Skip to content

Tornado Web Server

Tornado Web Server 4.5.3#

This is an old version of tornado

  • Python web framework and asynchronous networking library
  • Non-blocking network io, tornado can scale to 10s of thousands of open connections
  • Ideal for long polling and web sockets
  • Is not based on WSGI (Web Server Gateway Interface) spec defined in PEP3333
  • Is run one thread per process
  • Can use tornado.wsgi but better to use tornado.web
  • Integration with asyncio only happened in tornado>=6


pip install tornado

Hello World Example#

import asyncio

import tornado.web

class MainHandler(tornado.web.RequestHandler):
    def get(self):
        self.write("Hello, world")

def make_app():
    return tornado.web.Application([
        (r"/", MainHandler),

async def main():
    app = make_app()
    await asyncio.Event().wait()

if __name__ == "__main__":


$ http localhost:8888      
HTTP/1.1 200 OK
Content-Length: 12
Content-Type: text/html; charset=UTF-8
Date: Mon, 11 Jul 2022 06:43:32 GMT
Etag: "e02aa1b106d5c7c6a98def2b13005d5b84fd8dc8"
Server: TornadoServer/6.2

Hello, world

This example does not use tornado’s async features…

Load Test Hello World#
$ ab -c 100 -n 1000 http://localhost:8888/
Requests per second:    5915.93 [#/sec] (mean)
Time per request:       16.903 [ms] (mean)
Time per request:       0.169 [ms] (mean, across all concurrent requests)
Transfer rate:          1184.34 [Kbytes/sec] received

Connection Times (ms)
            min  mean[+/-sd] median   max
Connect:        0    2   1.5      2       7
Processing:     5   14   3.5     13      23
Waiting:        2   14   3.5     13      23
Total:          6   16   3.6     16      24

Percentage of the requests served within a certain time (ms)
50%     16
66%     17
75%     18
80%     19
90%     21
95%     22
98%     24
99%     24
100%     24 (longest request)

ab or tornado seems to break (on mac) when there are more than 200 concurrect requests:

apr_socket_recv: Connection reset by peer (54)

Structure of a Tornado web application#

  • One or more RequestHandler subclasses
  • An Application object - that routes incoming requests to handlers
  • a main() function to start the server

The Application object - global config and routing requests to handlers. A list of URLSpec objects - mapping a regular expression to a handler, initialization arguments it sends to RequestHandler.initialize and it also has a name so it can be reversed with RequestHandler.reverse_url. Regular expression variables are sent as a string to the get function.

Subclassing RequestHandler - the main entry point is the HTTP method being handled get(), post(). To return a response RequestHandler.render or RequestHandler.write are called. Render needs a template. Write accepts dicts that will return as json.


It is common to define a BaseHandler class that overrides methods such as write_error and get_current_user and then subclass your own BaseHandler instead of RequestHandler for all your specific handlers

Handling request input#

You can access the HttpServerRequest with self.request.

Tornado does not parse json request bodies - if you want to do that you can override prepare:

def prepare(self):
    if self.request.headers["Content-Type"].startswith("application/json"):
        self.json_args = json.loads(self.request.body)
        self.json_args = None
Sequence of Calls#
  1. A new RequestHandler object is created
  2. initialize() is called with the argument defined in main
  3. prepare() is called - usually in the base class - may produce output; if it calls finish (or redirect, etc), processing stops here
  4. One of the http methods is called get(), post(), put()
  5. When finished on_finish() is called. Synchronously it runs after get(), asynchronously it runs after finish()

Commonly overidden things of RequestHandler:

  • write_error - outputs HTML for use on error pages.
  • on_connection_close - called when the client disconnects; applications may choose to detect this case and halt further processing. Note that there is no guarantee that a closed connection can be detected promptly.
  • get_current_user - see User authentication
  • get_user_locale - returns Locale object to use for the current user
  • set_default_headers - may be used to set additional headers on the response (such as a custom Server header)
Error Handling#
  • RequestHandler.write_error generates an error page. Override this for a custom error messag (in the base handler).
  • RequestHandler.redirect - use redirect()
  • point to a RedirectHandler instead of a RequestHandler
Asynchronous handlers#

Tornado handlers are synchronous by default: when the get()/post() method returns, the request is considered finished and the response is sent. Since all other requests are blocked while one handler is running, any long-running handler should be made asynchronous so it can call its slow operations in a non-blocking way.

The simplest way to make asynchornous server handler according to the docs is using coroutines but there is no example there.

