skip to navigation
skip to content

Python Wiki

Python Insider Blog

Python 2 or 3?

Help Fund Python

[Python resources in languages other than English]

Non-English Resources

Add an event to this calendar.

Times are shown in UTC/GMT.

Add an event to this calendar.

Title:Asynchronous IO Support Rebooted: the "asyncio" Module
Last-Modified:2016-05-03 12:03:16 +0300 (Tue, 03 May 2016)
Author:Guido van Rossum <guido at>
BDFL-Delegate:Antoine Pitrou <>
Discussions-To:<python-tulip at>
Type:Standards Track


This is a proposal for asynchronous I/O in Python 3, starting at Python 3.3. Consider this the concrete proposal that is missing from PEP 3153. The proposal includes a pluggable event loop, transport and protocol abstractions similar to those in Twisted, and a higher-level scheduler based on yield from (PEP 380). The proposed package name is asyncio.



A reference implementation exists under the code name Tulip. The Tulip repo is linked from the References section at the end. Packages based on this repo will be provided on PyPI (see References) to enable using the asyncio package with Python 3.3 installations.

As of October 20th 2013, the asyncio package has been checked into the Python 3.4 repository and released with Python 3.4-alpha-4, with "provisional" API status. This is an expression of confidence and intended to increase early feedback on the API, and not intended to force acceptance of the PEP. The expectation is that the package will keep provisional status in Python 3.4 and progress to final status in Python 3.5. Development continues to occur primarily in the Tulip repo, with changes occasionally merged into the CPython repo.


Python 3.3 is required for many of the proposed features. The reference implementation (Tulip) requires no new language or standard library features beyond Python 3.3, no third-party modules or packages, and no C code, except for the (optional) IOCP support on Windows.

Module Namespace

The specification here lives in a new top-level package, asyncio. Different components live in separate submodules of the package. The package will import common APIs from their respective submodules and make them available as package attributes (similar to the way the email package works). For such common APIs, the name of the submodule that actually defines them is not part of the specification. Less common APIs may have to explicitly be imported from their respective submodule, and in this case the submodule name is part of the specification.

Classes and functions defined without a submodule name are assumed to live in the namespace of the top-level package. (But do not confuse these with methods of various classes, which for brevity are also used without a namespace prefix in certain contexts.)


The event loop is the place where most interoperability occurs. It should be easy for (Python 3.3 ports of) frameworks like Twisted, Tornado, or even gevents to either adapt the default event loop implementation to their needs using a lightweight adapter or proxy, or to replace the default event loop implementation with an adaptation of their own event loop implementation. (Some frameworks, like Twisted, have multiple event loop implementations. This should not be a problem since these all have the same interface.)

In most cases it should be possible for two different third-party frameworks to interoperate, either by sharing the default event loop implementation (each using its own adapter), or by sharing the event loop implementation of either framework. In the latter case two levels of adaptation would occur (from framework A's event loop to the standard event loop interface, and from there to framework B's event loop). Which event loop implementation is used should be under control of the main program (though a default policy for event loop selection is provided).

For this interoperability to be effective, the preferred direction of adaptation in third party frameworks is to keep the default event loop and adapt it to the framework's API. Ideally all third party frameworks would give up their own event loop implementation in favor of the standard implementation. But not all frameworks may be satisfied with the functionality provided by the standard implementation.

In order to support both directions of adaptation, two separate APIs are specified:

  • An interface for managing the current event loop
  • The interface of a conforming event loop

An event loop implementation may provide additional methods and guarantees, as long as these are called out in the documentation as non-standard. An event loop implementation may also leave certain methods unimplemented if they cannot be implemented in the given environment; however, such deviations from the standard API should be considered only as a last resort, and only if the platform or environment forces the issue. (An example would be a platform where there is a system event loop that cannot be started or stopped; see "Embedded Event Loops" below.)

The event loop API does not depend on yield from. Rather, it uses a combination of callbacks, additional interfaces (transports and protocols), and Futures. The latter are similar to those defined in PEP 3148, but have a different implementation and are not tied to threads. In particular, the result() method raises an exception instead of blocking when a result is not yet ready; the user is expected to use callbacks (or yield from) to wait for the result.

All event loop methods specified as returning a coroutine are allowed to return either a Future or a coroutine, at the implementation's choice (the standard implementation always returns coroutines). All event loop methods documented as accepting coroutine arguments must accept both Futures and coroutines for such arguments. (A convenience function, async(), exists to convert an argument that is either a coroutine or a Future into a Future.)

For users (like myself) who don't like using callbacks, a scheduler is provided for writing asynchronous I/O code as coroutines using the PEP 380 yield from expressions. The scheduler is not pluggable; pluggability occurs at the event loop level, and the standard scheduler implementation should work with any conforming event loop implementation. (In fact this is an important litmus test for conforming implementations.)

For interoperability between code written using coroutines and other async frameworks, the scheduler defines a Task class that behaves like a Future. A framework that interoperates at the event loop level can wait for a Future to complete by adding a callback to the Future. Likewise, the scheduler offers an operation to suspend a coroutine until a callback is called.

The event loop API provides limited interoperability with threads: there is an API to submit a function to an executor (see PEP 3148) which returns a Future that is compatible with the event loop, and there is a method to schedule a callback with an event loop from another thread in a thread-safe manner.

Transports and Protocols

For those not familiar with Twisted, a quick explanation of the relationship between transports and protocols is in order. At the highest level, the transport is concerned with how bytes are transmitted, while the protocol determines which bytes to transmit (and to some extent when).

A different way of saying the same thing: a transport is an abstraction for a socket (or similar I/O endpoint) while a protocol is an abstraction for an application, from the transport's point of view.

Yet another view is simply that the transport and protocol interfaces together define an abstract interface for using network I/O and interprocess I/O.

There is almost always a 1:1 relationship between transport and protocol objects: the protocol calls transport methods to send data, while the transport calls protocol methods to pass it data that has been received. Neither transport not protocol methods "block" -- they set events into motion and then return.

The most common type of transport is a bidirectional stream transport. It represents a pair of buffered streams (one in each direction) that each transmit a sequence of bytes. The most common example of a bidirectional stream transport is probably a TCP connection. Another common example is an SSL/TLS connection. But there are some other things that can be viewed this way, for example an SSH session or a pair of UNIX pipes. Typically there aren't many different transport implementations, and most of them come with the event loop implementation. However, there is no requirement that all transports must be created by calling an event loop method: a third party module may well implement a new transport and provide a constructor or factory function for it that simply takes an event loop as an argument or calls get_event_loop().

Note that transports don't need to use sockets, not even if they use TCP -- sockets are a platform-specific implementation detail.

A bidirectional stream transport has two "ends": one end talks to the network (or another process, or whatever low-level interface it wraps), and the other end talks to the protocol. The former uses whatever API is necessary to implement the transport; but the interface between transport and protocol is standardized by this PEP.

A protocol can represent some kind of "application-level" protocol such as HTTP or SMTP; it can also implement an abstraction shared by multiple protocols, or a whole application. A protocol's primary interface is with the transport. While some popular protocols (and other abstractions) may have standard implementations, often applications implement custom protocols. It also makes sense to have libraries of useful third party protocol implementations that can be downloaded and installed from PyPI.

There general notion of transport and protocol includes other interfaces, where the transport wraps some other communication abstraction. Examples include interfaces for sending and receiving datagrams (e.g. UDP), or a subprocess manager. The separation of concerns is the same as for bidirectional stream transports and protocols, but the specific interface between transport and protocol is different in each case.

Details of the interfaces defined by the various standard types of transports and protocols are given later.

Event Loop Interface Specification

Event Loop Policy: Getting and Setting the Current Event Loop

Event loop management is controlled by an event loop policy, which is a global (per-process) object. There is a default policy, and an API to change the policy. A policy defines the notion of context; a policy manages a separate event loop per context. The default policy's notion of context is defined as the current thread.

Certain platforms or programming frameworks may change the default policy to something more suitable to the expectations of the users of that platform or framework. Such platforms or frameworks must document their policy and at what point during their initialization sequence the policy is set, in order to avoid undefined behavior when multiple active frameworks want to override the default policy. (See also "Embedded Event Loops" below.)

