Coroutines in Python

Introduction

Every programmer is acquainted with functions - sequences of instructions grouped together as a single unit in order to perform predetermined tasks. They admit a single entry point, are capable of accepting arguments, may or may not have a return value, and can be called at any moment during a program's execution - including by other functions and themselves.

When a program calls a function its current execution context is saved before passing control over to the function and resuming execution. The function then creates a new context - from there on out newly created data exists exclusively during the functions runtime.

As soon as the task is complete, control is transferred back to the caller - the new context is effectively deleted and replaced by the previous one.

Coroutines

Coroutines are a special type of function that deliberately yield control over to the caller, but does not end its context in the process, instead maintaining it in an idle state.

They benefit from the ability to keep their data throughout their lifetime and, unlike functions, can have several entry points for suspending and resuming execution.

Coroutines in Python work in a very similar way to Generators. Both operate over data, so let's keep the main differences simple:

Generators produce data

Coroutines consume data

The distinct handling of the keyword yield determines whether we are manipulating one or the other.

Defining a Coroutine

With all the essentials out of the way, let us jump right in and code our first coroutine:

def bare_bones():
    while True:
        value = (yield)

It's clear to see the resemblance to a regular Python function. The while True: block guarantees the continuous execution of the coroutine for as long as it receives values.

The value is collected through the yield statement. We'll come back to this in a few moments...

It's clear to see that this code is practically useless, so we'll round it off with a few print statements:

def bare_bones():
    print("My first Coroutine!")
    while True:
        value = (yield)
        print(value)

Now, what happens when we try to call it like so:

coroutine = bare_bones()

If this were a normal Python function, one would expect it to produce some sort of output by this point. But if you run the code in its current state you will notice that not a single print() gets called.

That is because coroutines require the next() method to be called first:

def bare_bones():
    print("My first Coroutine!")
    while True:
        value = (yield)
        print(value)

coroutine = bare_bones()
next(coroutine)

This starts the execution of the coroutine until it reaches its first breakpoint - value = (yield). Then, it stops, returning the execution over to the main, and idles while awaiting new input:

My first Coroutine!

New input can be sent with send():

coroutine.send("First Value")

Our variable value will then receive the string First Value, print it, and a new iteration of the while True: loop forces the coroutine to once again wait for new values to be delivered. You can do this as many times as you like.

Finally, once you are done with the coroutine and no longer wish to make use of it you can free those resources by calling close(). This raises a GeneratorExit exception that needs to be dealt with:

def bare_bones():
    print("My first Coroutine!")
    try:
        while True:
            value = (yield)
            print(value)
    except GeneratorExit:
        print("Exiting coroutine...")

coroutine = bare_bones()
next(coroutine)
coroutine.send("First Value")
coroutine.send("Second Value")
coroutine.close()

Output:

My first Coroutine!
First Value
Second Value
Exiting coroutine...

Passing Arguments

Much like functions, coroutines are also capable of receiving arguments:

def filter_line(num):
    while True:
        line = (yield)
        if num in line:
            print(line)

cor = filter_line("33")
next(cor)
cor.send("Jessica, age:24")
cor.send("Marco, age:33")
cor.send("Filipe, age:55")

Output:

Marco, age:33

Applying Several Breakpoints

Multiple yield statements can be sequenced together in the same individual coroutine:

def joint_print():
    while True:
        part_1 = (yield)
        part_2 = (yield)
        print("{} {}".format(part_1, part_2))

cor = joint_print()
next(cor)
cor.send("So Far")
cor.send("So Good")

Output:

So Far So Good

The StopIteration Exception

After a coroutine is closed, calling send() again will generate a StopIteration exception:

def test():
    while True:
        value = (yield)
        print(value)
try:
    cor = test()
    next(cor)
    cor.close()
    cor.send("So Good")
except StopIteration:
    print("Done with the basics")

Output:

Done with the basics

Coroutines with Decorators

This is all well and good! But when working in larger projects initiating every single coroutine manually can be such a huge drag!

Worry not, its just the matter of exploiting the power of Decorators so we no longer need to use the next() method:

def coroutine(func):
    def start(*args, **kwargs):
        cr = func(*args, **kwargs)
        next(cr)
        return cr
    return start

@coroutine
def bare_bones():
    while True:
        value = (yield)
        print(value)

cor = bare_bones()
cor.send("Using a decorator!")

