/a-tale-of-event-loops

Notes describing my understanding of how modern Python event loops work

A tale of event loops

I was recently stuck by the amazing insight from Nathaniel J. Smith, a fellow at Berkely, about asyncio. His post Some thoughts on asynchronous API design in a post-async/await world helped me put some words on the (slightly) mixed feelings I have about asyncio after using it a lot recently.

While this is more or less the main focus in his post, he argues that curio has a much simpler implementation than asyncio because it assumes upfront that you have a version of Python that implements PEP 492 (async and await keywords).

I was intrigued and took a look at curio's source and... due to it gaining a lot of features (hopefully briging it closer to a production-ready library), I think it has enough features now that the main essence is diluted. This is definitely not a bad thing, but if, like me, you'd like to understand what Nathaniel was talking about, and learn a few neat Python tricks, reading curio's source code is hardly a good starting point.

Brett Cannon, a Python core developer gives some very good insight on how to use coroutine objects (e.g. if you were implementing an event loop) in his post How the heck does async/await work inPython 3.5?. His focus is on the implementation details and, at least for me, there was still a small piece missing to grok curio.

TL; DR: This is my attempt at trying to capture the core fundamentals on which curio is based. Hopefully, it will be a good stepping stone to get acquainted with curio's source code.

Let's get started :-)

The real nature of coroutines

First, let's take a look at what happens when you call a coroutine.

Hint: you get a "coroutine object", which has a .send() and a .throw() method, just like generator objects do.

>>> async def hello(name):
...     print('Hello, %s!' % (name,))
>>>
>>> coro = hello('world')
>>> type(coro)
<class 'coroutine'>
>>> type(coro.send)
<class 'builtin_function_or_method'>
>>> type(coro.throw)
<class 'builtin_function_or_method'>
>>>
...

Of course you usually don't use these methods because they are hidden behind await coro and asyncio.get_event_loop().run_until_complete(), but since we're trying to look into how that works... :-)

Notice that the code above never printed our "Hello, world!" message. That's because we never actually executed statements inside the coroutine function -- we simply created the coroutine object. In fact, if you actually run this code in an interpreter, you will get a warning stating that our hello coroutine never completed.

If you want to execute the coroutine function, you need to schedule it somehow. In order to do that, we're going to call the .send() method.

>>> async def hello(name):
...     print('Hello, %s!' % (name,))
>>>
>>> task = hello('world')
>>> try:
...     task.send(None)
... except StopIteration:
...     pass
Hello, world!

Just like for a generator object, a couroutine object's .send() method raises StopIteration when the coroutine ends.

We'll see later that we can use .send() to pass information to the coroutine when resuming it with a 2nd, 3rd or Nth call to .send().

If, instead of returning normally, the coroutine fails with an exception, then that exception will be propagated back through .send().

>>> class Hello(Exception):
...    def __init__(self, name):
...        self._name = name
...    def __str__(self):
...        return 'Hello, %s!' % (self._name,)
>>>
>>> async def hello(name):
...     raise Hello(name)
>>>
>>> task = hello('world')
>>> try:
...     task.send(None)
... except Exception as error:
...     # NEW: exception will be propagated here!
...     print(error)
Hello, world!

I also mentioned a .throw() method. Like .send(), it resumes the coroutine, but instead of passing it a value, it raises an exception inside the coroutine at the point where it was suspended.

>>> class Hello(Exception):
...    def __init__(self, name):
...        self._name = name
...    def __str__(self):
...        return 'Hello, %s!' % (self._name,)
>>>
>>> async def hello():
...     pass
>>>
>>> task = hello()
>>> try:
...     # NEW: inject exception.
...     task.throw(Hello('world'))
... except Exception as error:
...     print(error)
Hello, world!

At this point, you should be comfortable with the fact that coroutine objects are very, very similar to generator objects which exist since Python 2.2 (PEP 255) and have had .send() and .throw() methods since Python 2.5 (PEP 342).