To get the event loop for current context, use get_event_loop(). This returns an event loop object implementing the interface specified below, or raises an exception in case no event loop has been set for the current context and the current policy does not specify to create one. It should never return None.

To set the event loop for the current context, use set_event_loop(event_loop), where event_loop is an event loop object, i.e. an instance of AbstractEventLoop, or None. It is okay to set the current event loop to None, in which case subsequent calls to get_event_loop() will raise an exception. This is useful for testing code that should not depend on the existence of a default event loop.

It is expected that get_event_loop() returns a different event loop object depending on the context (in fact, this is the definition of context). It may create a new event loop object if none is set and creation is allowed by the policy. The default policy will create a new event loop only in the main thread (as defined by, which uses a special subclass for the main thread), and only if get_event_loop() is called before set_event_loop() is ever called. (To reset this state, reset the policy.) In other threads an event loop must be explicitly set. Other policies may behave differently. Event loop by the default policy creation is lazy; i.e. the first call to get_event_loop() creates an event loop instance if necessary and specified by the current policy.

For the benefit of unit tests and other special cases there's a third policy function: new_event_loop(), which creates and returns a new event loop object according to the policy's default rules. To make this the current event loop, you must call set_event_loop() with it.

To change the event loop policy, call set_event_loop_policy(policy), where policy is an event loop policy object or None. If not None, the policy object must be an instance of AbstractEventLoopPolicy that defines methods get_event_loop(), set_event_loop(loop) and new_event_loop(), all behaving like the functions described above.

Passing a policy value of None restores the default event loop policy (overriding the alternate default set by the platform or framework). The default event loop policy is an instance of the class DefaultEventLoopPolicy. The current event loop policy object can be retrieved by calling get_event_loop_policy().

Passing an Event Loop Around Explicitly

It is possible to write code that uses an event loop without relying on a global or per-thread default event loop. For this purpose, all APIs that need access to the current event loop (and aren't methods on an event class) take an optional keyword argument named loop. If this argument is None or unspecified, such APIs will call get_event_loop() to get the default event loop, but if the loop keyword argument is set to an event loop object, they will use that event loop, and pass it along to any other such APIs they call. For example, Future(loop=my_loop) will create a Future tied to the event loop my_loop. When the default current event is None, the loop keyword argument is effectively mandatory.

Note that an explicitly passed event loop must still belong to the current thread; the loop keyword argument does not magically change the constraints on how an event loop can be used.

Specifying Times

As usual in Python, all timeouts, intervals and delays are measured in seconds, and may be ints or floats. However, absolute times are not specified as POSIX timestamps. The accuracy, precision and epoch of the clock are up to the implementation.

The default implementation uses time.monotonic(). Books could be written about the implications of this choice. Better read the docs for the standard library time module.

Embedded Event Loops

On some platforms an event loop is provided by the system. Such a loop may already be running when the user code starts, and there may be no way to stop or close it without exiting from the program. In this case, the methods for starting, stopping and closing the event loop may not be implementable, and is_running() may always return True.

Event Loop Classes

There is no actual class named EventLoop. There is an AbstractEventLoop class which defines all the methods without implementations, and serves primarily as documentation. The following concrete classes are defined:

  • SelectorEventLoop is a concrete implementation of the full API based on the selectors module (new in Python 3.4). The constructor takes one optional argument, a selectors.Selector object. By default an instance of selectors.DefaultSelector is created and used.
  • ProactorEventLoop is a concrete implementation of the API except for the I/O event handling and signal handling methods. It is only defined on Windows (or on other platforms which support a similar API for "overlapped I/O"). The constructor takes one optional argument, a Proactor object. By default an instance of IocpProactor is created and used. (The IocpProactor class is not specified by this PEP; it is just an implementation detail of the ProactorEventLoop class.)

Event Loop Methods Overview

The methods of a conforming event loop are grouped into several categories. The first set of categories must be supported by all conforming event loop implementations, with the exception that embedded event loops may not implement the methods for starting, stopping and closing. (However, a partially-conforming event loop is still better than nothing. :-)

  • Starting, stopping and closing: run_forever(), run_until_complete(), stop(), is_running(), close().
  • Basic and timed callbacks: call_soon(), call_later(), call_at(), time().
  • Thread interaction: call_soon_threadsafe(), run_in_executor(), set_default_executor().
  • Internet name lookups: getaddrinfo(), getnameinfo().
  • Internet connections: create_connection(), create_server(), create_datagram_endpoint().
  • Wrapped socket methods: sock_recv(), sock_sendall(), sock_connect(), sock_accept().

The second set of categories may be supported by conforming event loop implementations. If not supported, they will raise NotImplementedError. (In the default implementation, SelectorEventLoop on UNIX systems supports all of these; SelectorEventLoop on Windows supports the I/O event handling category; ProactorEventLoop on Windows supports the pipes and subprocess category.)

  • I/O callbacks: add_reader(), remove_reader(), add_writer(), remove_writer().
  • Pipes and subprocesses: connect_read_pipe(), connect_write_pipe(), subprocess_shell(), subprocess_exec().
  • Signal callbacks: add_signal_handler(), remove_signal_handler().

Event Loop Methods

Starting, Stopping and Closing

An (unclosed) event loop can be in one of two states: running or stopped. These methods deal with starting and stopping an event loop:

  • run_forever(). Runs the event loop until stop() is called. This cannot be called when the event loop is already running. (This has a long name in part to avoid confusion with earlier versions of this PEP, where run() had different behavior, in part because there are already too many APIs that have a method named run(), and in part because there shouldn't be many places where this is called anyway.)
  • run_until_complete(future). Runs the event loop until the Future is done. If the Future is done, its result is returned, or its exception is raised. This cannot be called when the event loop is already running.
  • stop(). Stops the event loop as soon as it is convenient. It is fine to restart the loop with run_forever() or run_until_complete() subsequently; no scheduled callbacks will be lost if this is done. Note: stop() returns normally and the current callback is allowed to continue. How soon after this point the event loop stops is up to the implementation, but the intention is to stop short of polling for I/O, and not to run any callbacks scheduled in the future; the major freedom an implementation has is how much of the "ready queue" (callbacks already scheduled with call_soon()) it processes before stopping.
  • is_running(). Returns True if the event loop is currently running, False if it is stopped.
  • close(). Closes the event loop, releasing any resources it may hold, such as the file descriptor used by epoll() or kqueue(), and the default executor. This should not be called while the event loop is running. After it has been called the event loop should not be used again. It may be called multiple times; subsequent calls are no-ops.

Basic Callbacks

Callbacks associated with the same event loop are strictly serialized: one callback must finish before the next one will be called. This is an important guarantee: when two or more callbacks use or modify shared state, each callback is guaranteed that while it is running, the shared state isn't changed by another callback.

  • call_soon(callback, *args). This schedules a callback to be called as soon as possible. Returns a Handle (see below) representing the callback, whose cancel() method can be used to cancel the callback. It guarantees that callbacks are called in the order in which they were scheduled.
  • call_later(delay, callback, *args). Arrange for callback(*args) to be called approximately delay seconds in the future, once, unless cancelled. Returns a Handle representing the callback, whose cancel() method can be used to cancel the callback. Callbacks scheduled in the past or at exactly the same time will be called in an undefined order.
  • call_at(when, callback, *args). This is like call_later(), but the time is expressed as an absolute time. Returns a similar Handle. There is a simple equivalency: loop.call_later(delay, callback, *args) is the same as loop.call_at(loop.time() + delay, callback, *args).
  • time(). Returns the current time according to the event loop's clock. This may be time.time() or time.monotonic() or some other system-specific clock, but it must return a float expressing the time in units of approximately one second since some epoch. (No clock is perfect -- see PEP 418.)

Note: A previous version of this PEP defined a method named call_repeatedly(), which promised to call a callback at regular intervals. This has been withdrawn because the design of such a function is overspecified. On the one hand, a simple timer loop can easily be emulated using a callback that reschedules itself using call_later(); it is also easy to write coroutine containing a loop and a sleep() call (a toplevel function in the module, see below). On the other hand, due to the complexities of accurate timekeeping there are many traps and pitfalls here for the unaware (see PEP 418), and different use cases require different behavior in edge cases. It is impossible to offer an API for this purpose that is bullet-proof in all cases, so it is deemed better to let application designers decide for themselves what kind of timer loop to implement.

