References I wish I'd already found

# Python Asynchronous Programming Fundamentals

Posted: 2022-05-29 2022-05-30 John T. Wodder II Programming Python, async [link]

Python introduced asynchronous programming capabilities in version 3.4 in 2014, with further notable improvements in almost every minor version since. However, to many Python programmers, this area of the language remains esoteric, misunderstood, and underutilized. This article aims to elucidate the fundamental concepts of asynchronous programming as part of the first step towards mastery.

There won’t be many code samples in this article, but reading it should make it easier to grok the asyncio documentation and figure out how to piece things together.

## High-Level Overview

Asynchronous programming provides a way to interleave execution of multiple functions (coroutines) at once; at any given point, only one of the functions is actually doing operations, while the others are waiting for things like blocking I/O to complete (or, if the I/O has already completed, they’re waiting for the current coroutine to be suspended so that they get a chance to resume).

Note

In the Python standard library and all third-party async libraries that I am aware of, the kinds of low-level operations that one can ultimately suspend execution waiting for are almost exclusively I/O-related. If you’re seeking to add concurrency to CPU-bound code (e.g., anything involving number crunching), you are likely to be better off using multiprocessing instead.

Specifically, whenever execution of a currently-running coroutine reaches an await expression, the coroutine may be suspended, and another previously-suspended coroutine may resume execution if what it was suspended on has since returned a value. Suspension can also happen when an async for block requests the next value from an asynchronous iterator or when an async with block is entered or exited, as these operations use await under the hood.

Note that, although multiple coroutines can be processed at once, effectively finishing in any order, operations within a single coroutine continue to be executed in the order they’re written. For example, given the following code:

async def f():
...

async def g():
...

async def amain():
await f()
await g()

asyncio.run(amain())


When await f() is reached, amain() will be suspended until f() is finished, and only after that will execution proceed to the next line, starting coroutine g(). If you want to execute f() and g() concurrently, you need to use asyncio.create_task(), asyncio.gather(), or similar. See this article for more details.

In general, coroutines can only be called or scheduled by other coroutines. To run a “top-level” coroutine from inside synchronous code (i.e., either inside a non-coroutine function or at module level), the simplest & preferred way is to use the asyncio.run() function introduced in Python 3.7. Code meant to run on older Pythons must use the lower-level loop.run_until_complete() instead, along with other loop methods for cleanup.

## Definitions

Asynchronous programming features the execution of multiple tasks concurrently, with one task being run while waiting for others to complete. Asynchronous programming in Python is an example of cooperative multitasking [wiki], as it requires the running coroutines to cooperate and yield control on their own; if the current coroutine goes for a long time without calling await, execution will remain on that coroutine all the while, and all other coroutines in the thread will remain in a suspended state. This is contrast to the preemptive multitasking [wiki] of multithreaded and multiprocess programs, where the Python interpreter or OS scheduler decides on its own when to switch between running contexts, with the points at which switches can occur being difficult or impossible to predict in general.

A coroutine function is a function defined with async def instead of just def. (In the context of asynchronous programming, a function defined with just def is called a synchronous function.) Only coroutine functions can contain await, async for, and async with constructs, and, as of Python 3.10, they cannot contain yield from constructs.

• Note that just calling a coroutine function does not cause it to start running; you need to either schedule it for concurrent execution with asyncio.create_task() or similar or else await on it directly.

A coroutine object is the result of calling a coroutine function. This is not the value returned (or raised) by the function; rather, it is a pending computation that can be suspended & resumed at any point that the coroutine function uses await, async for, or async with. The actual return value or exception is obtained either by awaiting on the coroutine object (possibly wrapped in a task and/or something like asyncio.gather()) from within another coroutine function or by running it as a “top-level” entry point using asyncio.run() within synchronous code.

• Asynchronous generators — coroutine functions that use yield — are a bit of an exception. You do not await the result of the function; instead, you iterate through it using either async for ... in ...: or await anext(...).

• Confusingly, both coroutine functions and coroutine objects can be referred to as just “coroutines.”