A dialog with the event loop

If you look around (or try it), you will notice that coroutine functions, in contast with generator functions, cannot use the yield expression. This raises (not begs) the question: how exactly can coroutine functions yield control back to the code that called .send()?

The answer implies using await on an awaitable object. For an object to be awaitable, it must implement the special __await__() method that returns an iterable. In practice, this is a little awkward, so there is a @types.coroutine decorator in the standard library that allows you to create awaitable objects in a style reminescent of @contextlib.contextmanager.

>>> from types import coroutine
>>>
>>> # NEW: this is an awaitable object!
>>> @coroutine
... def nice():
...     yield
>>>
>>> async def hello(name):
...     # NEW: this makes ``.send()`` return!
...     await nice()
...     print('Hello, %s!' % (name,))
>>>
>>> task = hello('world')
>>> # NEW: call send twice!
>>> task.send(None)
>>> try:
...     task.send(None)
... except StopIteration:
...     pass
Hello, world!

Of course, our nice() object is pretty useless. Don't fret, we'll be doing some more useful things shortly.

Looping

Our previous example calls .send() exactly twice because it nows that hello() will only yield control once. When we don't know (the common case), we need to put this in a loop.

>>> from types import coroutine
>>>
>>> @coroutine
... def nice():
...     yield
>>>
>>> async def hello(name):
...     # NEW: yield many times!
...     for i in range(5):
...         await nice()
...     print('Hello, %s!' % (name,))
>>>
>>> task = hello('world')
>>> try:
...     # NEW: loop!
...     while True:
...         task.send(None)
... except StopIteration:
...     pass
Hello, world!

We're starting to get something that feels like the simplest possible implementation of asyncio.get_event_loop().run_until_complete(), so let's make it more syntactically similar.

>>> from types import coroutine
>>>
>>> @coroutine
... def nice():
...     yield
>>>
>>> async def hello(name):
...     for i in range(5):
...         await nice()
...     print('Hello, %s!' % (name,))
>>>
>>> # NEW: now a reusable function!
>>> def run_until_complete(task):
...     try:
...         while True:
...             task.send(None)
...     except StopIteration:
...         pass
>>>
>>> # NEW: call it as a function!
>>> run_until_complete(hello('world'))
Hello, world!

Spawning child tasks

Now that we've got an event loop that can complete a single task, we'll probably want to start doing useful things. There are many different things we expect to do, but since this is about concurrency, let's start by allowing creation of child tasks.

The main thing we need to do here is to introduce a new spawn() primitive that schedules the new child task. Once the task is scheduled, we want to return control to the parent task so that it can continue on its way.

Note: this example is deliberaly incomplete. We'll see how to join tasks later.

>>> from inspect import iscoroutine
>>> from types import coroutine
>>>
>>> # NEW: awaitable object that sends a request to launch a child task!
>>> @coroutine
... def spawn(task):
...     yield task
>>>
>>> async def hello(name):
...     await nice()
...     print('Hello, %s!' % (name,))
>>>
>>> # NEW: parent task!
>>> async def main():
...      # NEW: create a child task!
...     await spawn(hello('world'))
>>>
>>> def run_until_complete(task):
...     # NEW: schedule the "root" task.
...     tasks = [(task, None)]
...     while tasks:
...         # NEW: round-robin between a set tasks (we may now have more
...         #      than one and we'd like to be as "fair" as possible).
...         queue, tasks = tasks, []
...         for task, data in queue:
...             # NEW: resume the task *once*.
...             try:
...                 data = task.send(data)
...             except StopIteration:
...                 pass
...             except Exception as error:
...                 # NEW: prevent crashed task from ending the loop.
...                 print(error)
...             else:
...                 # NEW: schedule the child task.
...                 if iscoroutine(data):
...                     tasks.append((data, None))
...                 # NEW: reschedule the parent task.
...                 tasks.append((task, None))
>>>
>>> run_until_complete(main())
Hello, world!