Thread interaction

  • call_soon_threadsafe(callback, *args). Like call_soon(callback, *args), but when called from another thread while the event loop is blocked waiting for I/O, unblocks the event loop. Returns a Handle. This is the only method that is safe to call from another thread. (To schedule a callback for a later time in a threadsafe manner, you can use loop.call_soon_threadsafe(loop.call_later, when, callback, *args).) Note: this is not safe to call from a signal handler (since it may use locks). In fact, no API is signal-safe; if you want to handle signals, use add_signal_handler() described below.
  • run_in_executor(executor, callback, *args). Arrange to call callback(*args) in an executor (see PEP 3148). Returns an asyncio.Future instance whose result on success is the return value of that call. This is equivalent to wrap_future(executor.submit(callback, *args)). If executor is None, the default executor set by set_default_executor() is used. If no default executor has been set yet, a ThreadPoolExecutor with a default number of threads is created and set as the default executor. (The default implementation uses 5 threads in this case.)
  • set_default_executor(executor). Set the default executor used by run_in_executor(). The argument must be a PEP 3148 Executor instance or None, in order to reset the default executor.

See also the wrap_future() function described in the section about Futures.

Internet name lookups

These methods are useful if you want to connect or bind a socket to an address without the risk of blocking for the name lookup. They are usually called implicitly by create_connection(), create_server() or create_datagram_endpoint().

  • getaddrinfo(host, port, family=0, type=0, proto=0, flags=0). Similar to the socket.getaddrinfo() function but returns a Future. The Future's result on success will be a list of the same format as returned by socket.getaddrinfo(), i.e. a list of (address_family, socket_type, socket_protocol, canonical_name, address) where address is a 2-tuple (ipv4_address, port) for IPv4 addresses and a 4-tuple (ipv4_address, port, flow_info, scope_id) for IPv6 addresses. If the family argument is zero or unspecified, the list returned may contain a mixture of IPv4 and IPv6 addresses; otherwise the addresses returned are constrained by the family value (similar for proto and flags). The default implementation calls socket.getaddrinfo() using run_in_executor(), but other implementations may choose to implement their own DNS lookup. The optional arguments must be specified as keyword arguments.

    Note: implementations are allowed to implement a subset of the full socket.getaddrinfo() interface; e.g. they may not support symbolic port names, or they may ignore or incompletely implement the type, proto and flags arguments. However, if type and proto are ignored, the argument values passed in should be copied unchanged into the return tuples' socket_type and socket_protocol elements. (You can't ignore family, since IPv4 and IPv6 addresses must be looked up differently. The only permissible values for family are socket.AF_UNSPEC (0), socket.AF_INET and socket.AF_INET6, and the latter only if it is defined by the platform.)

  • getnameinfo(sockaddr, flags=0). Similar to socket.getnameinfo() but returns a Future. The Future's result on success will be a tuple (host, port). Same implementation remarks as for getaddrinfo().

Internet connections

These are the high-level interfaces for managing internet connections. Their use is recommended over the corresponding lower-level interfaces because they abstract away the differences between selector-based and proactor-based event loops.

Note that the client and server side of stream connections use the same transport and protocol interface. However, datagram endpoints use a different transport and protocol interface.

  • create_connection(protocol_factory, host, port, <options>). Creates a stream connection to a given internet host and port. This is a task that is typically called from the client side of the connection. It creates an implementation-dependent bidirectional stream Transport to represent the connection, then calls protocol_factory() to instantiate (or retrieve) the user's Protocol implementation, and finally ties the two together. (See below for the definitions of Transport and Protocol.) The user's Protocol implementation is created or retrieved by calling protocol_factory() without arguments(*). The coroutine's result on success is the (transport, protocol) pair; if a failure prevents the creation of a successful connection, an appropriate exception will be raised. Note that when the coroutine completes, the protocol's connection_made() method has not yet been called; that will happen when the connection handshake is complete.

    (*) There is no requirement that protocol_factory is a class. If your protocol class needs to have specific arguments passed to its constructor, you can use lambda. You can also pass a trivial lambda that returns a previously constructed Protocol instance.

    The <options> are all specified using optional keyword arguments:

    • ssl: Pass True to create an SSL/TLS transport (by default a plain TCP transport is created). Or pass an ssl.SSLContext object to override the default SSL context object to be used. If a default context is created it is up to the implementation to configure reasonable defaults. The reference implementation currently uses PROTOCOL_SSLv23 and sets the OP_NO_SSLv2 option, calls set_default_verify_paths() and sets verify_mode to CERT_REQUIRED. In addition, whenever the context (default or otherwise) specifies a verify_mode of CERT_REQUIRED or CERT_OPTIONAL, if a hostname is given, immediately after a successful handshake ssl.match_hostname(peercert, hostname) is called, and if this raises an exception the conection is closed. (To avoid this behavior, pass in an SSL context that has verify_mode set to CERT_NONE. But this means you are not secure, and vulnerable to for example man-in-the-middle attacks.)
    • family, proto, flags: Address family, protocol and flags to be passed through to getaddrinfo(). These all default to 0, which means "not specified". (The socket type is always SOCK_STREAM.) If any of these values are not specified, the getaddrinfo() method will choose appropriate values. Note: proto has nothing to do with the high-level Protocol concept or the protocol_factory argument.
    • sock: An optional socket to be used instead of using the host, port, family, proto and flags arguments. If this is given, host and port must be explicitly set to None.
    • local_addr: If given, a (host, port) tuple used to bind the socket to locally. This is rarely needed but on multi-homed servers you occasionally need to force a connection to come from a specific address. This is how you would do that. The host and port are looked up using getaddrinfo().
    • server_hostname: This is only relevant when using SSL/TLS; it should not be used when ssl is not set. When ssl is set, this sets or overrides the hostname that will be verified. By default the value of the host argument is used. If host is empty, there is no default and you must pass a value for server_hostname. To disable hostname verification (which is a serious security risk) you must pass an empty string here and pass an ssl.SSLContext object whose verify_mode is set to ssl.CERT_NONE as the ssl argument.
  • create_server(protocol_factory, host, port, <options>). Enters a serving loop that accepts connections. This is a coroutine that completes once the serving loop is set up to serve. The return value is a Server object which can be used to stop the serving loop in a controlled fashion (see below). Multiple sockets may be bound if the specified address allows both IPv4 and IPv6 connections.

    Each time a connection is accepted, protocol_factory is called without arguments(**) to create a Protocol, a bidirectional stream Transport is created to represent the network side of the connection, and the two are tied together by calling protocol.connection_made(transport).

    (**) See previous footnote for create_connection(). However, since protocol_factory() is called once for each new incoming connection, it should return a new Protocol object each time it is called.

    The <options> are all specified using optional keyword arguments:

    • ssl: Pass an ssl.SSLContext object (or an object with the same interface) to override the default SSL context object to be used. (Unlike for create_connection(), passing True does not make sense here -- the SSLContext object is needed to specify the certificate and key.)

    • backlog: Backlog value to be passed to the listen() call. The default is implementation-dependent; in the default implementation the default value is 100.

    • reuse_address: Whether to set the SO_REUSEADDR option on the socket. The default is True on UNIX, False on Windows.

    • family, flags: Address family and flags to be passed

      through to getaddrinfo(). The family defaults to AF_UNSPEC; the flags default to AI_PASSIVE. (The socket type is always SOCK_STREAM; the socket protocol always set to 0, to let getaddrinfo() choose.)

    • sock: An optional socket to be used instead of using the host, port, family and flags arguments. If this is given, host and port must be explicitly set to None.

  • create_datagram_endpoint(protocol_factory, local_addr=None, remote_addr=None, <options>). Creates an endpoint for sending and receiving datagrams (typically UDP packets). Because of the nature of datagram traffic, there are no separate calls to set up client and server side, since usually a single endpoint acts as both client and server. This is a coroutine that returns a (transport, protocol) pair on success, or raises an exception on failure. If the coroutine returns successfully, the transport will call callbacks on the protocol whenever a datagram is received or the socket is closed; it is up to the protocol to call methods on the protocol to send datagrams. The transport returned is a DatagramTransport. The protocol returned is a DatagramProtocol. These are described later.

