Waitress WSGI server
Waitress is meant to be a production-quality pure-Python WSGI server with very acceptable performance. It has no dependencies except ones which live in the Python standard library. It runs on CPython on Unix and Windows under Python 2.6+ and Python 3.2+. It is also known to run on PyPy 1.6.0+ on UNIX. It supports HTTP/1.0 and HTTP/1.1.
For more information, see the “docs” directory of the Waitress package or http://docs.pylonsproject.org/projects/waitress/en/latest/ .
- Fix tests under Windows. NB: to run tests under Windows, you cannot run “setup.py test” or “setup.py nosetests”. Instead you must run python.exe -c "import nose; nose.main()". If you try to run the tests using the normal method under Windows, each subprocess created by the test suite will attempt to run the test suite again. See https://github.com/nose-devs/nose/issues/407 for more information.
- Give the WSGI app_iter generated when wsgi.file_wrapper is used (ReadOnlyFileBasedBuffer) a close method. Do not call close on an instance of such a class when it’s used as a WSGI app_iter, however. This is part of a fix which prevents a leakage of file descriptors; the other part of the fix was in WebOb (https://github.com/Pylons/webob/commit/951a41ce57bd853947f842028bccb500bd5237da).
- Allow trusted proxies to override wsgi.url_scheme via a request header, X_FORWARDED_PROTO. Allows proxies which serve mixed HTTP / HTTPS requests to control signal which are served as HTTPS. See https://github.com/Pylons/waitress/pull/42.
- Fix some cases where the creation of extremely large output buffers (greater than 2GB, suspected to be buffers added via wsgi.file_wrapper) might cause an OverflowError on Python 2. See https://github.com/Pylons/waitress/issues/47.
- When the url_prefix adjustment starts with more than one slash, all slashes except one will be stripped from its beginning. This differs from older behavior where more than one leading slash would be preserved in url_prefix.
- If a client somehow manages to send an empty path, we no longer convert the empty path to a single slash in PATH_INFO. Instead, the path remains empty. According to RFC 2616 section “5.1.2 Request-URI”, the scenario of a client sending an empty path is actually not possible because the request URI portion cannot be empty.
- If the url_prefix adjustment matches the request path exactly, we now compute SCRIPT_NAME and PATH_INFO properly. Previously, if the url_prefix was /foo and the path received from a client was /foo, we would set both SCRIPT_NAME and PATH_INFO to /foo. This was incorrect. Now in such a case we set PATH_INFO to the empty string and we set SCRIPT_NAME to /foo. Note that the change we made has no effect on paths that do not match the url_prefix exactly (such as /foo/bar); these continue to operate as they did. See https://github.com/Pylons/waitress/issues/46
- Preserve header ordering of headers with the same name as per RFC 2616. See https://github.com/Pylons/waitress/pull/44
- When waitress receives a Transfer-Encoding: chunked request, we no longer send the TRANSFER_ENCODING nor the HTTP_TRANSFER_ENCODING value to the application in the environment. Instead, we pop this header. Since we cope with chunked requests by buffering the data in the server, we also know when a chunked request has ended, and therefore we know the content length. We set the content-length header in the environment, such that applications effectively never know the original request was a T-E: chunked request; it will appear to them as if the request is a non-chunked request with an accurate content-length.
- Cope with the fact that the Transfer-Encoding value is case-insensitive.
- When the --unix-socket-perms option was used as an argument to waitress-serve, a TypeError would be raised. See https://github.com/Pylons/waitress/issues/50.
- The HTTP version of the response returned by waitress when it catches an exception will now match the HTTP request version.
- Fix: CONNECTION header will be HTTP_CONNECTION and not CONNECTION_TYPE (see https://github.com/Pylons/waitress/issues/13)
- Do alternate type of checking for UNIX socket support, instead of checking for platform == windows.
- Functional tests now use multiprocessing module instead of subprocess module, speeding up test suite and making concurrent execution more reliable.
- Runner now appends the current working directory to sys.path to support running WSGI applications from a directory (i.e., not installed in a virtualenv).
- Add a url_prefix adjustment setting. You can use it by passing script_name='/foo' to waitress.serve or you can use it in a PasteDeploy ini file as script_name = /foo. This will cause the WSGI SCRIPT_NAME value to be the value passed minus any trailing slashes you add, and it will cause the PATH_INFO of any request which is prefixed with this value to be stripped of the prefix. You can use this instead of PasteDeploy’s prefixmiddleware to always prefix the path.
- Fix runner multisegment imports in some Python 2 revisions (see https://github.com/Pylons/waitress/pull/34).
- For compatibility, WSGIServer is now an alias of TcpWSGIServer. The signature of BaseWSGIServer is now compatible with WSGIServer pre-0.8.4.
- Add a command-line runner called waitress-serve to allow Waitress to run WSGI applications without any addional machinery. This is essentially a thin wrapper around the waitress.serve() function.
- Allow parallel testing (e.g., under detox or nosetests --processes) using PID-dependent port / socket for functest servers.
- Fix integer overflow errors on large buffers. Thanks to Marcin Kuzminski for the patch. See: https://github.com/Pylons/waitress/issues/22
- Add support for listening on Unix domain sockets.
