Armin Ronacher: I don't understand Python's Asyncio

Event Loops

The event loop in asyncio is a bit different than you would expect from first look. On the surface it looks like each thread has one event loop but that's not really how it works. Here is how I think this works:

• if you are the main thread an event loop is created when you call asyncio.get_event_loop()
• if you are any other thread, a runtime error is raised from asyncio.get_event_loop()
• You can at any point asyncio.set_event_loop() to bind an event loop with the current thread. Such an event loop can be created with the asyncio.new_event_loop() function.
• Event loops can be used without being bound to the current thread.
• asyncio.get_event_loop() returns the thread bound event loop, it does not return the currently running event loop.

The combination of these behaviors is super confusing for a few reasons. First of all you need to know that these functions are delegates to the underlying event loop policy which is globally set. The default is to bind the event loop to the thread. Alternatively one could in theory bind the event loop to a greenlet or something similar if one would so desire. However it's important to know that library code does not control the policy and as such cannot reason that asyncio will scope to a thread.

Secondly asyncio does not require event loops to be bound to the context through the policy. An event loop can work just fine in isolation. However this is the first problem for library code as a coroutine or something similar does not know which event loop is responsible for scheduling it. This means that if you call asyncio.get_event_loop() from within a coroutine you might not get the event loop back that ran you. This is also the reason why all APIs take an optional explicit loop parameter. So for instance to figure out which coroutine is currently running one cannot invoke something like this:

defget_task():loop=asyncio.get_event_loop()try:returnasyncio.Task.get_current(loop)exceptRuntimeError:returnNone

Instead the loop has to be passed explicitly. This furthermore requires you to pass through the loop explicitly everywhere in library code or very strange things will happen. Not sure what the thinking for that design is but if this is not being fixed (that for instance get_event_loop() returns the actually running loop) then the only other change that makes sense is to explicitly disallow explicit loop passing and require it to be bound to the current context (thread etc.).

Since the event loop policy does not provide an identifier for the current context it also is impossible for a library to "key" to the current context in any way. There are also no callbacks that would permit to hook the tearing down of such a context which further limits what can be done realistically.

Awaitables and Coroutines

In my humble opinion the biggest design mistake of Python was to overload iterators so much. They are now being used not just for iteration but also for various types of coroutines. One of the biggest design mistakes of iterators in Python is that StopIteration bubbles if not caught. This can cause very frustrating problems where an exception somewhere can cause a generator or coroutine elsewhere to abort. This is a long running issue that Jinja for instance has to fight with. The template engine internally renders into a generator and when a template for some reason raises a StopIteration the rendering just ends there.

Python is slowly learning the lesson of overloading this system more. First of all in 3.something the asyncio module landed and did not have language support. So it was decorators and generators all the way down. To implemented the yield from support and more, the StopIteration was overloaded once more. This lead to surprising behavior like this:

>>> deffoo(n):... ifnin(0,1):... return[1]... foriteminrange(n):... yielditem*2...>>> list(foo(0))[]>>> list(foo(1))[]>>> list(foo(2))[0, 2]

No error, no warning. Just not the behavior you expect. This is because a return with a value from a function that is a generator actually raises a StopIteration with a single arg that is not picked up by the iterator protocol but just handled in the coroutine code.

With 3.5 and 3.6 a lot changed because now in addition to generators we have coroutine objects. Instead of making a coroutine by wrapping a generator there is no a separate object which creates a coroutine directly. It's implemented by prefixing a function with async. For instance async def x() will make such a coroutine. Now in 3.6 there will be separate async generators that will raise AsyncStopIteration to keep it apart. Additionally with Python 3.5 and later there is now a future import (generator_stop) that will raise a RuntimeError if code raises StopIteration in an iteration step.

Why am I mentioning all this? Because the old stuff does not really go away. Generators still have send and throw and coroutines still largely behave like generators. That is a lot of stuff you need to know now for quite some time going forward.

To unify a lot of this duplication we have a few more concepts in Python now:

• awaitable: an object with an __await__ method. This is for instance implemented by native coroutines and old style coroutines and some others.
• coroutinefunction: a function that returns a native coroutine. Not to be confused with a function returning a coroutine.
• a coroutine: a native coroutine. Note that old asyncio coroutines are not considered coroutines by the current documentation as far as I can tell. At the very least inspect.iscoroutine does not consider that a coroutine. It's however picked up by the future/awaitable branches.

In particularly confusing is that asyncio.iscoroutinefunction and inspect.iscoroutinefunction are doing different things. Same with inspect.iscoroutine and inspect.iscoroutinefunction. Note that even though inspect does not know anything about asycnio legacy coroutine functions in the type check, it is apparently aware of them when you check for awaitable status even though it does not conform to __await__.