Running this piece of code will yield:

Using a decorator!

Building Pipelines

A pipeline is a sequence of processing elements organized so that the output of each element is the input of the next.

Data gets pushed through the pipe until it is eventually consumed. Every pipeline requires at least one source and one sink.

The remaining stages of the pipe can perform several different operations, from filtering to modifying, routing, and reducing data:

Coroutines are natural candidates for performing these operations, they can pass data between one another with send() operations and can also serve as the end-point consumer. Let's look at the following example:

def producer(cor):
    n = 1
    while n < 100:
        cor.send(n)
        n = n * 2

@coroutine
def my_filter(num, cor):
    while True:
        n = (yield)
        if n < num:
            cor.send(n)

@coroutine
def printer():
    while True:
        n = (yield)
        print(n)

prnt = printer()
filt = my_filter(50, prnt)
producer(filt)

Output:

1
2
4
8
16
32

So, what we have here is the producer() acting as the source, creating some values that are then filtered before being printed by the sink, in this case, the printer() coroutine.

my_filter(50, prnt) acts as the single intermediary step in the pipeline and receives its own coroutine as an argument.

This chaining perfectly illustrates the strength of coroutines: they are scalable for bigger projects (all that is required is to add more stages to the pipeline) and easily maintainable (changes to one don't force an entire rewrite of the source code).

Similarities to Objects

A sharp-eyed programmer might catch on that coroutines contain a certain conceptual similarity to Python objects. From the required prior definition to instance declaration and management. The obvious question arises of why one would use coroutines over the tried and true paradigm of object-oriented programming.

Well, aside the obvious fact that coroutines require but a single function definition, they also benefit from being significantly faster. Let's examine the following code:

class obj:
    def __init__(self, value):
        self.i = value
    def send(self, num):
        print(self.i + num)

inst = obj(1)
inst.send(5)

def coroutine(value):
    i = value
    while True:
        num = (yield)
        print(i + num)

cor = coroutine(1)
next(cor)
cor.send(5)

Here's how these two hold up against each other, when ran through the timeit module, 10,000 times:

Object	Coroutine
0.791811	0.6343617
0.7997058	0.6383156
0.8579286	0.6365501
0.838439	0.648442
0.9604255	0.7242559

Both perform the same menial task but the second example is quicker. Speed gains advent from the absence of the object's self lookups.

For more system-taxing tasks, this feature makes for a compelling reason to use coroutines instead of the conventional handler objects.

Caution when Using Coroutines

The send() Method is Not Thread-Safe

import threading
from time import sleep

def print_number(cor):
    while True:
        cor.send(1)

def coroutine():
    i = 1
    while True:
        num = (yield)
        print(i)
        sleep(3)
        i += num

cor = coroutine()
next(cor)

t = threading.Thread(target=print_number, args=(cor,))
t.start()

while True:
    cor.send(5)

Because send() was not properly synchronized, nor does it have inherent protection against thread related miscalls, the following error was raised: ValueError: generator already executing.

Mixing coroutines with concurrency should be done with extreme caution.

It's not Possible to Loop Coroutines

def coroutine_1(value):
    while True:
        next_cor = (yield)
        print(value)
        value = value - 1
        if next_cor != None:
            next_cor.send(value)

def coroutine_2(next_cor):
    while True:
        value = (yield)
        print(value)
        value = value - 2
        if next != None:
            next_cor.send(value)

cor1 = coroutine_1(20)
next(cor1)
cor2 = coroutine_2(cor1)
next(cor2)
cor1.send(cor2)

The same ValueError shows its face. From these simple examples we can infer that the send() method builds a sort of call-stack that doesn't return until the target reaches its yield statement.

So, using coroutines is not all sunshine and rainbows, careful thought must be had before application.

Conclusion

Coroutines provide a powerful alternative to the usual data processing mechanisms. Units of code can be easily combined, modified and rewritten, all the while profiting from variable persistence across its life cycle.

In the hands of a crafty programmer, coroutines become meaningful new tools by allowing simpler design and implementation, all the while providing significant performance gains.

Stripping ideas down into straightforward processes saves the programmer's effort and time, all the while avoiding stuffing code with superfluous objects that do nothing more than elementary tasks.