    Mandatory positional argument:

    • protocol_factory: A class or factory function that will be called exactly once, without arguments, to construct the protocol object to be returned. The interface between datagram transport and protocol is described below.

    Optional arguments that may be specified positionally or as keyword arguments:

    • local_addr: An optional tuple indicating the address to which the socket will be bound. If given this must be a (host, port) pair. It will be passed to getaddrinfo() to be resolved and the result will be passed to the bind() method of the socket created. If getaddrinfo() returns more than one address, they will be tried in turn. If omitted, no bind() call will be made.
    • remote_addr: An optional tuple indicating the address to which the socket will be "connected". (Since there is no such thing as a datagram connection, this just specifies a default value for the destination address of outgoing datagrams.) If given this must be a (host, port) pair. It will be passed to getaddrinfo() to be resolved and the result will be passed to sock_connect() together with the socket created. If getaddrinfo() returns more than one address, they will be tried in turn. If omitted, no sock_connect() call will be made.

    The <options> are all specified using optional keyword arguments:

    • family, proto, flags: Address family, protocol and flags to be passed through to getaddrinfo(). These all default to 0, which means "not specified". (The socket type is always SOCK_DGRAM.) If any of these values are not specified, the getaddrinfo() method will choose appropriate values.

    Note that if both local_addr and remote_addr are present, all combinations of local and remote addresses with matching address family will be tried.

Wrapped Socket Methods

The following methods for doing async I/O on sockets are not for general use. They are primarily meant for transport implementations working with IOCP through the ProactorEventLoop class. However, they are easily implementable for other event loop types, so there is no reason not to require them. The socket argument has to be a non-blocking socket.

  • sock_recv(sock, n). Receive up to n bytes from socket sock. Returns a Future whose result on success will be a bytes object.
  • sock_sendall(sock, data). Send bytes data to socket sock. Returns a Future whose result on success will be None. Note: the name uses sendall instead of send, to reflect that the semantics and signature of this method echo those of the standard library socket method sendall() rather than send().
  • sock_connect(sock, address). Connect to the given address. Returns a Future whose result on success will be None.
  • sock_accept(sock). Accept a connection from a socket. The socket must be in listening mode and bound to an address. Returns a Future whose result on success will be a tuple (conn, peer) where conn is a connected non-blocking socket and peer is the peer address.

I/O Callbacks

These methods are primarily meant for transport implementations working with a selector. They are implemented by SelectorEventLoop but not by ProactorEventLoop. Custom event loop implementations may or may not implement them.

The fd arguments below may be integer file descriptors, or "file-like" objects with a fileno() method that wrap integer file descriptors. Not all file-like objects or file descriptors are acceptable. Sockets (and socket file descriptors) are always accepted. On Windows no other types are supported. On UNIX, pipes and possibly tty devices are also supported, but disk files are not. Exactly which special file types are supported may vary by platform and per selector implementation. (Experimentally, there is at least one kind of pseudo-tty on OS X that is supported by select and poll but not by kqueue: it is used by Emacs shell windows.)

  • add_reader(fd, callback, *args). Arrange for callback(*args) to be called whenever file descriptor fd is deemed ready for reading. Calling add_reader() again for the same file descriptor implies a call to remove_reader() for the same file descriptor.
  • add_writer(fd, callback, *args). Like add_reader(), but registers the callback for writing instead of for reading.
  • remove_reader(fd). Cancels the current read callback for file descriptor fd, if one is set. If no callback is currently set for the file descriptor, this is a no-op and returns False. Otherwise, it removes the callback arrangement and returns True.
  • remove_writer(fd). This is to add_writer() as remove_reader() is to add_reader().

Pipes and Subprocesses

These methods are supported by SelectorEventLoop on UNIX and ProactorEventLoop on Windows.

The transports and protocols used with pipes and subprocesses differ from those used with regular stream connections. These are described later.

Each of the methods below has a protocol_factory argument, similar to create_connection(); this will be called exactly once, without arguments, to construct the protocol object to be returned.

Each method is a coroutine that returns a (transport, protocol) pair on success, or raises an exception on failure.

  • connect_read_pipe(protocol_factory, pipe): Create a unidrectional stream connection from a file-like object wrapping the read end of a UNIX pipe, which must be in non-blocking mode. The transport returned is a ReadTransport.
  • connect_write_pipe(protocol_factory, pipe): Create a unidrectional stream connection from a file-like object wrapping the write end of a UNIX pipe, which must be in non-blocking mode. The transport returned is a WriteTransport; it does not have any read-related methods. The protocol returned is a BaseProtocol.
  • subprocess_shell(protocol_factory, cmd, <options>): Create a subprocess from cmd, which is a string using the platform's "shell" syntax. This is similar to the standard library subprocess.Popen() class called with shell=True. The remaining arguments and return value are described below.
  • subprocess_exec(protocol_factory, *args, <options>): Create a subprocess from one or more string arguments, where the first string specifies the program to execute, and the remaining strings specify the program's arguments. (Thus, together the string arguments form the sys.argv value of the program, assuming it is a Python script.) This is similar to the standard library subprocess.Popen() class called with shell=False and the list of strings passed as the first argument; however, where Popen() takes a single argument which is list of strings, subprocess_exec() takes multiple string arguments. The remaining arguments and return value are described below.

Apart from the way the program to execute is specified, the two subprocess_*() methods behave the same. The transport returned is a SubprocessTransport which has a different interface than the common bidirectional stream transport. The protocol returned is a SubprocessProtocol which also has a custom interface.

The <options> are all specified using optional keyword arguments:

  • stdin: Either a file-like object representing the pipe to be connected to the subprocess's standard input stream using connect_write_pipe(), or the constant subprocess.PIPE (the default). By default a new pipe will be created and connected.
  • stdout: Either a file-like object representing the pipe to be connected to the subprocess's standard output stream using connect_read_pipe(), or the constant subprocess.PIPE (the default). By default a new pipe will be created and connected.
  • stderr: Either a file-like object representing the pipe to be connected to the subprocess's standard error stream using connect_read_pipe(), or one of the constants subprocess.PIPE (the default) or subprocess.STDOUT. By default a new pipe will be created and connected. When subprocess.STDOUT is specified, the subprocess's standard error stream will be connected to the same pipe as the standard output stream.
  • bufsize: The buffer size to be used when creating a pipe; this is passed to subprocess.Popen(). In the default implementation this defaults to zero, and on Windows it must be zero; these defaults deviate from subprocess.Popen().
  • executable, preexec_fn, close_fds, cwd, env, startupinfo, creationflags, restore_signals, start_new_session, pass_fds: These optional arguments are passed to subprocess.Popen() without interpretation.

Signal callbacks

These methods are only supported on UNIX.

  • add_signal_handler(sig, callback, *args). Whenever signal sig is received, arrange for callback(*args) to be called. Specifying another callback for the same signal replaces the previous handler (only one handler can be active per signal). The sig must be a valid signal number defined in the signal module. If the signal cannot be handled this raises an exception: ValueError if it is not a valid signal or if it is an uncatchable signal (e.g. SIGKILL), RuntimeError if this particular event loop instance cannot handle signals (since signals are global per process, only an event loop associated with the main thread can handle signals).
  • remove_signal_handler(sig). Removes the handler for signal sig, if one is set. Raises the same exceptions as add_signal_handler() (except that it may return False instead raising RuntimeError for uncatchable signals). Returns True if a handler was removed successfully, False if no handler was set.

Note: If these methods are statically known to be unsupported, they may raise NotImplementedError instead of RuntimeError.

Mutual Exclusion of Callbacks

An event loop should enforce mutual exclusion of callbacks, i.e. it should never start a callback while a previously callback is still running. This should apply across all types of callbacks, regardless of whether they are scheduled using call_soon(), call_later(), call_at(), call_soon_threadsafe(), add_reader(), add_writer(), or add_signal_handler().


There are two categories of exceptions in Python: those that derive from the Exception class and those that derive from BaseException. Exceptions deriving from Exception will generally be caught and handled appropriately; for example, they will be passed through by Futures, and they will be logged and ignored when they occur in a callback.