Whoa! This is way more complicated than our previous version of run_until_complete(). Where did all of that come from?

Well... now that we can run multiple tasks, we need to worry about things like:

  1. waiting for all the child tasks to complete (recursively), despite errors in any of the tasks
  2. alternating between tasks to let all tasks concurrently progress toward completion

Notice that we now have a nested loop:

  • the outer loop takes care of checking if we're done
  • the inner loop takes care of one scheduler "tick"

There are other ways to do this and it may not be immediately obvious why we're doing it this way, but it's important because of two critical pieces of the event loop that are still missing: timers and I/O. When we add support these later on, we're going to need to schedule internal checks in a "fair" manner too. The outer loop gives us a convenient spot for checking timers and polling the status of I/O operations:

  1. check timers to resume sleeping tasks for which the delay has elapsed;
  2. check I/O operations and schedule tasks for which the pending I/O operations have completed;
  3. perform one scheduler "tick" to resume all tasks we just scheduled.

In short, that's the gist of the coroutine-based scheduler loop.

However, before we get into the more complicated timers & I/O... remember earlier when I mentionned that the example was deliberately incomplete? We know how to spawn new child tasks, but we don't yet know how to wait for them to complete. This a good opportunity to see how we extend the vocabulary of event loop "requests" sent by our awaitable objects.

>>> from collections import defaultdict
>>> from types import coroutine
>>>
>>> @coroutine
... def nice():
...     yield
>>>
>>> @coroutine
... def spawn(task):
...     # NEW: recover the child task handle to pass it back to the parent.
...     child = yield ('spawn', task)
...     return child
>>>
>>> # NEW: awaitable object that sends a request to be notified when a
>>> #      concurrent task completes.
>>> @coroutine
... def join(task):
...     yield ('join', task)
>>>
>>> async def hello(name):
...     await nice()
...     print('Hello, %s!' % (name,))
>>>
>>> async def main():
...     # NEW: recover the child task handle.
...     child = await spawn(hello('world'))
...     # NEW: wait for the child task to complete.
...     await join(child)
...     print('(after join)')
>>>
>>> def run_until_complete(task):
...     tasks = [(task, None)]
...     # NEW: keep track of tasks to resume when a task completes.
...     watch = defaultdict(list)
...     while tasks:
...         queue, tasks = tasks, []
...         for task, data in queue:
...             try:
...                 data = task.send(data)
...             except StopIteration:
...                 # NEW: wait up tasks waiting on this one.
...                 tasks.extend((t, None) for t in watch.pop(task, []))
...             else:
...                 # NEW: dispatch request sent by awaitable object since
...                 #      we now have 3 diffent types of requests.
...                 if data and data[0] == 'spawn':
...                     tasks.append((data[1], None))
...                     tasks.append((task, data[1]))
...                 elif data and data[0] == 'join':
...                     watch[data[1]].append(task)
...                 else:
...                     tasks.append((task, None))
>>>
>>> run_until_complete(main())
Hello, world!
(after join)

For practical reasons, we'll probably want to have some kind Task wrapper for coroutine objects. This comes handy to expose an API for cancellation and to handle some race conditions such as the child task ending before the parent task attempts to join() it (can you spot the bug?)

Passing the child task's return value back as the result of await join() and propagating the exception that crashed the child are left as exercices to the reader.

Sleeping & timers

Now that we have task scheduling under control, we can start tackling some more advanced stuff like timers & I/O. I/O is the ultimate goal, but it pulls in a lot of new stuff, so we'll look at timers first.

If you need to sleep, you can't just call time.sleep() because you'll be blocking all tasks, not just the one you want to suspend.

You have probably spotted the pattern now. We'll be adding two things:

  1. a new type of request
  2. a bit of dispatch code based on the return value of task.send().

We'll also add some bookkeeping to track tasks that while they're suspended. Remember that tasks is a list of coroutines that are scheduled to run in the next tick, but sleeping tasks will likely skip one or more ticks before they are ready to run again.

