Waiting for File Descriptor Events

Title:Waiting for File Descriptor Events
Author:Christopher Stawarz <chris@pseudogreen.org>
Discussions-To:Python Web-SIG <web-sig@python.org>
Status:Proposed
Created:11-May-2008

Abstract

This specification defines a set of extensions that allow a WSGI application to suspend its execution until an event occurs on a specified file descriptor.

Rationale

The architecture of asynchronous (aka event driven) servers requires all I/O operations, including both interprocess and network communication, to be non-blocking. For a WSGI-compliant server, this requirement extends to all applications run on the server. However, the WSGI specification does not provide sufficient facilities for an application to ensure that its I/O is non-blocking. Specifically, it lacks a mechanism by which an application can suspend its execution until an arbitrary file descriptor (such as one belonging to a socket or pipe opened by the application) is ready for reading or writing. This specification defines a standard interface by which servers can provide such a mechanism to applications.

Specification

This specification introduces three new variables to the WSGI environment: x-wsgiorg.fdevent.readable, x-wsgiorg.fdevent.writable, and x-wsgiorg.fdevent.timeout.

The variables x-wsgiorg.fdevent.readable and x-wsgiorg.fdevent.writable are callable objects that accept two positional arguments, one required and one optional. In the following description, these arguments are given the names fd and timeout, but they are not required to have these names, and the application must invoke the callables using positional arguments.

The first argument, fd, is either an integer representing a file descriptor or an object with a fileno method that returns such an integer. The set of acceptable file descriptors is defined to be those accepted by select.select. (Note that this set is platform dependent: only sockets are allowed on Windows, whereas sockets, pipes, and files are acceptable on Unix-like systems.) The second, optional argument, timeout, is either None or a floating-point value in seconds. If omitted, it defaults to None.

When called, x-wsgiorg.fdevent.readable and x-wsgiorg.fdevent.writable return the empty string (''), which must be yielded by the application iterable to the server. (The result of calling x-wsgiorg.fdevent.readable or x-wsgiorg.fdevent.writable and yielding a non-empty string, or making multiple calls to x-wsgiorg.fdevent.readable and/or x-wsgiorg.fdevent.writable before yielding the empty string, is undefined.) The server then suspends execution of the application until one of the following conditions is met:

  • The specified file descriptor is ready for reading (if the application called x-wsgiorg.fdevent.readable) or writing (if the application called x-wsgiorg.fdevent.writable).
  • timeout seconds have elapsed without the desired file descriptor becoming readable (if the application called x-wsgiorg.fdevent.readable) or writable (if the application called x-wsgiorg.fdevent.writable), unless the value of timeout is None, in which case the wait will never timeout.
  • The server detects an error or “exceptional” condition (such as out-of-band data) on the file descriptor.

Put another way, if the application calls x-wsgiorg.fdevent.readable and yields the empty string, it will be suspended until select.select([fd],[],[fd],timeout) would return. If the application calls x-wsgiorg.fdevent.writable and yields the empty string, it will be suspended until select.select([],[fd],[fd],timeout) would return.

The variable x-wsgiorg.fdevent.timeout is an object whose truth value can be changed by the server. (For example, it could be a list instance, whose truth value is false when empty, true otherwise.) If timeout seconds elapse without the desired file descriptor event occurring, x-wsgiorg.fdevent.timeout will be true when the application resumes; otherwise, it will be false. The truth value of x-wsgiorg.fdevent.timeout when the application is first started or after it yields each response-body string is undefined.

The server may use any technique it desires to detect events on an application’s file descriptors. (Most likely, it will add them to the same event loop that it uses for accepting new client connections, receiving requests, and sending responses.)

Handling of the Input Stream

While technically outside the scope of this specification, the application’s input stream (wsgi.input) is another source of potentially blocking I/O that deserves mention.

The methods provided by the input stream follow the semantics of the corresponding methods of the file class. In particular, each of these methods can invoke the underlying I/O function (in this case, recv on the socket connected to the client) more than once, without giving the application the opportunity to check whether each invocation will block. Although authors of asynchronous servers may be tempted to provide a non-standard input stream that supports on-demand, non-blocking reads, such an input stream would be incompatible with WSGI middleware.