The actual scheduling of coroutine execution is managed by an event loop. An event loop is created by asyncio.run() or asyncio.new_event_loop(), handed one or more coroutines and/or synchronous callbacks, and then set off to run either forever or until completion of a “top-level” coroutine. It’s the event loop’s job to execute the current coroutine until it suspends on an await, after which it looks to see if any suspended coroutines are now done with their suspension and either picks one to resume or, if none are ready, waits until one is.

An awaitable is any value that the await keyword can be applied to; this includes coroutine objects, futures, future-likes, and tasks (see below). Awaiting on an awaitable causes the current coroutine to be suspended until the awaitable is ready to provide a value or raise an exception.

A future (class asyncio.Future) is a low-level container for the result of a computation (a value or exception) that starts out empty and is assigned a value or exception later. Awaiting on a future will suspend the current coroutine until something else either stores a result in the future or cancels it.

• You may already be familiar with futures in the form of the Future class from the concurrent.futures module, which provides access to the results of operations evaluated in other threads or processes. The asyncio.Future class is similar in spirit, but has a different API.

A future-like is an object with an __await__() method, which must return an iterator. Awaiting on a future-like causes the current coroutine to be suspended until the iterator is exhausted, at which point the value attribute of the terminating StopIteration exception is returned as the result of the await expression.

A task (class asyncio.Task) represents a running coroutine. Creating a task from a coroutine object with asyncio.create_task() will cause the coroutine to be scheduled for execution concurrently with other running coroutines. The task instance can later be awaited on to suspend the current coroutine until the wrapped coroutine finishes executing, returning its result.

A running task can be cancelled by calling its Task.cancel() method. This will cause the underlying coroutine to receive an asyncio.CancelledError exception the next time it awaits, likely putting an end to the task’s execution.

## Syntax

Asynchronous programming in Python takes place inside coroutines, functions defined using async def instead of just def. Within a coroutine, the await keyword can be applied to any awaitable expression (such as a call to another coroutine) to suspend execution of the coroutine until the awaitable has a value or exception ready, at which point the coroutine is resumed and the await expression returns that value or raises that exception.

Here’s a basic example you’ve seen in all the tutorials:

import asyncio

async def waiter():
print("Before sleep")
await asyncio.sleep(5)
print("After sleep")

asyncio.run(waiter())


This code prints a message, waits for a five-second sleep to elapse, and then prints another message. As written, it’s rather pointless; we’re only running one coroutine at once, so there’s no advantage over using a synchronous function with time.sleep(). Here’s something slightly more involved:

import asyncio

async def operate(time, result):
print(f"Spending {time} seconds doing operations ...")
await asyncio.sleep(time)
print(f"Operations done after {time} seconds!")
return result

async def amain():
x, y = await asyncio.gather(operate(5, 42), operate(2, 23))
print(f"Got {x=}, {y=}")
assert x == 42
assert y == 23

asyncio.run(amain())


This code mocks spending time on two blocking operations in parallel. If you run the script (Python 3.8+ required) and time it, you’ll see that it only takes about 5 seconds in total, and the 2-second task completes three seconds before the 5-second one. After both tasks are done, the final “Got x=42, y=23” message is printed.

Besides await, there are two other syntactical constructs specific to coroutines: async for ... in ...: (for iterating over asynchronous iterables) and async with ...: (for entering & exiting asynchronous context managers). These work the same way as their non-async counterparts, except that the iterables and context managers in question need to support asynchronous usage; for example, an async for cannot iterate over a list, and an async with cannot operate on an ordinary filehandle. Similarly, a regular for cannot be applied to an asynchronous iterator, and a regular with cannot be applied to, say, asyncio.Lock.

Speaking of asynchronous iteration, this works pretty much how you’d expect: by using yield inside a coroutine, it becomes an asynchronous generator that can be iterated over with async for or await anext(...). Note that, in contrast to non-generator coroutines, you do not apply await to an asynchronous generator. For example, given this function:

async def aiterator():
for i in range(5):
await asyncio.sleep(i)
yield i


you use it like this:

async for x in aiterator():
print(x)


No await anywhere in the async for loop.