Keep in mind that sleeping tasks are unlikely to be rescheduled in FIFO order, so we'll need somthing a little more evolved. The most practical way to keep timers (until you allow cancelling them) is to use a priority queue and the standard library's heapq module thankfully makes that super easy.

>>> from heapq import heappop, heappush
>>> from time import sleep as _sleep
>>> from timeit import default_timer
>>> from types import coroutine
>>>
>>> # NEW: we need to keep track of elasped time.
>>> clock = default_timer
>>>
>>> # NEW: request that the event loop reschedule us "later".
>>> @coroutine
... def sleep(seconds):
...     yield ('sleep', seconds)
>>>
>>> # NEW: verify elapsed time matches our request.
>>> async def hello(name):
...     ref = clock()
...     await sleep(3.0)
...     now = clock()
...     assert (now - ref) >= 3.0
...     print('Hello, %s!' % (name,))
>>>
>>> def run_until_complete(task):
...     tasks = [(task, None)]
...
...     # NEW: keep track of tasks that are sleeping.
...     timers = []
...
...     # NEW: watch out, all tasks might be suspended at the same time.
...     while tasks or timers:
...
...         # NEW: if we have nothing to do for now, don't spin.
...         if not tasks:
...             _sleep(max(0.0, timers[0][0] - clock()))
...
...         # NEW: schedule tasks when their timer has elapsed.
...         while timers and timers[0][0] < clock():
...             _, task = heappop(timers)
...             tasks.append((task, None))
...
...         queue, tasks = tasks, []
...         for task, data in queue:
...             try:
...                 data = task.send(data)
...             except StopIteration:
...                 pass
...             else:
...                 # NEW: set a timer and don't reschedule right away.
...                 if data and data[0] == 'sleep':
...                     heappush(timers, (clock() + data[1], task))
...                 else:
...                     tasks.append((task, None))
>>>
>>> run_until_complete(hello('world'))
Hello, world!

Wow, we're really getting the hang of this! Maybe this async stuff is not so hard after all?

Let's see what we can do for I/O!

Handling I/O

Now that we've been through all the other stuff, it's time for the ultimate showdown: I/O.

Scalable I/O normally using native C APIs for multiplexing I/O. Usually, this is the most difficult part of I/O libraries, but thankfully, Python's selectors module makes this quite accessible.

As for all the other operations we've added so far, we'll add some new I/O requests and corresponding request handlers in the event loop. Also, like timers, we'll need to do some internal checks at the start of each scheduler tick.

>>> from selectors import (
...     DefaultSelector,
...     EVENT_READ,
...     EVENT_WRITE,
... )
>>> from socket import socketpair as _socketpair
>>> from types import coroutine
>>>
>>> # NEW: request that the event loop tell us when we can read.
>>> @coroutine
... def recv(stream, size):
...     yield (EVENT_READ, stream)
...     return stream.recv(size)
>>>
>>> # NEW: request that the event loop tell us when we can write.
>>> @coroutine
... def send(stream, data):
...     while data:
...         yield (EVENT_WRITE, stream)
...         size = stream.send(data)
...         data = data[size:]
>>>
>>> # NEW: connect sockets, make sure they never, ever block the loop.
>>> @coroutine
... def socketpair():
...     lhs, rhs = _socketpair()
...     lhs.setblocking(False)
...     rhs.setblocking(False)
...     yield
...     return lhs, rhs
>>>
>>> # NEW: send a message through the socket pair.
>>> async def hello(name):
...     lhs, rhs = await socketpair()
...     await send(lhs, 'Hello, world!'.encode('utf-8'))
...     data = await recv(rhs, 1024)
...     print(data.decode('utf-8'))
...     lhs.close()
...     rhs.close()
>>>
>>> def run_until_complete(task):
...     tasks = [(task, None)]
...
...     # NEW: prepare for I/O multiplexing.
...     selector = DefaultSelector()
...
...     # NEW: watch out, all tasks might be suspended at the same time.
...     while tasks or selector.get_map():
...
...         # NEW: poll I/O operation status and resume tasks when ready.
...         timeout = 0.0 if tasks else None
...         for key, events in selector.select(timeout):
...             tasks.append((key.data, None))
...             selector.unregister(key.fileobj)
...
...         queue, tasks = tasks, []
...         for task, data in queue:
...             try:
...                 data = task.send(data)
...             except StopIteration:
...                 pass
...             else:
...                 # NEW: register for I/O and suspend the task.
...                 if data and data[0] == EVENT_READ:
...                     selector.register(data[1], EVENT_READ, task)
...                 elif data and data[0] == EVENT_WRITE:
...                     selector.register(data[1], EVENT_WRITE, task)
...                 else:
...                     tasks.append((task, None))
>>>
>>> run_until_complete(hello('world'))
Hello, world!

