23 KiB

Raw Blame History

giambio - Asynchronous Python made easy (and friendly)

Giambio is an event-driven concurrency library meant* to perform efficient and high-performant I/O multiplexing. This library implements what is known as a stackless mode of execution, or "green threads", though the latter term is misleading as no multithreading is involved (at least not by default).

*: The library works (sometimes), but its still in its very early stages and is nowhere close being production ready, so be aware that it is likely (if not guaranteed) that you'll find bugs and race conditions

Disclaimer

Right now this is nothing more than a toy implementation to help me understand how this whole async/await thing works and it is pretty much guaranteed to explode spectacularly badly while using it. If you find any bugs, please report them!

This project was hugely inspired by the curio and the trio projects, you might want to have a look at their amazing work if you need a rock-solid and structured concurrency framework (I personally recommend trio and that's definitely not related to the fact that most of the content of this document is ~~stolen~~ inspired from its documentation)

What the hell is async anyway?

Libraries like giambio shine the most when it comes to performing asynchronous I/O (reading from a socket, writing to a file, that sort of thing). The most common example of this is a network server that needs to handle multiple connections at the same time. One possible approach to achieve concurrency is to use threads, and despite their bad reputation in Python, they actually might be a good choice when it comes to I/O for reasons that span far beyond the scope of this document. If you choose to use threads, there are a couple things you can do, involving what is known as thread synchronization primitives and thread pools, but once again that is beyond the purposes of this quickstart guide. A library like giambio comes into play when you need to perform lots of blocking operations, and network servers happen to be heavily based on I/O: a blocking operation. Starting to see where we're heading?

A deeper dive

Giambio has been designed with simplicity in mind, so this document won't explain all the gritty details about how async is implemented in Python (you might want to check out this article if you want to learn more about all the implementation details). For the sake of this tutorial, all you need to know is that giambio is all about a feature added in Python 3.5: asynchronous functions, or 'async' for short.

Async functions are functions defined with async def instead of the regular def, like so:

async def async_fun():   # An async function
    print("Hello, world!")

def sync_fun():     # A regular (sync) function
    print("Hello, world!")

First of all, async functions like to stick together: to call an async function you need to put await in front of it, like below:


async def async_two():
    print("Hello from async_two!")
   
async def async_one():
    print("Hello from async_one!")
    await async_two()  # This is an async call

It has to be noted that using await outside of an async function is a SyntaxError, so basically async functions have a unique superpower: they, and no-one else, can call other async functions.

This already presents a chicken-and-egg problem, because when you fire up Python, it is running plain ol' synchronous code; So how do we enter the async context in the first place?

That is done via a special synchronous function, giambio.run in our case, that has the ability to call asynchronous functions and can therefore initiate the async context. For this reason, giambio.run must be called from a synchronous context, to avoid a horrible deadlock.

Now that you know all of this, you might be wondering why on earth would one use async functions instead of regular functions: after all, their ability to call other async functions seems pretty pointless in itself, doesn't it? Take a look at this example below:

import giambio

async def foo():
    print("Hello, world!")

giambio.run(foo)   # Prints 'Hello, world!'

This could as well be written the following way and would produce the same output:

def foo():
    print("Hello, world!")

foo()   # Prints 'Hello, world!'

To answer this question, we have to dig a bit deeper about what giambio gives you in exchange for all this async/await madness.

We already introduced giambio.run, a special runner function that can start the async context from a synchronous one, but giambio provides also a set of tools, mainly for doing I/O. These functions, as you might have guessed, are async functions and they're useful! So if you wanna take advantage of giambio, and hopefully you will after reading this guide, you need to write async code. As an example, take this function using giambio.sleep (giambio.sleep is like time.sleep, but with an async flavor):

Note: If you have decent knowledge about asynchronous python, you might have noticed that we haven't mentioned coroutines so far. Don't worry, that is intentional: giambio never lets a user deal with coroutines on the surface because the whole async model is much simpler if we take coroutines out of the game, and everything works just the same.

import giambio

async def sleep_double(n):
    await giambio.sleep(2 * n)

giambio.run(sleep_double, 2)  # This hangs for 4 seconds and returns

As it turns out, this function is one that's actually worth making async: because it calls another async function. Not that there's nothing wrong with our foo from before, it surely works, but it doesn't really make sense to make it async in the first place.

Don't forget the `await`!

As we already learned, async functions can only be called with the await keyword, and it would be logical to think that forgetting to do so would raise an error, but it's actually a little bit trickier than that.