However, exceptions deriving only from BaseException are typically not caught, and will usually cause the program to terminate with a traceback. In some cases they are caught and re-raised. (Examples of this category include KeyboardInterrupt and SystemExit; it is usually unwise to treat these the same as most other exceptions.)


The various methods for registering one-off callbacks (call_soon(), call_later(), call_at() and call_soon_threadsafe()) all return an object representing the registration that can be used to cancel the callback. This object is called a Handle. Handles are opaque and have only one public method:

  • cancel(): Cancel the callback.

Note that add_reader(), add_writer() and add_signal_handler() do not return Handles.


The create_server() method returns a Server instance, which wraps the sockets (or other network objects) used to accept requests. This class has two public methods:

  • close(): Close the service. This stops accepting new requests but does not cancel requests that have already been accepted and are currently being handled.
  • wait_closed(): A coroutine that blocks until the service is closed and all accepted requests have been handled.


The asyncio.Future class here is intentionally similar to the concurrent.futures.Future class specified by PEP 3148, but there are slight differences. Whenever this PEP talks about Futures or futures this should be understood to refer to asyncio.Future unless concurrent.futures.Future is explicitly mentioned. The supported public API is as follows, indicating the differences with PEP 3148:

  • cancel(). If the Future is already done (or cancelled), do nothing and return False. Otherwise, this attempts to cancel the Future and returns True. If the cancellation attempt is successful, eventually the Future's state will change to cancelled (so that cancelled() will return True) and the callbacks will be scheduled. For regular Futures, cancellation will always succeed immediately; but for Tasks (see below) the task may ignore or delay the cancellation attempt.
  • cancelled(). Returns True if the Future was successfully cancelled.
  • done(). Returns True if the Future is done. Note that a cancelled Future is considered done too (here and everywhere).
  • result(). Returns the result set with set_result(), or raises the exception set with set_exception(). Raises CancelledError if cancelled. Difference with PEP 3148: This has no timeout argument and does not wait; if the future is not yet done, it raises an exception.
  • exception(). Returns the exception if set with set_exception(), or None if a result was set with set_result(). Raises CancelledError if cancelled. Difference with PEP 3148: This has no timeout argument and does not wait; if the future is not yet done, it raises an exception.
  • add_done_callback(fn). Add a callback to be run when the Future becomes done (or is cancelled). If the Future is already done (or cancelled), schedules the callback to using call_soon(). Difference with PEP 3148: The callback is never called immediately, and always in the context of the caller -- typically this is a thread. You can think of this as calling the callback through call_soon(). Note that in order to match PEP 3148, the callback (unlike all other callbacks defined in this PEP, and ignoring the convention from the section "Callback Style" below) is always called with a single argument, the Future object. (The motivation for strictly serializing callbacks scheduled with call_soon() applies here too.)
  • remove_done_callback(fn). Remove the argument from the list of callbacks. This method is not defined by PEP 3148. The argument must be equal (using ==) to the argument passed to add_done_callback(). Returns the number of times the callback was removed.
  • set_result(result). The Future must not be done (nor cancelled) already. This makes the Future done and schedules the callbacks. Difference with PEP 3148: This is a public API.
  • set_exception(exception). The Future must not be done (nor cancelled) already. This makes the Future done and schedules the callbacks. Difference with PEP 3148: This is a public API.

The internal method set_running_or_notify_cancel() is not supported; there is no way to set the running state. Likewise, the method running() is not supported.

The following exceptions are defined:

  • InvalidStateError. Raised whenever the Future is not in a state acceptable to the method being called (e.g. calling set_result() on a Future that is already done, or calling result() on a Future that is not yet done).
  • InvalidTimeoutError. Raised by result() and exception() when a nonzero timeout argument is given.
  • CancelledError. An alias for concurrent.futures.CancelledError. Raised when result() or exception() is called on a Future that is cancelled.
  • TimeoutError. An alias for concurrent.futures.TimeoutError. May be raised by run_until_complete().

A Future is associated with an event loop when it is created.

A asyncio.Future object is not acceptable to the wait() and as_completed() functions in the concurrent.futures package. However, there are similar APIs asyncio.wait() and asyncio.as_completed(), described below.

A asyncio.Future object is acceptable to a yield from expression when used in a coroutine. This is implemented through the __iter__() interface on the Future. See the section "Coroutines and the Scheduler" below.

When a Future is garbage-collected, if it has an associated exception but neither result() nor exception() has ever been called, the exception is logged. (When a coroutine uses yield from to wait for a Future, that Future's result() method is called once the coroutine is resumed.)

In the future (pun intended) we may unify asyncio.Future and concurrent.futures.Future, e.g. by adding an __iter__() method to the latter that works with yield from. To prevent accidentally blocking the event loop by calling e.g. result() on a Future that's not done yet, the blocking operation may detect that an event loop is active in the current thread and raise an exception instead. However the current PEP strives to have no dependencies beyond Python 3.3, so changes to concurrent.futures.Future are off the table for now.

There are some public functions related to Futures:

  • asyncio.async(arg). This takes an argument that is either a coroutine object or a Future (i.e., anything you can use with yield from) and returns a Future. If the argument is a Future, it is returned unchanged; if it is a coroutine object, it wraps it in a Task (remember that Task is a subclass of Future).
  • asyncio.wrap_future(future). This takes a PEP 3148 Future (i.e., an instance of concurrent.futures.Future) and returns a Future compatible with the event loop (i.e., a asyncio.Future instance).


Transports and protocols are strongly influenced by Twisted and PEP 3153. Users rarely implement or instantiate transports -- rather, event loops offer utility methods to set up transports.

Transports work in conjunction with protocols. Protocols are typically written without knowing or caring about the exact type of transport used, and transports can be used with a wide variety of protocols. For example, an HTTP client protocol implementation may be used with either a plain socket transport or an SSL/TLS transport. The plain socket transport can be used with many different protocols besides HTTP (e.g. SMTP, IMAP, POP, FTP, IRC, SPDY).

The most common type of transport is a bidirectional stream transport. There are also unidirectional stream transports (used for pipes) and datagram transports (used by the create_datagram_endpoint() method).

Methods For All Transports

  • get_extra_info(name, default=None). This is a catch-all method that returns implementation-specific information about a transport. The first argument is the name of the extra field to be retrieved. The optional second argument is a default value to be returned. Consult the implementation documentation to find out the supported extra field names. For an unsupported name, the default is always returned.

Bidirectional Stream Transports

A bidrectional stream transport is an abstraction on top of a socket or something similar (for example, a pair of UNIX pipes or an SSL/TLS connection).

Most connections have an asymmetric nature: the client and server usually have very different roles and behaviors. Hence, the interface between transport and protocol is also asymmetric. From the protocol's point of view, writing data is done by calling the write() method on the transport object; this buffers the data and returns immediately. However, the transport takes a more active role in reading data: whenever some data is read from the socket (or other data source), the transport calls the protocol's data_received() method.

Nevertheless, the interface between transport and protocol used by bidirectional streams is the same for clients as it is for servers, since the connection between a client and a server is essentially a pair of streams, one in each direction.

Bidirectional stream transports have the following public methods:

  • write(data). Write some bytes. The argument must be a bytes object. Returns None. The transport is free to buffer the bytes, but it must eventually cause the bytes to be transferred to the entity at the other end, and it must maintain stream behavior. That is, t.write(b'abc'); t.write(b'def') is equivalent to t.write(b'abcdef'), as well as to:

  • writelines(iterable). Equivalent to:

    for data in iterable:
  • write_eof(). Close the writing end of the connection. Subsequent calls to write() are not allowed. Once all buffered data is transferred, the transport signals to the other end that no more data will be received. Some protocols don't support this operation; in that case, calling write_eof() will raise an exception. (Note: This used to be called half_close(), but unless you already know what it is for, that name doesn't indicate which end is closed.)

  • can_write_eof(). Return True if the protocol supports write_eof(), False if it does not. (This method typically returns a fixed value that depends only on the specific Transport class, not on the state of the Transport object. It is needed because some protocols need to change their behavior when write_eof() is unavailable. For example, in HTTP, to send data whose size is not known ahead of time, the end of the data is typically indicated using write_eof(); however, SSL/TLS does not support this, and an HTTP protocol implementation would have to use the "chunked" transfer encoding in this case. But if the data size is known ahead of time, the best approach in both cases is to use the Content-Length header.)

