Asyncio

What is it

asyncio is a library included in Python 3.5 that supports a programming model where sometimes, operations that would normally block the thread until some other event happened (like getting a response from a network connection) instead allow other code to run on that thread while waiting.

asyncio takes a very, very explicit approach to asynchronous programming: only code written in methods flagged as async can call any code in an asynchronous way.

Which creates a chicken/egg problem: your async methods can only be called by other async methods, so how do you call the first one?

The answer: you don’t. What you have to do instead is turn over control of the thread to an event loop, after arranging for the loop to (sooner or later) invoke your async code.

Then once you start the loop running, it can invoke the async code.

What good is it

Note first that you can use threads to accomplish the same things as asyncio in most cases, with better performance. So what good is asyncio?

For one thing, it leads to more straightforward code than managing multiple threads, protecting data structures from concurrent access, etc. There’s only one thread and no preemptive multitasking.

If you want to play with async programming in Python, asyncio looks easier to work with and understand than Twisted, but that’s not a very practical reason.

More significantly, threads won’t scale as well if you need to wait for many, many things at the same time - asyncio might be somewhat slower, but might be the only way that some tasks can be run at all. Each thread can take 50K of memory, while a coroutine might take only 3K.

Event loops

Async code can only run inside an event loop. The event loop is the driver code that manages the cooperative multitasking.

(I think) a typical pattern would be to get or create an event loop, set up some things to be run by it, then start the event loop running and have it run until the program is finished.

If it’s useful for some reason, you can create multiple threads and run different event loops in each of them. For example, Django uses the main thread to wait for incoming requests, so we can’t run an asyncio event loop there, but we can start a separate worker thread for our event loop.

Coroutines

coroutines

  • Python distinguishes between a coroutine function and a coroutine object

  • Write a coroutine function by putting async in front of the def.

  • Only a coroutine function can use await, non-coroutine functions cannot.

  • Calling a coroutine function does not execute it, but rather returns a coroutine object. (This is analogous to generator functions - calling them doesn’t execute the function, it returns a generator object, which we then use later.)

  • To execute a coroutine object, either:

Example with await:

async def coro_function():
    return 2 + 2

coro = coro_function()
# not executed yet; coro is a coroutine, not 4

print(await coro)
# prints "4"

Example of scheduling it:

async def coro_function(hostname):
    conn = await .... connect async to hostname somehow...

coro = coro_function("example.com")
asyncio.ensure_future(coro)

Of course, usually you wouldn’t split it onto two lines with a temp variable:

asyncio.ensure_future(coro_function("example.com"))

or:

asyncio.get_event_loop().create_task(coro_function("example.com"))

Futures

A future is an object that represents something uncompleted. It makes it easy for code in one place to indicate when the work is done, and optionally what the result was, and for code elsewhere that was interested in it to find out about it.

In other words, you can use future objects to manage synchronization more explicitly.

Create one on the fly by calling loop.create_future():

future = loop.create_future()

Arrange for something to be called when the future becomes done:

future.add_done_callback(fn)

You can add lots of callbacks. They’ll all be called (one at a time).

The callback receives the future object as an argument. Use functools.partial as usual if you want to pass other arguments.

When the future is done, mark it done and set its result:

future.set_result(value)

The callbacks can call future.result() to find out what the result was if they care.

Tasks

A Task is a way to arrange for a coroutine to be executed by an event loop, while also providing the caller a way to find out what the result was.

A task is automatically scheduled for execution when it is created.

There are two ways to do this, which seem equivalent as far as I can tell:

future = loop.create_task(coroutine)
future = asyncio.ensure_future(coroutine[, loop=loop])

Now you can add callbacks if you want:

future.add_done_callback(fn1)

Also, if the loop isn’t already running and you just want to run the loop for this one thing, you can now:

loop.run_until_complete(future)

Awaitables

Coroutine objects and future objects are called awaitables - either can be used with await.

Note: You can only invoke an awaitable once; after that, it’s completed, done, it runs no more.

Event loops

Creating/getting one

  • To get the current thread’s default event loop object, call asyncio.get_event_loop()
  • get_event_loop will not create an event loop object unless you’re on the main thread, and otherwise will raise an exception if the current thread doesn’t have a default loop set.
  • To create a new event loop: new_event_loop()
  • To make a loop the default loop for the current thread: set_event_loop(loop)

So, to use an event loop in the main thread, you can just do:

loop = asyncio.get_event_loop()
# use loop....

But to run an event loop in another thread, you would do something like:

loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
# use loop...

You don’t have to set your loop as the thread’s default, though, if you’re willing to pass your loop object to all the APIs that otherwise use the default loop. But that’s a pain.

Running a loop

If you want a long-running loop that keeps responding to events until it’s told to stop, use loop.run_forever().

If you want to compute some finite work using coroutines and then stop, use loop.run_until_complete(<future or coroutine>).

Stopping a loop

Use loop.stop().

Getting a loop to call a synchronous callable

By a synchronous callable, I mean a callable that is not an awaitable as described above.

This is more like Javascript’s callback-style async programming than in the spirit of Python’s coroutines, but sometimes you need it.

To call the callable as soon as possible, use loop.call_soon(callback). If you want to pass args to the callable, use functools.partial:

loop.call_soon(functools.partial(callable, arg1, arg2))

To delay for N seconds before calling it, use loop.call_later(delay, callable).

To schedule a callback from a different thread, the AbstractEventLoop.call_soon_threadsafe() method should be used. Example:

loop.call_soon_threadsafe(callback, *args)

Getting a loop to call an awaitable

Use asyncio.ensure_future(awaitable, *, loop=None).

Or loop.run_until_complete, but as noted above, that just runs the loop as long as it takes to complete the awaitable.

If you’re doing this from another thread, then you need to use a different method, asyncio.run_coroutine_threadsafe(coro, loop):

future = asyncio.run_coroutine_threadsafe(coroutine, loop)

Running blocking code in another thread

If you need to call some blocking code from a coroutine, and don’t want to block the whole thread, you can make it run in another thread using coroutine AbstractEventLoop.run_in_executor(executor, func, *args):

fn = functools.partial(method, *args)
result = await loop.run_in_executor(None, fn)

Sleep

Calling asyncio.sleep(seconds) does not sleep; it returns a coroutine object. When you execute it by invoking it with await etc, it will complete after <seconds> seconds. So, mostly you’d do:

await asyncio.sleep(10)  # pause 10 seconds