Note that there is no way to get a value out of a coroutine without awaiting on it (either directly or via something like asyncio.gather()); if a coroutine is never awaited and never converted into a task, asyncio will complain when it is garbage-collected. Moreover, await (and async for and async with) cannot be used outside of a coroutine; in order to start the awaiting on a “top-level” coroutine, you need to use asyncio.run().

Note that the body of a coroutine isn’t required to contain any awaits or similar, though if it doesn’t, there often isn’t much point in making it a coroutine in the first place. An exception is the __aenter__() special method of asynchronous context managers; usually, the body will just be return self, but it’s still required to define the method with async def.

### Type Annotations

Adding type annotations to asynchronous code works the same way as for synchronous code. If an asynchronous func() take an integer x and returns a string, you write its annotated signature as async def func(x: int) -> str. However, if you pass around an unawaited coroutine object (not always the best idea), you annotate it as Awaitable[T], where T is the return type of the coroutine.

Async callables do not have their own type; they are instead annotated as Callable[..., Awaitable[T]], where T is the return type of the coroutine function.

Asynchronous iterators, iterables, and context managers, though, do get their own types: AsyncIterator, AsyncIterable, and typing.AsyncContextManager/contextlib.AbstractAsyncContextManager.

### Special Methods

__aiter__() and __anext__()

These methods are used to implement asynchronous iterator & iterable classes as an alternative to writing asynchronous generator functions, analogously to how a class can be defined with __iter__() and __next__() methods to implement a synchronous iterator as an alternative to writing a generator function.

__aiter__() must be a synchronous function that returns an object with __aiter__() and __anext__() methods. __anext__() must be a coroutine that either returns a value or raises a StopAsyncIteration exception.

__aenter__() and __aexit__()

These methods are used to implement asynchronous context managers. They are defined the same way as the __enter__() and __exit__() methods of synchronous context managers, except that the asynchronous versions must be coroutines.

__await__()

This method is used to create a future-like class that can be awaited directly. There is generally very little need to implement this, but we include it here for completeness.

__await__() must be a synchronous function that returns an iterator. This iterator will be advanced each time the event loop checks to see if the future-like is ready to return a value. If the future-like is not ready, the iterator must yield a special value (see below). If the future-like is ready, it must either return a result or raise an exception; this result or exception will then be the result of the await expression acting on the future-like.

The values that the iterator yields depend on what async implementation the code is running under (See “Alternative Async Implementations” below). When using the Python standard library’s asyncio, you generally want to yield None. When using trio, you need to yield an instance of one of several private classes internal to trio. When using curio, you need to yield a ("trap_name", *trap_args) tuple instructing the kernel to invoke a special “trap” function; yielding ("trap_sleep", 0) instructs the kernel to do nothing special.

For example, if you want to implement your own Future class for use with asyncio, you might start out writing it like so:

class MyFuture:
def __init__(self):
self.value = None
self.is_set = False

def set_value(self, value):
self.value = valuse
self.is_set = True

def __await__(self):
while not self.is_set:
yield
return self.value


### Historical Note: Generator-Based Coroutines

When asyncio was first introduced in Python 3.4, the async and await keywords were not yet present. Instead, coroutine functions were created by applying the @asyncio.coroutine decorator to a normal generator function, and awaiting was done using yield from. There were no asynchronous iterators or asynchronous context managers in 3.4, either. Even after async and await were introduced in Python 3.5, the older generator-based coroutines could not use them.

This style of writing coroutines was deprecated in Python 3.8 and removed entirely in Python 3.11.

## Running More than One Coroutine at Once

Now for the part you’ve been waiting for, and the part that makes asynchronous programming worth it: actually running multiple functions concurrently.

The simplest way to start running another coroutine concurrently is to pass a coroutine object to asyncio.create_task(); this schedules the coroutine for execution, but it won’t actually start running until the next time the current coroutine calls await (and maybe not even then). asyncio.create_task() returns an asyncio.Task object that can be used to query the state of the coroutine or cancel it; awaiting on the task object will cause the current coroutine to be suspended until the task is complete, at which point the return value of the task’s underlying coroutine is returned.