In order to avoid these problems, it is strongly recommended that asynchronous servers pre-read the entire request body (to an in-memory buffer or temporary file) before invoking the application, either by default or as a configurable option. Doing so will ensure that the input stream is compatible with middleware and that reads from it will not block waiting for data from the client.

Examples

The following application acts as a proxy to python.org. It uses a pycurl.CurlMulti instance to perform the outgoing HTTP request in a non-blocking fashion. When the CurlMulti.perform() method detects that its next I/O operation would block, it returns control to the application, which then yields until the file descriptor of interest becomes readable or writable as required. If the descriptor is not ready after one second, the application sends a 504 Gateway Timeout response to the client and terminates:

def pyorg_proxy(environ, start_response):
    result = StringIO()

    c = pycurl.Curl()
    c.setopt(pycurl.URL, 'http://python.org' + environ['PATH_INFO'])
    c.setopt(pycurl.WRITEFUNCTION, result.write)

    m = pycurl.CurlMulti()
    m.add_handle(c)

    while True:
        while True:
            ret, num_handles = m.perform()
            if ret != pycurl.E_CALL_MULTI_PERFORM:
                break
        if not num_handles:
            break

        read, write, exc = m.fdset()
        if read:
            yield environ['x-wsgiorg.fdevent.readable'](read[0], 1.0)
        else:
            yield environ['x-wsgiorg.fdevent.writable'](write[0], 1.0)

        if environ['x-wsgiorg.fdevent.timeout']:
            msg = 'The request to python.org timed out.'
            start_response('504 Gateway Timeout',
                           [('Content-Type', 'text/plain'),
                            ('Content-Length', str(len(msg)))])
            yield msg
            return

    start_response('200 OK', [('Content-Type', 'application/octet-stream'),
                              ('Content-Length', str(result.len))])
    yield result.getvalue()

The following adapter allows an application that uses the x-wsgiorg.fdevent extensions to run on a server that does not support them, without any modification to the application’s code:

def with_fdevent(application):
    def wrapper(environ, start_response):
        select_args = [None]

        def readable(fd, timeout=None):
            assert (not select_args[0])
            select_args[0] = ([fd], [], [fd], timeout)
            return ''

        def writable(fd, timeout=None):
            assert (not select_args[0])
            select_args[0] = ([], [fd], [fd], timeout)
            return ''

        environ['x-wsgiorg.fdevent.readable'] = readable
        environ['x-wsgiorg.fdevent.writable'] = writable

        timeout = False

        class TimeoutWrapper(object):
            def __nonzero__(self):
                return timeout

        environ['x-wsgiorg.fdevent.timeout'] = TimeoutWrapper()

        for result in application(environ, start_response):
            assert (not (result and select_args[0]))
            if result or (not select_args[0]):
                yield result
            else:
                ready = select.select(*select_args[0])
                timeout = (ready == ([], [], []))
                select_args[0] = None

    return wrapper

Problems

  • The empty string yielded by an application after calling x-wsgiorg.fdevent.readable or x-wsgiorg.fdevent.writable must pass through any intervening middleware and be detected by the server. Although WSGI explicitly requires middleware to relay such strings to the server (see Middleware Handling of Block Boundaries), some components may not, making them incompatible with this specification.

Other Possibilities

  • To prevent an application that does blocking I/O from blocking the entire server, an asynchronous server could run each instance of the application in a separate thread. However, since asynchronous servers achieve high levels of concurrency by expressly avoiding multithreading, this technique will almost always be unacceptable.
  • The greenlet package enables the use of cooperatively-scheduled micro-threads in Python programs, and a WSGI server could potentially use it to pause and resume applications around blocking I/O operations. However, such micro-threading is not part of the Python language or standard library, and some server authors may be unwilling or unable to make use of it.

Open Issues

  • Some third-party libraries (such as PycURL) provide non-blocking interfaces that may need to monitor multiple file descriptors for events simultaneously. Since this specification allows an application to wait on only one file descriptor at a time, application authors may find it difficult or impossible to use such libraries, or they may be limited to a subset of the libraries’ capabilities.

    Although this specification could be extended to include an interface for waiting on multiple file descriptors, it is unclear whether it would be easy (or even possible) for all servers to implement it. Also, the appropriate behavior for a multi-descriptor wait is not obvious. (Should the application be resumed when a single descriptor is ready? All of them? Some minimum number?)