Toggle navigation

What’s new in Tornado 4.0

July 15, 2014


  • The tornado.web.stream_request_body decorator allows large files to be uploaded with limited memory usage.
  • Coroutines are now faster and are used extensively throughout Tornado itself. More methods now return Futures, including most IOStream methods and RequestHandler.flush.
  • Many user-overridden methods are now allowed to return a Future for flow control.
  • HTTP-related code is now shared between the tornado.httpserver, tornado.simple_httpclient and tornado.wsgi modules, making support for features such as chunked and gzip encoding more consistent. HTTPServer now uses new delegate interfaces defined in tornado.httputil in addition to its old single-callback interface.
  • New module tornado.tcpclient creates TCP connections with non-blocking DNS, SSL handshaking, and support for IPv6.

Backwards-compatibility notes

  • tornado.concurrent.Future is no longer thread-safe; use concurrent.futures.Future when thread-safety is needed.
  • Tornado now depends on the certifi package instead of bundling its own copy of the Mozilla CA list. This will be installed automatically when using pip or easy_install.
  • This version includes the changes to the secure cookie format first introduced in version 3.2.1, and the xsrf token change in version 3.2.2. If you are upgrading from an earlier version, see those versions’ release notes.
  • WebSocket connections from other origin sites are now rejected by default. To accept cross-origin websocket connections, override the new method WebSocketHandler.check_origin.
  • WebSocketHandler no longer supports the old draft 76 protocol (this mainly affects Safari 5.x browsers). Applications should use non-websocket workarounds for these browsers.
  • Authors of alternative IOLoop implementations should see the changes to IOLoop.add_handler in this release.
  • The RequestHandler.async_callback and WebSocketHandler.async_callback wrapper functions have been removed; they have been obsolete for a long time due to stack contexts (and more recently coroutines).
  • curl_httpclient now requires a minimum of libcurl version 7.21.1 and pycurl 7.18.2.
  • Support for RequestHandler.get_error_html has been removed; override RequestHandler.write_error instead.

Other notes

  • The git repository has moved to All old links should be redirected to the new location.
  • An announcement mailing list is now available.
  • All Tornado modules are now importable on Google App Engine (although the App Engine environment does not allow the system calls used by IOLoop so many modules are still unusable).


  • Fixed a bug in .FacebookMixin on Python 3.
  • When using the Future interface, exceptions are more reliably delivered to the caller.


  • tornado.concurrent.Future is now always thread-unsafe (previously it would be thread-safe if the concurrent.futures package was available). This improves performance and provides more consistent semantics. The parts of Tornado that accept Futures will accept both Tornado’s thread-unsafe Futures and the thread-safe concurrent.futures.Future.
  • tornado.concurrent.Future now includes all the functionality of the old TracebackFuture class. TracebackFuture is now simply an alias for Future.


  • curl_httpclient now passes along the HTTP “reason” string in response.reason.


  • Performance of coroutines has been improved.
  • Coroutines no longer generate StackContexts by default, but they will be created on demand when needed.
  • The internals of the tornado.gen module have been rewritten to improve performance when using Futures, at the expense of some performance degradation for the older YieldPoint interfaces.
  • New function with_timeout wraps a Future and raises an exception if it doesn’t complete in a given amount of time.
  • New object moment can be yielded to allow the IOLoop to run for one iteration before resuming.
  • Task is now a function returning a Future instead of a YieldPoint subclass. This change should be transparent to application code, but allows Task to take advantage of the newly-optimized Future handling.


  • New module contains the HTTP implementation shared by tornado.httpserver and tornado.simple_httpclient.


  • The command-line HTTP client (python -m tornado.httpclient $URL) now works on Python 3.
  • Fixed a memory leak in AsyncHTTPClient shutdown that affected applications that created many HTTP clients and IOLoops.
  • New client request parameter decompress_response replaces the existing use_gzip parameter; both names are accepted.


  • tornado.httpserver.HTTPRequest has moved to tornado.httputil.HTTPServerRequest.
  • HTTP implementation has been unified with tornado.simple_httpclient in tornado.http1connection.
  • Now supports Transfer-Encoding: chunked for request bodies.
  • Now supports Content-Encoding: gzip for request bodies if decompress_request=True is passed to the HTTPServer constructor.
  • The connection attribute of HTTPServerRequest is now documented for public use; applications are expected to write their responses via the HTTPConnection interface.
  • The HTTPServerRequest.write and HTTPServerRequest.finish methods are now deprecated. (RequestHandler.write and RequestHandler.finish are not deprecated; this only applies to the methods on HTTPServerRequest)
  • HTTPServer now supports HTTPServerConnectionDelegate in addition to the old request_callback interface. The delegate interface supports streaming of request bodies.
  • HTTPServer now detects the error of an application sending a Content-Length error that is inconsistent with the actual content.
  • New constructor arguments max_header_size and max_body_size allow separate limits to be set for different parts of the request. max_body_size is applied even in streaming mode.
  • New constructor argument chunk_size can be used to limit the amount of data read into memory at one time per request.
  • New constructor arguments idle_connection_timeout and body_timeout allow time limits to be placed on the reading of requests.
  • Form-encoded message bodies are now parsed for all HTTP methods, not just POST, PUT, and PATCH.