Take this example here:

import giambio

async def sleep_double_broken(n):
    print("Taking a nap!")
    start = giambio.clock()
    giambio.sleep(2 * n)    # We forgot the await!
    end = giambio.clock() - start
    print(f"Slept for {end:.2f} seconds!")

giambio.run(sleep_double_broken, 2)

Running this code, will produce an output that looks like this:

Taking a nap!
Slept 0.00 seconds!
__main__:7: RuntimeWarning: coroutine 'sleep' was never awaited

Wait, what happened here? From this output, it looks like the code worked, but something clearly went wrong: the function didn't sleep. Python gives us a hint that we broke something by raising a warning, complaining that coroutine 'sleep' was never awaited (you might not see this warning because it depends on whether a garbage collection cycle occurred or not). I know I said we weren't going to talk about coroutines, but you have to blame Python, not me. Just know that if you see a warning like that, it means that somewhere in your code you forgot an await when calling an async function, so try fixing that before trying to figure out what could be the problem if you have a long traceback: most likely that's just collateral damage caused by the missing keyword.

If you're ok with just remembering to put await every time you call an async function, you can safely skip to the next section, but for the curios among y'all I might as well explain exactly what happened there.

When async functions are called without the await, they don't exactly do nothing: they return this weird 'coroutine' object


>>> giambio.sleep(1)
<coroutine object sleep at 0x1069520d0>

The reason for this is that while giambio tries to separate the async and sync worlds, therefore considering await giambio.sleep(1) as a single unit, when you await an async function Python does 2 things:

It creates this weird coroutine object
Passes that object to await, which runs the function

This is due to the fact that people started writing asynchronous Python code before the async/await syntax was added so many libraries (like asyncio), had to figure out some clever hacks to make it work without native support from the language itself, taking advantage of generator functions (we'll talk about those later on), and coroutines are heavily based on generators.

Something actually useful

Ok, so far you've learned that asynchronous functions can call other async functions, and that giambio has a special runner function that can start the whole async context, but we didn't really do anything useful. Our previous examples could be written using sync functions (like time.sleep) and they would work just fine, that isn't quite useful is it?

But here's the plot twist: giambio can run multiple async functions at the same time. Yep, you read that right.

To demonstrate this, have a look a this example

import giambio

async def child():
    print("[child] Child spawned! Sleeping for 2 seconds")
    await giambio.sleep(2)
    print("[child] Had a nice nap!")

async def child1():
    print("[child 1] Child spawned! Sleeping for 2 seconds")
    await giambio.sleep(2)
    print("[child 1] Had a nice nap!")


async def main():
    start = giambio.clock()
    async with giambio.create_pool() as pool:
        pool.spawn(child)
        pool.spawn(child1)
        print("[main] Children spawned, awaiting completion")
    print(f"[main] Children execution complete in {giambio.clock() - start:.2f} seconds")


if __name__ == "__main__":
    giambio.run(main)

There is a lot going on here, and we'll explain every bit of it step by step:

First, we imported giambio and defined two async functions: child and child1
These two functions will just print something and then sleep for 2 seconds
Here comes the real fun: async with? What's going on there? As it turns out, Python 3.5 didn't just add async functions, but also quite a bit of related new syntax. One of the things that was added is asynchronous context managers. You might have already encountered context managers in python, but in case you didn't, a line such as with foo as sth tells the Python interpreter to call foo.__enter__() at the beginning of the block, and foo.__exit__() at the end of the block. The as keyword just assigns the return value of foo.__enter__() to the variable sth. So context managers are a shorthand for calling functions, and since Python 3.5 added async functions, we also needed async context managers. While with foo as sth calls foo.__enter__(), async with foo as sth calls await foo.__aenter__(): easy huh?

Note: On a related note, Python 3.5 also added asynchronous for loops! The logic is the same though: while for item in container calls container.__next__() to fetch the next item, async for item in container calls await container.__anext__() to do so. It's that simple, mostly just remember to stick await everywhere and you'll be good.

Ok, so now we grasp async with, but what's with that create_pool()? In giambio, there are actually 2 ways to call async functions: one we've already seen (await fn()), while the other is trough an asynchronous pool. The cool part about pool.spawn() is that it will return immediately, without waiting for the async function to finish. So, now our functions are running in the background. After we spawn our tasks, we hit the call to print and the end of the block, so Python awaits the pool's __aexit__() method. What this does is pause the parent task (our main async function in this case) until all children tasks have exited, and as it turns out, that is a good thing. The reason why pools always wait for all children to have finished executing is that it makes easier propagating exceptions in the parent if something goes wrong: unlike many other frameworks, exceptions in giambio always behave as expected*

*: This is a WIP, it doesn't work right now!