- Add an asyncore_loop_timeout adjustment value, which controls the timeout value passed to asyncore.loop; defaults to 1.
- The default asyncore loop timeout is now 1 second. This prevents slow shutdown on Windows. See https://github.com/Pylons/waitress/issues/6 . This shouldn’t matter to anyone in particular, but it can be changed via the asyncore_loop_timeout adjustment (it used to previously default to 30 seconds).
- Don’t complain if there’s a response to a HEAD request that contains a Content-Length > 0. See https://github.com/Pylons/waitress/pull/7.
- Fix bug in HTTP Expect/Continue support. See https://github.com/Pylons/waitress/issues/9 .
- http://corte.si/posts/code/pathod/pythonservers/index.html pointed out that sending a bad header resulted in an exception leading to a 500 response instead of the more proper 400 response without an exception.
- Fix a race condition in the test suite.
- Allow “ident” to be used as a keyword to serve() as per docs.
- Add py33 to tox.ini.
- A brown-bag bug prevented request concurrency. A slow request would block subsequent the responses of subsequent requests until the slow request’s response was fully generated. This was due to a “task lock” being declared as a class attribute rather than as an instance attribute on HTTPChannel. Also took the opportunity to move another lock named “outbuf lock” to the channel instance rather than the class. See https://github.com/Pylons/waitress/pull/1 .
Support the WSGI wsgi.file_wrapper protocol as per http://www.python.org/dev/peps/pep-0333/#optional-platform-specific-file-handling. Here’s a usage example:
import os here = os.path.dirname(os.path.abspath(__file__)) def myapp(environ, start_response): f = open(os.path.join(here, 'myphoto.jpg'), 'rb') headers = [('Content-Type', 'image/jpeg')] start_response( '200 OK', headers ) return environ['wsgi.file_wrapper'](f, 32768)
The signature of the file wrapper constructor is (filelike_object, block_size). Both arguments must be passed as positional (not keyword) arguments. The result of creating a file wrapper should be returned as the app_iter from a WSGI application.
The object passed as filelike_object to the wrapper must be a file-like object which supports at least the read() method, and the read() method must support an optional size hint argument. It should support the seek() and tell() methods. If it does not, normal iteration over the filelike object using the provided block_size is used (and copying is done, negating any benefit of the file wrapper). It should support a close() method.
The specified block_size argument to the file wrapper constructor will be used only when the filelike_object doesn’t support seek and/or tell methods. Waitress needs to use normal iteration to serve the file in this degenerate case (as per the WSGI spec), and this block size will be used as the iteration chunk size. The block_size argument is optional; if it is not passed, a default value``32768`` is used.
Waitress will set a Content-Length header on the behalf of an application when a file wrapper with a sufficiently filelike object is used if the application hasn’t already set one.
The machinery which handles a file wrapper currently doesn’t do anything particularly special using fancy system calls (it doesn’t use sendfile for example); using it currently just prevents the system from needing to copy data to a temporary buffer in order to send it to the client. No copying of data is done when a WSGI app returns a file wrapper that wraps a sufficiently filelike object. It may do something fancier in the future.
- Default send_bytes value is now 18000 instead of 9000. The larger default value prevents asyncore from needing to execute select so many times to serve large files, speeding up file serving by about 15%-20% or so. This is probably only an optimization for LAN communications, and could slow things down across a WAN (due to higher TCP overhead), but we’re likely to be behind a reverse proxy on a LAN anyway if in production.
- Added an (undocumented) profiling feature to the serve() command.
- Remove performance-sapping call to pull_trigger in the channel’s write_soon method added mistakenly in 0.6.
- A logic error prevented the internal outbuf buffer of a channel from being flushed when the client could not accept the entire contents of the output buffer in a single succession of socket.send calls when the channel was in a “pending close” state. The socket in such a case would be closed prematurely, sometimes resulting in partially delivered content. This was discovered by a user using waitress behind an Nginx reverse proxy, which apparently is not always ready to receive data. The symptom was that he received “half” of a large CSS file (110K) while serving content via waitress behind the proxy.
- Fix PATH_INFO encoding/decoding on Python 3 (as per PEP 3333, tunnel bytes-in-unicode-as-latin-1-after-unquoting).
- Added “design” document to docs.
- Set default connection_limit back to 100 for benefit of maximal platform compatibility.
- Normalize setting of last_activity during send.
- Minor resource cleanups during tests.
- Channel timeout cleanup was broken.
- Dont hang a thread up trying to send data to slow clients.
- Use self.logger to log socket errors instead of self.log_info (normalize).
- Remove pointless handle_error method from channel.
- Queue requests instead of tasks in a channel.
- Expect: 100-continue responses were broken.
- Set up logging by calling logging.basicConfig() when serve is called (show tracebacks and other warnings to console by default).
- Disallow WSGI applications to set “hop-by-hop” headers (Connection, Transfer-Encoding, etc).
- Don’t treat 304 status responses specially in HTTP/1.1 mode.
- Remove out of date interfaces.py file.
- Normalize logging (all output is now sent to the waitress logger rather than in degenerate cases some output being sent directly to stderr).
- Support HTTP/1.1 Transfer-Encoding: chunked responses.
- Slightly better docs about logging.
- Initial release.