Tornado '@run_on_executor' is blocking

I would like to ask how tornado.concurrent.run_on_executor (later just run_on_executor) works, because I probably do not understand how to run synchronous task to not block the main IOLoop.

All the examples using run_on_executor, which I found, are using just time to block the loop. With time module it works fine, but when I try some time intesive calculations, using run_on_executor, the task blocks the IOLoop. I am able to see that the app uses multiple threads, but it is still blocking.

I want to use run_on_executor for hashing passwords using bcrypt, but replaced it with this calculation to gain some extra time for testing.

Here I have small app, to demonstrate my confusion.

from tornado.options import define, options
import tornado.web
import tornado.httpserver
from tornado import gen
from tornado.concurrent import run_on_executor
import tornado.httpclient
import tornado.escape
import time
import concurrent.futures
import urllib


executor = concurrent.futures.ThreadPoolExecutor(20)
define("port", default=8888, help="run on the given port", type=int)


# Should not be blocking ?
class ExpHandler(tornado.web.RequestHandler):
    _thread_pool = executor

    @gen.coroutine
    def get(self, num):
        i = int(num)
        result = yield self.exp(2, i)
        self.write(str(result))
        self.finish()

    @run_on_executor(executor="_thread_pool")
    def exp(self, x, y):
        result = x ** y
        return(result)


class NonblockingHandler(tornado.web.RequestHandler):
    @gen.coroutine
    def get(self):
        http_client = tornado.httpclient.AsyncHTTPClient()
        try:
            response = yield http_client.fetch("http://www.google.com/")
            self.write(response.body)
        except tornado.httpclient.HTTPError as e:
            self.write(("Error: " + str(e)))
        finally:
            http_client.close()
        self.finish()


class SleepHandler(tornado.web.RequestHandler):
    _thread_pool = executor

    @gen.coroutine
    def get(self, sec):
        sec = float(sec)
        start = time.time()
        res = yield self.sleep(sec)
        self.write("Sleeped for {} s".format((time.time() - start)))
        self.finish()

    @run_on_executor(executor="_thread_pool")
    def sleep(self, sec):
        time.sleep(sec)
        return(sec)


class Application(tornado.web.Application):
    def __init__(self):
        handlers = [
            (r'/exp/(?P<num>[^\/]+)?', ExpHandler),
            (r'/nonblocking/?', NonblockingHandler),
            (r'/sleep/(?P<sec>[^\/]+)?',SleepHandler)
        ]
        settings = dict(
            debug=True,
            logging="debug"
        )
        tornado.web.Application.__init__(self, handlers, **settings)


def  main():
    tornado.options.parse_command_line()
    http_server = tornado.httpserver.HTTPServer(Application())
    http_server.listen(options.port)
    io_loop = tornado.ioloop.IOLoop.instance()
    io_loop.start()


if __name__ == "__main__":
    main()

I would be very grateful for any explanation why ExpHandler, running in executor is blocking the loop.

Python (at least in the CPython implementation) has a Global Interpreter Lock which prevents multiple threads from executing Python code at the same time. In particular, anything which runs in a single Python opcode is uninterruptible unless it calls a C function which explicitly releases the GIL. A large exponentation with ** holds the GIL the whole time and thus blocks all other python threads, while a call to bcrypt() will release the GIL so other threads can continue to work.