Curio - Async Await Async Await Async Await - Bad Curio! Sit! Good Curio!
Clone or download
dabeaz Merge pull request #280 from alekibango/master
add join for tasks in semaphore example
Latest commit 190e79a Oct 9, 2018


curio - concurrent I/O

Curio is a library of building blocks for performing concurrent I/O and common system programming tasks such as launching subprocesses, working with files, and farming work out to thread and process pools. It uses Python coroutines and the explicit async/await syntax introduced in Python 3.5. Its programming model is based on cooperative multitasking and existing programming abstractions such as threads, sockets, files, subprocesses, locks, and queues. You'll find it to be small, fast, and fun.

Curio has no third-party dependencies and does not use the standard asyncio module. Most users will probably find it to be a bit too-low level--it's probably best to think of it as a library for building libraries. Although you might not use it directly, many of its ideas have influenced other libraries with similar functionality.

Important Disclaimer

Curio is experimental software that currently only works on POSIX systems (OS X, Linux, etc.). Although it is a work in progress, it is extensively documented and has a fairly comprehensive test suite. Just be aware that the programming API is fluid and could change at any time. Although curio can be installed via pip, the version uploaded on PyPI is only updated occasionally. You'll probably get better results using the version cloned from github. You'll also want to make sure you're using Python 3.6. Of course, your mileage might vary.

Quick install

pip install git+

A Simple Example

Curio provides the same basic primitives that you typically find with thread programming. For example, here is a simple concurrent TCP echo server implemented using sockets and curio:


from curio import run, spawn
from curio.socket import *

async def echo_server(address):
    sock = socket(AF_INET, SOCK_STREAM)
    sock.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
    print('Server listening at', address)
    async with sock:
        while True:
            client, addr = await sock.accept()
            await spawn(echo_client, client, addr, daemon=True)

async def echo_client(client, addr):
    print('Connection from', addr)
    async with client:
         while True:
             data = await client.recv(100000)
             if not data:
             await client.sendall(data)
    print('Connection closed')

if __name__ == '__main__':
    run(echo_server, ('',25000))

If you have programmed with threads, you'll find that curio looks similar. You'll also find that the above server can handle thousands of simultaneous client connections even though no threads are being used under the hood.

Of course, if you prefer something a little higher level, you can have curio take care of the fiddly bits related to setting up the server portion of the code:


from curio import run, tcp_server

async def echo_client(client, addr):
    print('Connection from', addr)
    while True:
        data = await client.recv(100000)
        if not data:
        await client.sendall(data)
    print('Connection closed')

if __name__ == '__main__':
    run(tcp_server, '', 25000, echo_client)

A Complex Example

The above example only illustrates a few basics. You'll find Curio to be a bit more interesting if you try something more complicated.

As an example, one such problem is that of making a client TCP connection on a dual IPv4/IPv6 network. On such networks, functions such as socket.getaddrinfo() return a series of possible connection endpoints. To make a connection, each endpoint is tried until one of them succeeds. However, serious usability problems arise if this is done as a purely sequential process since bad connection requests might take a considerable time to fail. It's better to try several concurrent connection requests and use the first one that succeeds.

One solution to this problem is the so-called "Happy Eyeballs" algorithm as described in RFC 6555. You can read the RFC for more details, but Nathaniel Smith's Pyninsula Talk talk gives a pretty good overview of the problem and one possible implementation solution. The gist of the algorithm is that a client makes concurrent time-staggered connection requests and uses the first connection that is successful. What makes it tricky is that the algorithm involves a combination of timing, concurrency, and task cancellation--something that would be pretty hard to coordinate using a classical approach involving threads.

Here is an example of how the problem can be solved with Curio:

from curio import socket, TaskGroup, ignore_after, run
import itertools

async def open_tcp_stream(hostname, port, delay=0.3):
    # Get all of the possible targets for a given host/port
    targets = await socket.getaddrinfo(hostname, port, type=socket.SOCK_STREAM)
    if not targets:
        raise OSError(f'nothing known about {hostname}:{port}')

    # Cluster the targets into unique address families (e.g., AF_INET, AF_INET6, etc.)
    # and make sure the first entries are from a different family.
    families = [ list(g) for _, g in itertools.groupby(targets, key=lambda t: t[0]) ]
    targets = [ fam.pop(0) for fam in families ]

    # List of accumulated errors to report in case of total failure
    errors = []

    # Task group to manage a collection concurrent tasks.
    # It waits for a single task to return a non-None object
    # and cancels all remaining tasks when complete.
    async with TaskGroup(wait=object) as group:

        # Attempt to make a connection request
        async def try_connect(sockargs, addr, errors):
            sock = socket.socket(*sockargs)
                await sock.connect(addr)
                return sock
            except Exception as e:
                await sock.close()