  • get_write_buffer_size(). Return the current size of the transport's write buffer in bytes. This only knows about the write buffer managed explicitly by the transport; buffering in other layers of the network stack or elsewhere of the network is not reported.

  • set_write_buffer_limits(high=None, low=None). Set the high- and low-water limits for flow control.

    These two values control when to call the protocol's pause_writing() and resume_writing() methods. If specified, the low-water limit must be less than or equal to the high-water limit. Neither value can be negative.

    The defaults are implementation-specific. If only the high-water limit is given, the low-water limit defaults to a implementation-specific value less than or equal to the high-water limit. Setting high to zero forces low to zero as well, and causes pause_writing() to be called whenever the buffer becomes non-empty. Setting low to zero causes resume_writing() to be called only once the buffer is empty. Use of zero for either limit is generally sub-optimal as it reduces opportunities for doing I/O and computation concurrently.

  • pause_reading(). Suspend delivery of data to the protocol until a subsequent resume_reading() call. Between pause_reading() and resume_reading(), the protocol's data_received() method will not be called.

  • resume_reading(). Restart delivery of data to the protocol via data_received(). Note that "paused" is a binary state -- pause_reading() should only be called when the transport is not paused, while resume_reading() should only be called when the transport is paused.

  • close(). Sever the connection with the entity at the other end. Any data buffered by write() will (eventually) be transferred before the connection is actually closed. The protocol's data_received() method will not be called again. Once all buffered data has been flushed, the protocol's connection_lost() method will be called with None as the argument. Note that this method does not wait for all that to happen.

  • abort(). Immediately sever the connection. Any data still buffered by the transport is thrown away. Soon, the protocol's connection_lost() method will be called with None as argument.

Unidirectional Stream Transports

A writing stream transport supports the write(), writelines(), write_eof(), can_write_eof(), close() and abort() methods described for bidrectional stream transports.

A reading stream transport supports the pause_reading(), resume_reading() and close() methods described for bidrectional stream transports.

A writing stream transport calls only connection_made() and connection_lost() on its associated protocol.

A reading stream transport can call all protocol methods specified in the Protocols section below (i.e., the previous two plus data_received() and eof_received()).

Datagram Transports

Datagram transports have these methods:

  • sendto(data, addr=None). Sends a datagram (a bytes object). The optional second argument is the destination address. If omitted, remote_addr must have been specified in the create_datagram_endpoint() call that created this transport. If present, and remote_addr was specified, they must match. The (data, addr) pair may be sent immediately or buffered. The return value is None.
  • abort(). Immediately close the transport. Buffered data will be discarded.
  • close(). Close the transport. Buffered data will be transmitted asynchronously.

Datagram transports call the following methods on the associated protocol object: connection_made(), connection_lost(), error_received() and datagram_received(). ("Connection" in these method names is a slight misnomer, but the concepts still exist: connection_made() means the transport representing the endpoint has been created, and connection_lost() means the transport is closed.)

Subprocess Transports

Subprocess transports have the following methods:

  • get_pid(). Return the process ID of the subprocess.
  • get_returncode(). Return the process return code, if the process has exited; otherwise None.
  • get_pipe_transport(fd). Return the pipe transport (a unidirectional stream transport) corresponding to the argument, which should be 0, 1 or 2 representing stdin, stdout or stderr (of the subprocess). If there is no such pipe transport, return None. For stdin, this is a writing transport; for stdout and stderr this is a reading transport. You must use this method to get a transport you can use to write to the subprocess's stdin.
  • send_signal(signal). Send a signal to the subprocess.
  • terminate(). Terminate the subprocess.
  • kill(). Kill the subprocess. On Windows this is an alias for terminate().
  • close(). This is an alias for terminate().

Note that send_signal(), terminate() and kill() wrap the corresponding methods in the standard library subprocess module.


Protocols are always used in conjunction with transports. While a few common protocols are provided (e.g. decent though not necessarily excellent HTTP client and server implementations), most protocols will be implemented by user code or third-party libraries.

Like for transports, we distinguish between stream protocols, datagram protocols, and perhaps other custom protocols. The most common type of protocol is a bidirectional stream protocol. (There are no unidirectional protocols.)

Stream Protocols

A (bidirectional) stream protocol must implement the following methods, which will be called by the transport. Think of these as callbacks that are always called by the event loop in the right context. (See the "Context" section way above.)

  • connection_made(transport). Indicates that the transport is ready and connected to the entity at the other end. The protocol should probably save the transport reference as an instance variable (so it can call its write() and other methods later), and may write an initial greeting or request at this point.

  • data_received(data). The transport has read some bytes from the connection. The argument is always a non-empty bytes object. There are no guarantees about the minimum or maximum size of the data passed along this way. p.data_received(b'abcdef') should be treated exactly equivalent to:

  • eof_received(). This is called when the other end called write_eof() (or something equivalent). If this returns a false value (including None), the transport will close itself. If it returns a true value, closing the transport is up to the protocol. However, for SSL/TLS connections this is ignored, because the TLS standard requires that no more data is sent and the connection is closed as soon as a "closure alert" is received.

    The default implementation returns None.

  • pause_writing(). Asks that the protocol temporarily stop writing data to the transport. Heeding the request is optional, but the transport's buffer may grow without bounds if you keep writing. The buffer size at which this is called can be controlled through the transport's set_write_buffer_limits() method.

  • resume_writing(). Tells the protocol that it is safe to start writing data to the transport again. Note that this may be called directly by the transport's write() method (as opposed to being called indirectly using call_soon()), so that the protocol may be aware of its paused state immediately after write() returns.

  • connection_lost(exc). The transport has been closed or aborted, has detected that the other end has closed the connection cleanly, or has encountered an unexpected error. In the first three cases the argument is None; for an unexpected error, the argument is the exception that caused the transport to give up.

Here is a table indicating the order and multiplicity of the basic calls:

  1. connection_made() -- exactly once
  2. data_received() -- zero or more times
  3. eof_received() -- at most once
  4. connection_lost() -- exactly once

Calls to pause_writing() and resume_writing() occur in pairs and only between #1 and #4. These pairs will not be nested. The final resume_writing() call may be omitted; i.e. a paused connection may be lost and never be resumed.

Datagram Protocols

Datagram protocols have connection_made() and connection_lost() methods with the same signatures as stream protocols. (As explained in the section about datagram transports, we prefer the slightly odd nomenclature over defining different method names to indicating the opening and closing of the socket.)

In addition, they have the following methods:

  • datagram_received(data, addr). Indicates that a datagram data (a bytes objects) was received from remote address addr (an IPv4 2-tuple or an IPv6 4-tuple).
  • error_received(exc). Indicates that a send or receive operation raised an OSError exception. Since datagram errors may be transient, it is up to the protocol to call the transport's close() method if it wants to close the endpoint.

Here is a chart indicating the order and multiplicity of calls:

  1. connection_made() -- exactly once
  2. datagram_received(), error_received() -- zero or more times
  3. connection_lost() -- exactly once

Subprocess Protocol

Subprocess protocols have connection_made(), connection_lost(), pause_writing() and resume_writing() methods with the same signatures as stream protocols. In addition, they have the following methods:

  • pipe_data_received(fd, data). Called when the subprocess writes data to its stdout or stderr. fd is the file descriptor (1 for stdout, 2 for stderr). data is a bytes object. (TBD: No pipe_eof_received()?)
  • pipe_connection_lost(fd, exc). Called when the subprocess closes its stdin, stdout or stderr. fd is the file descriptor. exc is an exception or None.
  • process_exited(). Called when the subprocess has exited. To retrieve the exit status, use the transport's get_returncode() method.

Note that depending on the behavior of the subprocess it is possible that process_exited() is called either before or after pipe_connection_lost(). For example, if the subprocess creates a sub-subprocess that shares its stdin/stdout/stderr and then itself exits, process_exited() may be called while all the pipes are still open. On the other hand when the subprocess closes its stdin/stdout/stderr but does not exit, pipe_connection_lost() may be called for all three pipes without process_exited() being called. If (as is the more common case) the subprocess exits and thereby implicitly closes all pipes, the calling order is undefined.

