests from Mongrel2 and send back the response.

It's simple to create a script that runs your Python WSGI app using m2wsgi:

from m2wsgi.io.standard import WSGIHandler
from mysuperapp import wsgiapp

zmq_port =  "tcp://127.0.0.1:9999"
handler = WSGIHandler(wsgiapp, zmq_port)
handler.serve()

Then to have Mongrel2 send stuff on that port. Something like:

wsgi_handler = Handler(send_spec='tcp://127.0.0.1:9999',
                    send_ident='be4ee7d-6a47-42dd-9acd-1707add81835',
                    recv_spec='tcp://127.0.0.1:9998',.
                    recv_ident='')

routes={ '/': wsgi_handler }

localhost = Host(name='localhost', routes=routes)
localip = Host(name='127.0.0.1', routes=routes)

main = Server(
    uuid='31bf6b07-a147-466c-87b5-961481b99201',
    access_log='/logs/access.log',
    error_log='/logs/error.log',
    chroot='/var/www/mongrel2/',
    pid_file='/run/mongrel2.pid',
    default_host='localhost',
    name='main',
    port=6767,
    hosts=[localhost, localip]
)

settings = {'zeromq.threads': 1}
servers = [main]

(Inspired from http://server.dzone.com/articles/deploying-graphite-mongrel2)

And finally, have Mongrel and m2wsgi processes managed by Circus:

[circus]
check_delay = 5
endpoint = tcp://127.0.0.1:5555
pubsub_endpoint = tcp://127.0.0.1:5556
stats_endpoint = tcp://127.0.0.1:5557

[watcher:mongrel2]
cmd = mongrel2 tests/config.sqlite 31bf6b07-a147-466c-87b5-961481b99201
warmup_delay = 0
numprocesses = 1
working_dir = /Users/tarek/Dev/github.com/circus-wsgi
stdout_stream.class = StdoutStream
stderr_stream.class = StdoutStream

[watcher:m2wsgi]
cmd = bin/python server.py
numprocesses = 2

Then. circusctl, circus-top and circushttpd give me a full control on my stack, live:

$ circusctl list
circusd-stats,m2wsgi,mongrel2

$ circusctl list m2wsgi
1

$ circusctl incr m2wsgi
2

$ circusctl stats m2wsgi
m2wsgi:
1: 10936  python tarek 0 N/A N/A N/A N/A N/A
2: 10946  python tarek 0 N/A N/A N/A N/A N/A

What's next

I'll bench a Mongrel2-Circus stack to see how it performs compared to our current Gunicorn stack.

If the results are good, I might try to write a small circus-wsgi integration package to make it easier to setup and configure everything together.