Cancelling tasks

The last, but not least, piece of the puzzle is to cancel tasks. This is where we take advantage of the coroutine object's .throw() method.

Since there is a race condition on cancellation (it could finish concurrently with the cancellation attempt), it requires keeping track of all running tasks to know their "state".

Otherwise, it's a simple extension of task spawning and joining.

Note: this implementation is deliberately imcomplete. It does not properly handle the possibility of cancelling a task that is already scheduled to run in the next tick.

>>> from inspect import iscoroutine
>>> from types import coroutine
>>>
>>> @coroutine
... def spawn(task):
...     task = yield ('spawn', task)
...     return task
>>>
>>> @coroutine
... def join(task):
...     yield ('join', task)
>>>
>>> # NEW: exception to be raised inside tasks when they are cancelled.
>>> class CancelledError(Exception):
...     pass
>>>
>>> # NEW: request that CancelledError be raised inside the task.
>>> @coroutine
... def cancel(task):
...     cancelled = yield ('cancel', task, CancelledError())
...     assert cancelled is True
>>>
>>> # NEW: pause the task without plans to reschedule it (this is simply to
>>> #      guarantee execution order in this demo).
>>> @coroutine
... def suspend():
...     yield ('suspend',)
>>>
>>> async def hello(name):
...     try:
...         await suspend()
...     except CancelledError:
...         print('Hello, %s!' % (name,))
...         raise
>>>
>>> # NEW: spawn a task and then cancel it.
>>> async def main():
...     child = await spawn(hello('world'))
...     await cancel(child)
...     await join(child)
>>>
>>> def run_until_complete(task):
...     tasks = [(task, None)]
...     watch = defaultdict(list)
...
...     # NEW: keep track of all tasks in the tree.
...     tree = {task}
...
...     while tasks:
...         queue, tasks = tasks, []
...         for task, data in queue:
...             try:
...                 # NEW: we may need to pass data or inject an exception.
...                 if isinstance(data, Exception):
...                     data = task.throw(data)
...                 else:
...                     data = task.send(data)
...             except (StopIteration, CancelledError):
...                 tasks.extend((t, None) for t in watch.pop(task, []))
...                 # NEW: update bookkeeping.
...                 tree.discard(task)
...             else:
...                 if data and data[0] == 'spawn':
...                     tasks.append((data[1], None))
...                     tasks.append((task, data[1]))
...                     # NEW: update bookkeeping.
...                     tree.add(data[1])
...                 elif data and data[0] == 'join':
...                     watch[data[1]].append(task)
...                 elif data and data[0] == 'cancel':
...                     # NEW: schedule to raise the exception in the task.
...                     if data[1] in tree:
...                         tasks.append((data[1], data[2]))
...                         tasks.append((task, True))
...                     else:
...                         tasks.append((task, False))
...                 elif data and data[0] == 'suspend':
...                     pass
...                 else:
...                     tasks.append((task, None))
>>>
>>> run_until_complete(main())
Hello, world!

License

This document is copyright Andre Caron <andre.l.caron@gmail.com> and is made available to you under a Creative Commons CC-BY-SA license.