Callback Style

Most interfaces taking a callback also take positional arguments. For instance, to arrange for foo("abc", 42) to be called soon, you call loop.call_soon(foo, "abc", 42). To schedule the call foo(), use loop.call_soon(foo). This convention greatly reduces the number of small lambdas required in typical callback programming.

This convention specifically does not support keyword arguments. Keyword arguments are used to pass optional extra information about the callback. This allows graceful evolution of the API without having to worry about whether a keyword might be significant to a callee somewhere. If you have a callback that must be called with a keyword argument, you can use a lambda. For example:

loop.call_soon(lambda: foo('abc', repeat=42))

Coroutines and the Scheduler

This is a separate toplevel section because its status is different from the event loop interface. Usage of coroutines is optional, and it is perfectly fine to write code using callbacks only. On the other hand, there is only one implementation of the scheduler/coroutine API, and if you're using coroutines, that's the one you're using.


A coroutine is a generator that follows certain conventions. For documentation purposes, all coroutines should be decorated with @asyncio.coroutine, but this cannot be strictly enforced.

Coroutines use the yield from syntax introduced in PEP 380, instead of the original yield syntax.

The word "coroutine", like the word "generator", is used for two different (though related) concepts:

  • The function that defines a coroutine (a function definition decorated with asyncio.coroutine). If disambiguation is needed we will call this a coroutine function.
  • The object obtained by calling a coroutine function. This object represents a computation or an I/O operation (usually a combination) that will complete eventually. If disambiguation is needed we will call it a coroutine object.

Things a coroutine can do:

  • result = yield from future -- suspends the coroutine until the future is done, then returns the future's result, or raises an exception, which will be propagated. (If the future is cancelled, it will raise a CancelledError exception.) Note that tasks are futures, and everything said about futures also applies to tasks.
  • result = yield from coroutine -- wait for another coroutine to produce a result (or raise an exception, which will be propagated). The coroutine expression must be a call to another coroutine.
  • return expression -- produce a result to the coroutine that is waiting for this one using yield from.
  • raise exception -- raise an exception in the coroutine that is waiting for this one using yield from.

Calling a coroutine does not start its code running -- it is just a generator, and the coroutine object returned by the call is really a generator object, which doesn't do anything until you iterate over it. In the case of a coroutine object, there are two basic ways to start it running: call yield from coroutine from another coroutine (assuming the other coroutine is already running!), or convert it to a Task (see below).

Coroutines (and tasks) can only run when the event loop is running.

Waiting for Multiple Coroutines

To wait for multiple coroutines or Futures, two APIs similar to the wait() and as_completed() APIs in the concurrent.futures package are provided:

  • asyncio.wait(fs, timeout=None, return_when=ALL_COMPLETED). This is a coroutine that waits for the Futures or coroutines given by fs to complete. Coroutine arguments will be wrapped in Tasks (see below). This returns a Future whose result on success is a tuple of two sets of Futures, (done, pending), where done is the set of original Futures (or wrapped coroutines) that are done (or cancelled), and pending is the rest, i.e. those that are still not done (nor cancelled). Note that with the defaults for timeout and return_when, done will always be an empty list. Optional arguments timeout and return_when have the same meaning and defaults as for concurrent.futures.wait(): timeout, if not None, specifies a timeout for the overall operation; return_when, specifies when to stop. The constants FIRST_COMPLETED, FIRST_EXCEPTION, ALL_COMPLETED are defined with the same values and the same meanings as in PEP 3148:

    • ALL_COMPLETED (default): Wait until all Futures are done (or until the timeout occurs).
    • FIRST_COMPLETED: Wait until at least one Future is done (or until the timeout occurs).
    • FIRST_EXCEPTION: Wait until at least one Future is done but not cancelled with an exception set. (The exclusion of cancelled Futures from the condition is surprising, but PEP 3148 does it this way.)
  • asyncio.as_completed(fs, timeout=None). Returns an iterator whose values are Futures or coroutines; waiting for successive values waits until the next Future or coroutine from the set fs completes, and returns its result (or raises its exception). The optional argument timeout has the same meaning and default as it does for concurrent.futures.wait(): when the timeout occurs, the next Future returned by the iterator will raise TimeoutError when waited for. Example of use:

    for f in as_completed(fs):
        result = yield from f  # May raise an exception.
        # Use result.

    Note: if you do not wait for the values produced by the iterator, your for loop may not make progress (since you are not allowing other tasks to run).

  • asyncio.wait_for(f, timeout). This is a convenience to wait for a single coroutine or Future with a timeout. When a timeout occurs, it cancels the task and raises TimeoutError. To avoid the task cancellation, wrap it in shield().

  • asyncio.gather(f1, f2, ...). Returns a Future which waits until all arguments (Futures or coroutines) are done and return a list of their corresponding results. If one or more of the arguments is cancelled or raises an exception, the returned Future is cancelled or has its exception set (matching what happened to the first argument), and the remaining arguments are left running in the background. Cancelling the returned Future does not affect the arguments. Note that coroutine arguments are converted to Futures using asyncio.async().

  • asyncio.shield(f). Wait for a Future, shielding it from cancellation. This returns a Future whose result or exception is exactly the same as the argument; however, if the returned Future is cancelled, the argument Future is unaffected.

    A use case for this function would be a coroutine that caches a query result for a coroutine that handles a request in an HTTP server. When the request is cancelled by the client, we could (arguably) want the query-caching coroutine to continue to run, so that when the client reconnects, the query result is (hopefully) cached. This could be written e.g. as follows:

    def handle_request(self, request):
        cached_query = self.get_cache(...)
        if cached_query is None:
            cached_query = yield from asyncio.shield(self.fill_cache(...))


The coroutine asyncio.sleep(delay) returns after a given time delay.


A Task is an object that manages an independently running coroutine. The Task interface is the same as the Future interface, and in fact Task is a subclass of Future. The task becomes done when its coroutine returns or raises an exception; if it returns a result, that becomes the task's result, if it raises an exception, that becomes the task's exception.

Cancelling a task that's not done yet throws an asyncio.CancelledError exception into the coroutine. If the coroutine doesn't catch this (or if it re-raises it) the task will be marked as cancelled (i.e., cancelled() will return True); but if the coroutine somehow catches and ignores the exception it may continue to execute (and cancelled() will return False).

Tasks are also useful for interoperating between coroutines and callback-based frameworks like Twisted. After converting a coroutine into a Task, callbacks can be added to the Task.

To convert a coroutine into a task, call the coroutine function and pass the resulting coroutine object to the asyncio.Task() constructor. You may also use asyncio.async() for this purpose.

You may ask, why not automatically convert all coroutines to Tasks? The @asyncio.coroutine decorator could do this. However, this would slow things down considerably in the case where one coroutine calls another (and so on), as switching to a "bare" coroutine has much less overhead than switching to a Task.

The Scheduler

The scheduler has no public interface. You interact with it by using yield from future and yield from task. In fact, there is no single object representing the scheduler -- its behavior is implemented by the Task and Future classes using only the public interface of the event loop, so it will work with third-party event loop implementations, too.

Convenience Utilities

A few functions and classes are provided to simplify the writing of basic stream-based clients and servers, such as FTP or HTTP. Thes are:

  • asyncio.open_connection(host, port): A wrapper for EventLoop.create_connection() that does not require you to provide a Protocol factory or class. This is a coroutine that returns a (reader, writer) pair, where reader is an instance of StreamReader and writer is an instance of StreamWriter (both described below).

  • asyncio.start_server(client_connected_cb, host, port): A wrapper for EventLoop.create_server() that takes a simple callback function rather than a Protocol factory or class. This is a coroutine that returns a Server object just as create_server() does. Each time a client connection is accepted, client_connected_cb(reader, writer) is called, where reader is an instance of StreamReader and writer is an instance of StreamWriter (both described below). If the result returned by client_connected_cb() is a coroutine, it is automatically wrapped in a Task.