Ok, so, let's try running this snippet and see what we get:

[child] Child spawned!! Sleeping for 2 seconds
[child 1] Child spawned!! Sleeping for 2 seconds
[... 2 seconds pass ...]
[child] Had a nice nap!
[child 1] Had a nice nap!
[main] Children execution complete in 2.01 seconds

(Your output might have some lines swapped compared to this)

You see how child and child1 both start and finish together? Moreover, even though each function slept for about 2 seconds (therefore 4 seconds total), the program just took 2 seconds to complete, so our children are really running at the same time.

If you've ever done thread programming, this will feel like home, and that's good: it's exactly what we want. But beware! No threads are involved here, giambio is running in a single thread. That's why we talked about tasks rather than threads so far. The difference between the two is that you can run a lot of tasks in a single thread, and that with threads Python can switch which thread is running at any time. Giambio, on the other hand, can switch tasks only at certain fixed points called checkpoints, more on that later.

A sneak peak into the async world

The basic idea behind libraries like giambio is that they can run a lot of tasks at the same time by switching back and forth between them at appropriate places. An example for that could be a web server: while the server is waiting for a response from a client, we can accept another connection. You don't necessarily need all these pesky details to use giambio, but it's good to have at least an high-level understanding of how this all works.

The peculiarity of asynchronous functions is that they can suspend their execution: that's what await does, it yields back the execution control to giambio, which can then decide what to do next.

To understand this better, take a look at this code:

def countdown(n: int) -> int:
    while n:
        yield n
        n -= 1

for x in countdown(5):
    print(x)

In the above snippet, countdown is a generator function. Generators are really useful because they allow to customize iteration. Running that code produces the following output:

The trick for this to work is yield. What yield does is return back to the caller and suspend itself: In our case, yield returns to the for loop, which calls countdown again. So, the generator resumes right after the yield, decrements n, and loops right back to the top for the while loop to execute again. It's that suspension part that allows the async magic to happen: the whole async/await logic overlaps a lot with generator functions.

Some libraries, like asyncio, take advantage of this yielding mechanism, because they were made way before Python 3.5 added that nice new syntax.

So, since only async functions can suspend themselves, the only places where giambio will switch tasks is where there is a call to await something(). If there is no await, then you can be sure that giambio will not switch tasks (because it can't): this makes the asynchronous model much easier to reason about, because you can know if a function will ever switch, and where will it do so, just by looking at its source code. That is very different from what threads do: they can (and will) switch whenever they feel like it.

Remember when we talked about checkpoints? That's what they are: calls to async functions that allow giambio to switch tasks. The problem with checkpoints is that if you don't have enough of them in your code, then giambio will switch less frequently, hurting concurrency. It turns out that a quick and easy fix for that is calling await giambio.sleep(0); This will implicitly let giambio kick in and do its job, and it will reschedule the caller almost immediately, because the sleep time is 0.

A closer look

In the above section we explained the theory behind async functions, but now we'll inspect the magic behind giambio.run() and its event loop to demistify how giambio makes this whole async thing happen. Luckily for us, giambio has some useful tooling that lets us sneak peak inside the machinery of the library to better help us understand what's going on, located at giambio.debug.BaseDebugger. That's an abstract class that we can customize for our purposes and that communicates with the event loop about everything it's doing, so let's code it:

class Debugger(giambio.debug.BaseDebugger):
    """
    A simple debugger for this test
    """

    def on_start(self):
        print("## Started running")

    def on_exit(self):
        print("## Finished running")

    def on_task_schedule(self, task, delay: int):
        print(f">> A task named '{task.name}' was scheduled to run in {delay:.2f} seconds")

    def on_task_spawn(self, task):
        print(f">> A task named '{task.name}' was spawned")
   
    def on_task_exit(self, task):
        print(f"<< Task '{task.name}' exited")
   
    def before_task_step(self, task):
        print(f"-> About to run a step for '{task.name}'")

    def after_task_step(self, task):
        print(f"<- Ran a step for '{task.name}'")

    def before_sleep(self, task, seconds):
        print(f"# About to put '{task.name}' to sleep for {seconds:.2f} seconds")
   
    def after_sleep(self, task, seconds):
        print(f"# Task '{task.name}' slept for {seconds:.2f} seconds")

    def before_io(self, timeout):
        print(f"!! About to check for I/O for up to {timeout:.2f} seconds")

    def after_io(self, timeout):
        print(f"!! Done I/O check (timeout {timeout:.2f} seconds)")

    def before_cancel(self, task):
        print(f"// About to cancel '{task.name}'")

    def after_cancel(self, task):
        print(f"// Cancelled '{task.name}'")

