load balancing socket vs ip:port about gunicorn

[jbergstroem]

Not really a issue - but to answer your question, gunicorn will balance incoming requests for you. Gunicorn is similar to Unicorn (unicorn.bogomips.org) in this regard - so feel free to read there for more info. Perhaps some architecture notes should be added to docs?

[benoitc]

Load balancing is done by the OS. You can still use NGINX to load balance between different machine. Or even if you need by launching 2 in instances of gunicorn on same app.

[davisp]

There's no difference between Unix and TCP sockets here. Each listen() socket gets a connection queue, and accept() just pops a connection from that queue. In gunicorn each worker just pops a new connection from that queue as its able. Its similar to the classic producer/consumer model in that the producer just happens to be managed by the kernel.

The load balancing in nginx is orthogonal to a certain degree. If you're used to doing something like using nginx to proxy to 4-12 mongrels on the same machine, you probably just need nginx to proxy to a single gunicorn instance with 4-12 workers.

[benoitc]

think answer have been done, feel free to reopen it if needed.

[radzhome]

The only difference I can note between the two is the overhead created by TCP/IP when using ip:port. Ip:port is a bit more flexible as it works on any OS. Unix sockets less portable but faster, everything else is the same. Maybe something can be added to the docs to note the differences.