Toggle navigation

Asynchronous and non-Blocking I/O

Real-time web features require a long-lived mostly-idle connection per user. In a traditional synchronous web server, this implies devoting one thread to each user, which can be very expensive.

To minimize the cost of concurrent connections, Tornado uses a single-threaded event loop. This means that all application code should aim to be asynchronous and non-blocking because only one operation can be active at a time.

The terms asynchronous and non-blocking are closely related and are often used interchangeably, but they are not quite the same thing.


A function blocks when it waits for something to happen before returning. A function may block for many reasons: network I/O, disk I/O, mutexes, etc. In fact, every function blocks, at least a little bit, while it is running and using the CPU (for an extreme example that demonstrates why CPU blocking must be taken as seriously as other kinds of blocking, consider password hashing functions like bcrypt, which by design use hundreds of milliseconds of CPU time, far more than a typical network or disk access).

A function can be blocking in some respects and non-blocking in others. In the context of Tornado we generally talk about blocking in the context of network I/O, although all kinds of blocking are to be minimized.


An asynchronous function returns before it is finished, and generally causes some work to happen in the background before triggering some future action in the application (as opposed to normal synchronous functions, which do everything they are going to do before returning). There are many styles of asynchronous interfaces:

  • Callback argument
  • Return a placeholder (Future, Promise, Deferred)
  • Deliver to a queue
  • Callback registry (e.g. POSIX signals)

Regardless of which type of interface is used, asynchronous functions by definition interact differently with their callers; there is no free way to make a synchronous function asynchronous in a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous).


Here is a sample synchronous function:

from tornado.httpclient import HTTPClient

def synchronous_fetch(url):
    http_client = HTTPClient()
    response = http_client.fetch(url)
    return response.body

And here is the same function rewritten to be asynchronous with a callback argument:

from tornado.httpclient import AsyncHTTPClient

def asynchronous_fetch(url, callback):
    http_client = AsyncHTTPClient()
    def handle_response(response):
    http_client.fetch(url, callback=handle_response)

And again with a Future instead of a callback:

from tornado.concurrent import Future

def async_fetch_future(url):
    http_client = AsyncHTTPClient()
    my_future = Future()
    fetch_future = http_client.fetch(url)
        lambda f: my_future.set_result(f.result()))
    return my_future

The raw Future version is more complex, but Futures are nonetheless recommended practice in Tornado because they have two major advantages. Error handling is more consistent since the Future.result method can simply raise an exception (as opposed to the ad-hoc error handling common in callback-oriented interfaces), and Futures lend themselves well to use with coroutines. Coroutines will be discussed in depth in the next section of this guide. Here is the coroutine version of our sample function, which is very similar to the original synchronous version:

from tornado import gen

def fetch_coroutine(url):
    http_client = AsyncHTTPClient()
    response = yield http_client.fetch(url)
    raise gen.Return(response.body)

The statement raise gen.Return(response.body) is an artifact of Python 2, in which generators aren’t allowed to return values. To overcome this, Tornado coroutines raise a special kind of exception called a Return. The coroutine catches this exception and treats it like a returned value. In Python 3.3 and later, a return response.body achieves the same result.