If you create multiple tasks and then await on them one by one, a given await will not return a result until the task in question is done; if task B finishes running while you’re still awaiting on task A, the coroutine doing the awaiting will continue to be suspended until A is done, and when it then later awaits on B, it will get back B’s return value immediately, because B is already done. For example, the following code:

import asyncio
from time import strftime

def hms():
return strftime("%H:%M:%S")

async def operate(time, result):
print(f"{hms()}: Spending {time} seconds doing operations ...")
await asyncio.sleep(time)
print(f"{hms()}: Operations done after {time} seconds!")
return result

async def amain():

asyncio.run(amain())


outputs something like the following:

17:12:56: Spending 5 seconds doing operations ...
17:12:56: Spending 2 seconds doing operations ...
17:12:58: Operations done after 2 seconds!
17:13:01: Operations done after 5 seconds!


If you need to await on multiple coroutines but don’t care about the exact order in which they finish, you can use asyncio.gather(), asyncio.as_completed(), or asyncio.wait() to await on them together; this article gives a good overview and explanation of the differences between the functions.

You don’t have to await on a task if you don’t need its return value or don’t need to be assured that it ever finishes, but if such a “background task” raises an uncaught exception, asyncio will complain. One way to address this is to attach a synchronous callback function to the task with Task.add_done_callback() that retrieves any uncaught exceptions with Task.exception(), like so:

import asyncio
from time import strftime

def hms():
return strftime("%H:%M:%S")

print(f"{hms()}: In the background")
raise RuntimeError("Ouch")

try:
else:
except asyncio.CancelledError:

print(f"{hms()}: Now we sleep and let bg_task do its thing")
await asyncio.sleep(2)
print(f"{hms()}: I'm awake!")



The output from the above will look like:

17:16:33: Now we sleep and let bg_task do its thing
17:16:33: In the background
17:16:35: I'm awake!


## Exception Handling

Whenever an exception occurs inside a coroutine, it propagates upwards to whatever’s awaiting on it; if unhandled, it will propagate all the way out through the asyncio.run() call, at which point all still-running tasks are cancelled. If there is no chain of awaits leading to the “top-level” coroutine (say, because you did asyncio.create_task() and then didn’t await on the result, letting it run in the background), asyncio will end up complaining when the coroutine is eventually garbage-collected. See the passage above on using Task.add_done_callback() to handle such errors.

For the specific case of KeyboardInterrupt, the exception is raised in whatever coroutine the main thread is currently running at the time.

## Example Code

This gist provides an example of using asynchronous programming in Python to download assets for one or more releases of a GitHub repository in parallel. Try it out with an invocation like:

python download-assets.py --download-dir jq stedolan/jq jq-1.5 jq-1.6


The script requires Python 3.8 or higher and the ghrepo and httpx packages on PyPI to run.

Asynchronous programming does not use threads; by default, all coroutines and their operations are run in whichever thread called asyncio.run(). The exception is when asyncio.to_thread() or loop.run_in_executor() is used to run a synchronous function in a separate thread (or, with the latter function, even a separate process), returning an object that can be awaited on to receive the function’s result.

If multiple threads call asyncio.run() separately, each thread will get its own event loop and collection of coroutines.

Note that each thread in a Python process has at most one event loop at a time, and an event loop can only belong to one thread. An important consequence of this is that, if you have a synchronous function foo() that calls asyncio.run() on some coroutine, then foo() cannot be called by another coroutine, because that would lead to two event loops in the same thread, which doesn’t work.

• In asynchronous programming, the coroutine being executed can only change when the current coroutine uses await or similar. This allows the programmer to be assured that, between awaits in the same coroutine, operations will not be interfered with and data will not be modified by other coroutines.

When using threads, on the other hand, the running thread [1] can change at almost any point as chosen by the interpreter, which necessitates careful programming and copious use of locks in order to ensure that variables are not modified by one thread “behind the back of” another thread that’s also using them.