The example they use:

class MainHandler(tornado.web.RequestHandler):
    def get(self):
        http = tornado.httpclient.AsyncHTTPClient()
        response = yield http.fetch("")
        json = tornado.escape.json_decode(response.body)
        self.write("Fetched " + str(len(json["entries"])) + " entries "
                "from the FriendFeed API")

You can also use the asynchornous decorator with a async client with a callback

class MainHandler(tornado.web.RequestHandler):
    def get(self):
        http = tornado.httpclient.AsyncHTTPClient()

    def on_response(self, response):
        if response.error: raise tornado.web.HTTPError(500)
        json = tornado.escape.json_decode(response.body)
        self.write("Fetched " + str(len(json["entries"])) + " entries "
                "from the FriendFeed API")

When get() returns, the request has not finished. When the HTTP client eventually calls on_response(), the request is still open, and the response is finally flushed to the client with the call to self.finish().

In the above examples when the request is seen - it immediately puts the work to a thread and deals with the next incoming requests.

You can also inspect the results in the callback or after the yield keyword with import ipdb; ipdb.set_trace()

Instead yield and the gen.coroutine decorator you can use async and await if you are on python3.5 and later:

class MainHandler(tornado.web.RequestHandler):
    async def get(self):
        http = tornado.httpclient.AsyncHTTPClient()
        response = await http.fetch("")
        json = tornado.escape.json_decode(response.body)
        self.write("Fetched " + str(len(json["entries"])) + " entries "
                "from the FriendFeed API")

Only tasks that will take a long time (usually IO - network and file based operations) need to be defferred. That means more requests can be deal with quickly and none will be blocked.

Running and deploying#

Tornado supplies its own HTTPServer - so it is different from other web frameworks (fast api, django).

Tornado is normally intended to be run on its own, without a WSGI container. In some environments only wsgi is allowed.

The features that are not allowed in WSGI mode include:

  • coroutines
  • the @asynchronous decorator
  • AsyncHTTPClient
  • the auth module
  • WebSockets

Instead of configuring a WSGI container to find your applciation, you write a main() function:

def main():
    app = make_app()

if __name__ == '__main__':

Configure your OS or process manager to run this program

Remember to increase the number of open files per process (to avoid “Too many open files”-Error)

Processes and ports#

Because of the Python GIL (Global Interpreter Lock) you need to run multiple processes to make full use of the number of cores on your computer or server.

Typically it is best to run one process per CPU

You can enable multiprocess mode with:

def main():
    app = make_app()
    server = tornado.httpserver.HTTPServer(app)
    server.start(0)  # forks one process per cpu

Note you must use server.bind() and not server.listen()


  • each child process will have its own IOLoop - so do not modify before forking
  • all processes share the same port - it is hard to monitor individually
  • difficult to do zero downtime updates

For more sophisticated deployments, it is recommended to start the processes independently, and have each one listen on a different port. The process groups feature of supervisord is one good way to arrange this

Debug mode#

Passing debug=True to the app constructor enables:

  • autoreload=True - automatic reload (only works in single process mode)
  • compiled_template_cache=False - templates will not cache
  • static_hash_cache=False - static files will not be cached
  • serve_traceback=True - an error page and stack trace will be generated

Automatic reload standalone: python -m tornado.autoreload

Templates and UI#

Info on templates and ui


Info on authentication

Major components of Tornado#

  • A web framework RequestHandler sub-classed / inherited from to create web applications
  • Client- and server-side implementions of HTTP (HTTPServer andAsyncHTTPClient).
  • An asynchronous networking library including the classes IOLoop and IOStream, which serve as the building blocks for the HTTP components and can also be used to implement other protocols.


  • Tornado is not thread-safe - manipulation of shared data structures between threads can have unintended consequences - like race conditions.
  • IOLoop.add_callback is the only method safe to call from other threads
  • IOLoop.run_in_executor is the recommended way to run synchronously blocking code in another thread
  • Integrated with standard lib ayncio (only in versions 6 and greater)

Asynchronous and non-blocking IO#

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.

  • An asynchronous function returns before it is finished
  • They do work in the background before triggering a future action
  • Synchronous functions do everything they need to do before returning
  • asynchronous functions by definition interact differently with their callers - there is no free way to make a synchronous function asynchronous
  • systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous

Styles of asynchronous interfaces:

  • Callback argument
  • Return a placeholder (Future, Promise, Deferred)
  • Deliver to a queue
  • Callback registry (e.g. POSIX signals)
  • A function blocks when it waits for something to happen before returning.
  • A function blocks for many reasons: network I/O, disk I/O, mutexes
  • In fact, every function blocks, at least a little bit, while it is running and using the CPU (be careful with password hashing like bcrypt)

More Info#

More info in the docs:

Sometimes it is hard to even get a time.sleep() working asynchronously as per the docs. The recommendation is to find a async library for your task made for tornado