To use our debugger class, we need to pass it to giambio.run() using the debugger keyword argument, like so:

...
if __name__ == "__main__":
    giambio.run(main, debugger=Debugger())

Note: We passed an instance (see the parentheses?) not a class

Running that modified code will produce a lot of output, and it should look something like this:

## Started running
-> About to run a step for 'main'
>> A task named 'child' was spawned
>> A task named 'child1' was spawned
[main] Children spawned, awaiting completion
<- Ran a step for 'main'
-> About to run a step for 'child'
[child] Child spawned!! Sleeping for 2 seconds
<- Ran a step for 'child'
# About to put 'child' to sleep for 2.00 seconds
-> About to run a step for 'child1'
[child 1] Child spawned!! Sleeping for 2 seconds
<- Ran a step for 'child1'
# About to put 'child1' to sleep for 2.00 seconds
[... 2 seconds pass ...]
# Task 'child' slept for 2.01 seconds
# Task 'child1' slept for 2.01 seconds
!! About to check for I/O for up to 0.00 seconds
!! Done I/O check (timeout 0.00 seconds)
-> About to run a step for 'child'
[child] Had a nice nap!
<< Task 'child' exited
-> About to run a step for 'child1'
[child 1] Had a nice nap!
<< Task 'child1' exited
-> About to run a step for 'main'
<- Ran a step for 'main'
-> About to run a step for 'main'
[main] Children execution complete in 2.01 seconds
<< Task 'main' exited
## Finished running

As expected, this prints a lot of stuff, but let's start going trough it:

First, we start the event loop: That's the call to giambio.run()
```
## Started running
```
After that, we start running the main function
```
-> About to run a step for 'main'
```

When we run main, that enters the async with block and spawns our children, as well as execute our call to print

>> A task named 'child' was spawned
>> A task named 'child1' was spawned
[main] Children spawned, awaiting completion

After that, we hit the end of the block, so we pause and wait for our children to complete: That's when we start switching, and child can now run
```
<- Ran a step for 'main'
-> About to run a step for 'child'
[child] Child spawned!! Sleeping for 2 seconds
```

We're now at await giambio.sleep(2) inside child, and that puts it to sleep

<- Ran a step for 'child'
# About to put 'child' to sleep for 2.00 seconds

Ok, so now child is asleep while main is waiting on its children, and child1 can now execute, so giambio switches again and runs that
```
-> About to run a step for 'child1'
[child 1] Child spawned!! Sleeping for 2 seconds
```
Now we hit the call to await giambio.sleep(2) inside child1, so that also goes to sleep
```
<- Ran a step for 'child1'
# About to put 'child1' to sleep for 2.00 seconds
```
Since there is no other work to do, giambio just waits until it wakes up the two children, 2 seconds later
```
# Task 'child' slept for 2.01 seconds
# Task 'child1' slept for 2.01 seconds
```
Even though we're not doing any I/O here, giambio doesn't know that, so it does some checks (and finds out there is no I/O to do)
```
!! About to check for I/O for up to 0.00 seconds
!! Done I/O check (timeout 0.00 seconds)
```

After 2 seconds have passed giambio wakes up our children and runs them until completion

-> About to run a step for 'child'
[child] Had a nice nap!
<< Task 'child' exited
-> About to run a step for 'child1'
[child 1] Had a nice nap!
<< Task 'child1' exited

As promised, once all children exit, the parent task resumes and runs until it exits. This also causes the entire event loop to exit because there is nothing else to do

-> About to run a step for 'main'
<- Ran a step for 'main'
-> About to run a step for 'main'
[main] Children execution complete in 2.01 seconds
<< Task 'main' exited
## Finished running

So, in our example, our children run until they hit a call to await giambio.sleep, then execution control goes back to giambio.run, which drives the execution for another step. This works because giambio.sleep and giambio.run (as well as many others) work together to make this happen: giambio.sleep can pause the execution of its children task and ask giambio.run to wake him up after a given amount of time.

Note: You may wonder whether you can mix async libraries: for instance, can we call trio.sleep in a giambio application? The answer is no, we can't, and this section explains why. When you call await giambio.sleep, it asks giambio.run to pause the current task, and to do so it talks a language that only giambio.run can understand. Other libraries have other private "languages", so mixing them is not possible: doing so will cause giambio to get very confused and most likely just explode spectacularly badly.

Contributing

This is a relatively young project and it is looking for collaborators! It's not rocket science, but writing a proper framework like this implies some non-trivial issues that require proper and optimized solutions, so if you feel like you want to challenge yourself don't hesitate to contact me on Telegram or by E-mail

23 KiB Raw Blame History