• If you’ve done serious work with threads, you’ve likely encountered the fact that you cannot “kill” a thread in the middle of its execution unless the “killable” thread is deliberately programmed to allow for this by, say, regularly checking some flag and exiting if it’s true. Asynchronous programming, on the other hand, makes it possible to cancel a running coroutine via the asyncio.Task.cancel() method; once a coroutine is cancelled, the next time the event loop checks on it while it’s suspended on an await or similar, the coroutine will be resumed, but instead of receiving the value it was awaiting for, an asyncio.CancelledError will be raised at the await expression, likely putting an end to the coroutine’s execution.

 [1] Recall that, due to Python’s global interpreter lock (GIL), regardless of how many threads a Python program uses or how many cores your machine has, only one thread will be executing Python bytecode at any moment.

## Async Version History

Here follows a list of the notable developments & changes in asynchronous programming across Python versions.

Python 3.4

The asyncio module is added, implementing PEP 3156. This enables the creation of coroutines via the @asyncio.coroutine decorator; within a coroutine, awaiting is performed with yield from. (Support for creating coroutines in this way would later be deprecated in Python 3.8 and removed in Python 3.11.)

Most of the functionality added in this version is now categorized as the “low level” part of asyncio.

Python 3.5

PEP 492 implemented:

• It is now possible to define coroutines via async def and await with await.

• In Python 3.5, yield cannot be used inside the body of an async def coroutine function.

• Asynchronous iteration with async for is now possible.

• Asynchronous context managers are now supported via async with.

• __await__(), __aiter__(), __anext__(), __aenter__(), and __aexit__() special methods added

• Originally, __aiter__() methods were expected to be coroutines (or anything else returning an awaitable) resolving to asynchronous iterators. This was changed in 3.5.2 to have __aiter__() instead return an asynchronous iterator directly. Returning an awaitable from __aiter__() produces a PendingDeprecationWarning starting in 3.5.2, a DeprecationWarning starting in 3.6, and a RuntimeError starting in 3.7.

Python 3.6

• yield can now be used in the body of an async def coroutine function, thereby enabling asynchronous generators (PEP 525). (yield from remains prohibited, though.)

• async for can now be used in list, set, & dict comprehensions and in generator expressions

• await expressions can now be used in any comprehension

• Using async or await as an identifier now generates a DeprecationWarning

Python 3.7

• async and await are now reserved keywords

Python 3.8

• Running python -m asyncio now starts an async REPL

• @asyncio.coroutine() is now deprecated

• Passing a loop parameter is now deprecated for most of asyncio’s high-level API

• asyncio.CancelledError now inherits directly from BaseException instead of Exception

Python 3.9

Python 3.10

• aiter() and anext() functions added

• The loop parameter (deprecated in Python 3.8) is now removed from most of asyncio’s high-level API

Python 3.11

• @asyncio.coroutine (deprecated in Python 3.8) is now removed

## Alternative Async Implementations

While all the code we’ve shown so far uses the Python standard library’s asyncio module, it’s not required to use this to work with coroutines. Alternative async library implementations exist that define their own event loops and primitive operations. The more notable implementations include:

• trio seeks to enable structured concurrency [wiki] in asynchronous code. In trio, a collection of tasks are run concurrently by grouping them together under a nursery (also known as a task group); if one of the tasks in a nursery raises an error, all the other tasks in the same nursery are automatically cancelled.

• curio is an asyncio alternative featuring a more streamlined API and intended to be easier to reason about

In general, different async implementations are incompatible, and features from different implementations cannot be used in the same code unless you make careful use of whatever compatibility facilities they may provide.

In fact, just being able to use the same code unmodified regardless of whether using implementation A or implementation B is tricky, as all implementations use different primitives. Libraries to help you with that include:

• sniffio can be used to detect which async library is in use

• anyio provides a common API (based on trio) that can be used to run the same code under both asyncio and trio (and previously curio, until it & anyio parted ways in anyio 3.0)