Coroutine Wrappers

Whenever you run async def Python invokes a thread local coroutine wrapper. It's set with sys.set_coroutine_wrapper and it's a function that can wrap this. Looks a bit like this:

>>> import sys
>>> sys.set_coroutine_wrapper(lambda x: 42)
>>> async def foo():
...  pass
...
>>> foo()
__main__:1: RuntimeWarning: coroutine 'foo' was never awaited
42

In this case I never actually invoke the original function and just give you a hint of what this can do. As far as I can tell this is always thread local so if you swap out the event loop policy you need to figure out separately how to make this coroutine wrapper sync up with the same context if that's something you want to do. New threads spawned will not inherit that flag from the parent thread.

This is not to be confused with the asyncio coroutine wrapping code.

Awaitables and Futures

Some things are awaitables. As far as I can see the following things are considered awaitable:

• native coroutines
• generators that have the fake CO_ITERABLE_COROUTINE flag set (we will cover that)
• objects with an __await__ method

Essentially these are all objects with an __await__ method except that the generators don't for legacy reasons. Where does the CO_ITERABLE_COROUTINE flag come from? It comes from a coroutine wrapper (now to be confused with sys.set_coroutine_wrapper) that is @asyncio.coroutine. That through some indirection will wrap the generator with types.coroutine (to to be confused with types.CoroutineType or asyncio.coroutine) which will re-create the internal code object with the additional flag CO_ITERABLE_COROUTINE.

So now that we know what those things are, what are futures? First we need to clear up one thing: there are actually two (completely incompatible) types of futures in Python 3. asyncio.futures.Future and concurrent.futures.Future. One came before the other but they are also also both still used even within asyncio. For instance asyncio.run_coroutine_threadsafe() will dispatch a coroutine to a event loop running in another thread but it will then return a concurrent.futures.Future object instead of a asyncio.futures.Future object. This makes sense because only the concurrent.futures.Future object is thread safe.

So now that we know there are two incompatible futures we should clarify what futures are in asyncio. Honestly I'm not entirely sure where the differences are but I'm going to call this "eventual" for the moment. It's an object that eventually will hold a value and you can do some handling with that eventual result while it's still computing. Some variations of this are called deferreds, others are called promises. What the exact difference is is above my head.

What can you do with a future? You can attach a callback that will be invoked once it's ready or you can attach a callback that will be invoked if the future fails. Additionally you can await it (it implements __await__ and is thus awaitable). Additionally futures can be cancelled.

So how do you get such a future? By calling asyncio.ensure_future on an awaitable object. This will also make a good old generator into such a future. However if you read the docs you will read that asyncio.ensure_future actually returns a Task. So what's a task?

Tasks

A task is a future that is wrapping a coroutine in particular. It works like a future but it also has some extra methods to extract the current stack of the contained coroutine. We already saw the tasks mentioned earlier because it's the main way to figure out what an event loop is currently doing via Task.get_current.

There is also a difference in how cancellation works for tasks and futures but that's beyond the scope of this. Cancellation is its own entire beast. If you are in a coroutine and you know you are currently running you can get your own task through Task.get_current as mentioned but this requires knowledge of what event loop you are dispatched on which might or might not be the thread bound one.

It's not possible for a coroutine to know which loop goes with it. Also the Task does not provide that information through a public API. However if you did manage to get hold of a task you can currently access task._loop to find back to the event loop.

Handles

In addition to all of this there are handles. Handles are opaque objects of pending executions that cannot be awaited but they can be cancelled. In particular if you schedule the execution of a call with call_soon or call_soon_threadsafe (and some others) you get that handle you can then use to cancel the execution as a best effort attempt but you can't wait for the call to actually take place.

Executors

Since you can have multiple event loops but it's not obvious what the use of more than one of those things per thread is the obvious assumption can be made that a common setup is to have N threads with an event loop each. So how do you inform another event loop about doing some work? You cannot schedule a callback into an event loop in another thread and get the result back. For that you need to use executors instead.

Executors come from concurrent.futures for instance and they allow you to schedule work into threads that itself is not evented. For instance if you use run_in_executor on the event loop to schedule a function to be called in another thread. The result is then returned as an asyncio coroutine instead of a concurrent coroutine like run_coroutine_threadsafe would do. I did not yet have enough mental capacity to figure out why those APIs exist, how you are supposed to use and when which one. The documentation suggests that the executor stuff could be used to build multiprocess things.

Transports and Protocols

I always though those would be the confusing things but that's basically a verbatim copy of the same concepts in Twisted. So read those docs if you want to understand them.