We use Django behind gunicorn for most of the stack. Not that Django is slow (we still happily use it for 99% of our public facing system) -- it's just that OpenResty is so fast[1].

Besides the fact that OpenResty is built on C/Lua, Nginx's evented architecture makes it scale concurrently really well. We use Gunicorn in a multiprocessing configuration for the rest of the stack -- so each worker consumes much more RAM than the equivalent evented configuration.

We were toying with the idea of implementing the service with Tornado (also evented, but in Python) -- but OpenResty/Lapis has been so great for our use case that it's tough to justify the test at this point.

[1] http://www.techempower.com/benchmarks/