  • StreamReader: A class offering an interface not unlike that of a read-only binary stream, except that the various reading methods are coroutines. It is normally driven by a StreamReaderProtocol instance. Note that there should be only one reader. The interface for the reader is:

    • readline(): A coroutine that reads a string of bytes representing a line of text ending in '\n', or until the end of the stream, whichever comes first.
    • read(n): A coroutine that reads up to n bytes. If n is omitted or negative, it reads until the end of the stream.
    • readexactly(n): A coroutine that reads exactly n bytes, or until the end of the stream, whichever comes first.
    • exception(): Return the exception that has been set on the stream using set_exception(), or None if no exception is set.

    The interface for the driver is:

    • feed_data(data): Append data (a bytes object) to the internal buffer. This unblocks a blocked reading coroutine if it provides sufficient data to fulfill the reader's contract.
    • feed_eof(): Signal the end of the buffer. This unblocks a blocked reading coroutine. No more data should be fed to the reader after this call.
    • set_exception(exc): Set an exception on the stream. All subsequent reading methods will raise this exception. No more data should be fed to the reader after this call.
  • StreamWriter: A class offering an interface not unlike that of a write-only binary stream. It wraps a transport. The interface is an extended subset of the transport interface: the following methods behave the same as the corresponding transport methods: write(), writelines(), write_eof(), can_write_eof(), get_extra_info(), close(). Note that the writing methods are _not_ coroutines (this is the same as for transports, but different from the StreamReader class). The following method is in addition to the transport interface:

    • drain(): This should be called with yield from after writing significant data, for the purpose of flow control. The intended use is like this:

      yield from writer.drain()

      Note that this is not technically a coroutine: it returns either a Future or an empty tuple (both can be passed to yield from). Use of this method is optional. However, when it is not used, the internal buffer of the transport underlying the StreamWriter may fill up with all data that was ever written to the writer. If an app does not have a strict limit on how much data it writes, it _should_ call yield from drain() occasionally to avoid filling up the transport buffer.

  • StreamReaderProtocol: A protocol implementation used as an adapter between the bidirectional stream transport/protocol interface and the StreamReader and StreamWriter classes. It acts as a driver for a specific StreamReader instance, calling its methods feed_data(), feed_eof(), and set_exception() in response to various protocol callbacks. It also controls the behavior of the drain() method of the StreamWriter instance.


Locks, events, conditions and semaphores modeled after those in the threading module are implemented and can be accessed by importing the asyncio.locks submodule. Queus modeled after those in the queue module are implemented and can be accessed by importing the asyncio.queues submodule.

In general these have a close correspondence to their threaded counterparts, however, blocking methods (e.g. acquire() on locks, put() and get() on queues) are coroutines, and timeout parameters are not provided (you can use asyncio.wait_for() to add a timeout to a blocking call, however).

The docstrings in the modules provide more complete documentation.


The following classes are provided by asyncio.locks. For all these except Event, the with statement may be used in combination with yield from to acquire the lock and ensure that the lock is released regardless of how the with block is left, as follows:

with (yield from my_lock):
  • Lock: a basic mutex, with methods acquire() (a coroutine), locked(), and release().
  • Event: an event variable, with methods wait() (a coroutine), set(), clear(), and is_set().
  • Condition: a condition variable, with methods acquire(), wait(), wait_for(predicate) (all three coroutines), locked(), release(), notify(), and notify_all().
  • Semaphore: a semaphore, with methods acquire() (a coroutine), locked(), and release(). The constructor argument is the initial value (default 1).
  • BoundedSemaphore: a bounded semaphore; this is similar to Semaphore but the initial value is also the maximum value.


The following classes and exceptions are provided by asyncio.queues.

  • Queue: a standard queue, with methods get(), put() (both coroutines), get_nowait(), put_nowait(), empty(), full(), qsize(), and maxsize().
  • PriorityQueue: a subclass of Queue that retrieves entries in priority order (lowest first).
  • LifoQueue: a subclass of Queue that retrieves the most recently added entries first.
  • JoinableQueue: a subclass of Queue with task_done() and join() methods (the latter a coroutine).
  • Empty, Full: exceptions raised when get_nowait() or put_nowait() is called on a queue that is empty or full, respectively.



All logging performed by the asyncio package uses a single logging.Logger object, asyncio.logger. To customize logging you can use the standard Logger API on this object. (Do not replace the object though.)

SIGCHLD handling on UNIX

Efficient implementation of the process_exited() method on subprocess protocols requires a SIGCHLD signal handler. However, signal handlers can only be set on the event loop associated with the main thread. In order to support spawning subprocesses from event loops running in other threads, a mechanism exists to allow sharing a SIGCHLD handler between multiple event loops. There are two additional functions, asyncio.get_child_watcher() and asyncio.set_child_watcher(), and corresponding methods on the event loop policy.

There are two child watcher implementation classes, FastChildWatcher and SafeChildWatcher. Both use SIGCHLD. The SafeChildWatcher class is used by default; it is inefficient when many subprocesses exist simultaneously. The FastChildWatcher class is efficient, but it may interfere with other code (either C code or Python code) that spawns subprocesses without using an asyncio event loop. If you are sure you are not using other code that spawns subprocesses, to use the fast implementation, run the following in your main thread:

watcher = asyncio.FastChildWatcher()

Wish List

(There is agreement that these features are desirable, but no implementation was available when Python 3.4 beta 1 was released, and the feature freeze for the rest of the Python 3.4 release cycle prohibits adding them in this late stage. However, they will hopefully be added in Python 3.5, and perhaps earlier in the PyPI distribution.)

  • Support a "start TLS" operation to upgrade a TCP socket to SSL/TLS.

Former wish list items that have since been implemented (but aren't specified by the PEP):

  • UNIX domain sockets.
  • A per-loop error handling callback.

Open Issues

(Note that these have been resolved de facto in favor of the status quo by the acceptance of the PEP. However, the PEP's provisional status allows revising these decisions for Python 3.5.)

  • Why do create_connection() and create_datagram_endpoint() have a proto argument but not create_server()? And why are the family, flag, proto arguments for getaddrinfo() sometimes zero and sometimes named constants (whose value is also zero)?
  • Do we need another inquiry method to tell whether the loop is in the process of stopping?
  • A fuller public API for Handle? What's the use case?
  • A debugging API? E.g. something that logs a lot of stuff, or logs unusual conditions (like queues filling up faster than they drain) or even callbacks taking too much time...
  • Do we need introspection APIs? E.g. asking for the read callback given a file descriptor. Or when the next scheduled call is. Or the list of file descriptors registered with callbacks. Right now these all require using internals.
  • Do we need more socket I/O methods, e.g. sock_sendto() and sock_recvfrom(), and perhaps others like pipe_read()? I guess users can write their own (it's not rocket science).
  • We may need APIs to control various timeouts. E.g. we may want to limit the time spent in DNS resolution, connecting, ssl/tls handshake, idle connection, close/shutdown, even per session. Possibly it's sufficient to add timeout keyword arguments to some methods, and other timeouts can probably be implemented by clever use of call_later() and Task.cancel(). But it's possible that some operations need default timeouts, and we may want to change the default for a specific operation globally (i.e., per event loop).



Apart from PEP 3153, influences include PEP 380 and Greg Ewing's tutorial for yield from, Twisted, Tornado, ZeroMQ, pyftpdlib, and wattle (Steve Dower's counter-proposal). My previous work on asynchronous support in the NDB library for Google App Engine provided an important starting point.

I am grateful for the numerous discussions on python-ideas from September through December 2012, and many more on python-tulip since then; a Skype session with Steve Dower and Dino Viehland; email exchanges with and a visit by Ben Darnell; an audience with Niels Provos (original author of libevent); and in-person meetings (as well as frequent email exchanges) with several Twisted developers, including Glyph, Brian Warner, David Reid, and Duncan McGreggor.

Contributors to the implementation include Eli Bendersky, Gustavo Carneiro (Gambit Research), Saúl Ibarra Corretgé, Geert Jansen, A. Jesse Jiryu Davis, Nikolay Kim, Charles-François Natali, Richard Oudkerk, Antoine Pitrou, Giampaolo Rodolá, Andrew Svetlov, and many others who submitted bugs and/or fixes.

I thank Antoine Pitrou for his feedback in his role of official PEP BDFL.