  • IOLoop.add_handler and related methods now accept file-like objects in addition to raw file descriptors. Passing the objects is recommended (when possible) to avoid a garbage-collection-related problem in unit tests.
  • New method IOLoop.clear_instance makes it possible to uninstall the singleton instance.
  • Timeout scheduling is now more robust against slow callbacks.
  • IOLoop.add_timeout is now a bit more efficient.
  • When a function run by the IOLoop returns a Future and that Future has an exception, the IOLoop will log the exception.
  • New method IOLoop.spawn_callback simplifies the process of launching a fire-and-forget callback that is separated from the caller’s stack context.
  • New methods IOLoop.call_later and IOLoop.call_at simplify the specification of relative or absolute timeouts (as opposed to add_timeout, which used the type of its argument).


  • The callback argument to most IOStream methods is now optional. When called without a callback the method will return a Future for use with coroutines.
  • New method IOStream.start_tls converts an IOStream to an SSLIOStream.
  • No longer gets confused when an IOError or OSError without an errno attribute is raised.
  • BaseIOStream.read_bytes now accepts a partial keyword argument, which can be used to return before the full amount has been read. This is a more coroutine-friendly alternative to streaming_callback.
  • BaseIOStream.read_until and read_until_regex now acept a max_bytes keyword argument which will cause the request to fail if it cannot be satisfied from the given number of bytes.
  • IOStream no longer reads from the socket into memory if it does not need data to satisfy a pending read. As a side effect, the close callback will not be run immediately if the other side closes the connection while there is unconsumed data in the buffer.
  • The default chunk_size has been increased to 64KB (from 4KB)
  • The IOStream constructor takes a new keyword argument max_write_buffer_size (defaults to unlimited). Calls to BaseIOStream.write will raise StreamBufferFullError if the amount of unsent buffered data exceeds this limit.
  • ETIMEDOUT errors are no longer logged. If you need to distinguish timeouts from other forms of closed connections, examine stream.error from a close callback.


  • When bind_sockets chooses a port automatically, it will now use the same port for IPv4 and IPv6.
  • TLS compression is now disabled by default on Python 3.3 and higher (it is not possible to change this option in older versions).


  • It is now possible to disable the default logging configuration by setting options.logging to None instead of the string "none".


  • Now works on Python 2.6.
  • Now works with Trollius version 0.3.



  • simple_httpclient has better support for IPv6, which is now enabled by default.
  • Improved default cipher suite selection (Python 2.7+).
  • HTTP implementation has been unified with tornado.httpserver in tornado.http1connection
  • Streaming request bodies are now supported via the body_producer keyword argument to tornado.httpclient.HTTPRequest.
  • The expect_100_continue keyword argument to tornado.httpclient.HTTPRequest allows the use of the HTTP Expect: 100-continue feature.
  • simple_httpclient now raises the original exception (e.g. an IOError) in more cases, instead of converting everything to HTTPError.


  • The stack context system now has less performance overhead when no stack contexts are active.


  • New module which creates TCP connections and IOStreams, including name resolution, connecting, and SSL handshakes.


  • AsyncTestCase now attempts to detect test methods that are generators but were not run with @gen_test or any similar decorator (this would previously result in the test silently being skipped).
  • Better stack traces are now displayed when a test times out.
  • The @gen_test decorator now passes along *args, **kwargs so it can be used on functions with arguments.
  • Fixed the test suite when unittest2 is installed on Python 3.


  • It is now possible to support streaming request bodies with the stream_request_body decorator and the new RequestHandler.data_received method.
  • RequestHandler.flush now returns a Future if no callback is given.
  • New exception Finish may be raised to finish a request without triggering error handling.
  • When gzip support is enabled, all text/* mime types will be compressed, not just those on a whitelist.
  • Application now implements the HTTPMessageDelegate interface.
  • HEAD requests in StaticFileHandler no longer read the entire file.
  • StaticFileHandler now streams response bodies to the client.
  • New setting compress_response replaces the existing gzip setting; both names are accepted.
  • XSRF cookies that were not generated by this module (i.e. strings without any particular formatting) are once again accepted (as long as the cookie and body/header match). This pattern was common for testing and non-browser clients but was broken by the changes in Tornado 3.2.2.


  • WebSocket connections from other origin sites are now rejected by default. Browsers do not use the same-origin policy for WebSocket connections as they do for most other browser-initiated communications. This can be surprising and a security risk, so we disallow these connections on the server side by default. To accept cross-origin websocket connections, override the new method WebSocketHandler.check_origin.
  • WebSocketHandler.close and WebSocketClientConnection.close now support code and reason arguments to send a status code and message to the other side of the connection when closing. Both classes also have close_code and close_reason attributes to receive these values when the other side closes.
  • The C speedup module now builds correctly with MSVC, and can support messages larger than 2GB on 64-bit systems.
  • The fallback mechanism for detecting a missing C compiler now works correctly on Mac OS X.
  • Arguments to are now decoded in the same way as arguments to RequestHandler.get and similar methods.
  • It is now allowed to override prepare in a WebSocketHandler, and this method may generate HTTP responses (error pages) in the usual way. The HTTP response methods are still not allowed once the WebSocket handshake has completed.