       # Walk the list of targets and try connections with a staggered delay
        for *sockargs, _, addr in targets:
            await group.spawn(try_connect, sockargs, addr, errors)
            async with ignore_after(delay):
                 sock = await group.next_result()
                 if sock:
                     return sock

    if group.completed:
        return group.completed.result
        raise OSError(errors)

# Example use:
async def main():
    result = await open_tcp_stream('', 80)


This might require a bit of study, but the key to this solution is the Curio TaskGroup instance which represents a collection of managed concurrently executing tasks. Tasks created in the group aren't allowed to live beyond the lifetime of the code defined in the associated async with context manager block. Inside this block, you'll find statements that spawn tasks and wait for a result to come back with a time delay. When a successful connection is made, it is returned and any remaining tasks are magically cancelled (the wait=object controls this behavior). That's pretty neat.

Thread Interoperability Example

One of the more notable features of Curio is how it can interoperate with traditional synchronous code. For example, maybe you have a standard function that reads off a queue like this:

def consumer(queue):
    while True:
        item = queue.get()
        if item is None:
        print('Got:', item)

There is nothing too special here. This is something you might write using standard thread-programming. However, it's easy to make this code read data sent from a Curio async task. Use a UniversalQueue object like this:

from curio import UniversalQueue, run, sleep, spawn
from threading import Thread

async def producer(n, queue):
    for x in range(n):
        await queue.put(x)
        await sleep(1)
    await queue.put(None)

async def main():
    q = UniversalQueue()
    Thread(target=consumer, args=(q,)).start()
    t = await spawn(producer, 10, q)
    await t.join()


As the name implies, UniversalQueue is a queue that can be used in both synchronous and asynchronous code. The API is the same. It just works.

Additional Features

Curio provides additional support for SSL connections, synchronization primitives (events, locks, recursive locks, semaphores, and condition variables), queues, Unix signals, subprocesses, as well as running tasks in threads and processes. The task model fully supports cancellation, timeouts, monitoring, and other features critical to writing reliable code.

The two examples shown are only a small sample of what's possible. Read the official documentation for more in-depth coverage. The tutorial is a good starting point. The howto describes how to carry out various tasks. The developer guide describes the general design of Curio and how to use it in more detail.

Talks Related to Curio

Most of the principles behind Curio's design and general issues related to async programming have been described in various conference talks:

Additional Resources

The Big Question: Why?

Python already has a variety of libraries for async and event driven I/O. So, why create yet another library? There is no simple answer to that question, but here are a few of the motivations for creating Curio.

  • Python 3 has evolved considerably as a programming language and has adopted many new language features that are well-suited to cleanly writing a library like this. For example, improved support for non-blocking I/O, support for delegation to subgenerators (yield from) and the introduction of explicit async and await syntax in Python 3.5. Curio takes full advantage of these features and is not encumbered by issues of backwards compatibility with legacy Python code written 15 years ago.
  • Existing I/O libraries are mainly built on event-loops, callback functions, futures, and various abstractions that predate Python's proper support for coroutines. As a result, they are either overly complicated or dependent on esoteric magic involving C extensions, monkeypatching, or reimplementing half of the TCP flow-control protocol. Curio is a ground-up implementation that takes a different approach to the problem while relying upon known programming techniques involving sockets and files. If you have previously written synchronous code using processes or threads, curio will feel familiar. That is by design.
  • Simplicity is an important part of writing reliable systems software. Some of this simplicity comes from making intuitive programming APIs, but simplicity also comes from details of the implementation itself. Although parts of Curio may appear magical, it's actually built around a very small core of functionality centered on task scheduling. There is considerably less design complexity in the internals of Curio than what's typically found in a normal async framework. This is also a big reason why Curio is fast.
  • It's fun.

Questions and Answers

Q: Is curio implemented using the asyncio module?

A: No. Curio is a standalone library. Although the core of the library uses the same basic machinery as asyncio to poll for I/O events, the handling of those events is carried out in a completely different manner.

Q: Is curio meant to be a clone of asyncio?

A: No. Although curio provides a significant amount of overlapping functionality, the API is different. Compatibility with other libaries is not a goal.

Q: Is there any kind of overarching design philosophy?

A: Yes and no. The "big picture" design of curio is mainly inspired by the kernel/user space isolation found in operating systems. Beyond that, curio takes a generally pragmatic view towards concurrent programming techniques. It's probably best to view curio as providing a base set of primitives upon which you can build all sorts of interesting things. Yes, you can use it to shoot yourself in the foot.

Q: How many tasks can be created?

A: Each task involves an instance of a Task class that encapsulates a generator. No threads are used. As such, you're really only limited by the memory of your machine--potentially you could have hundreds of thousands of tasks. The I/O functionality in curio is implemented using the built-in selectors module. Thus, the number of open sockets allowed would be subject to the limits of that library combined with any per-user limits imposed by the operating system.

Q: Can curio interoperate with other event loops?

A: It depends on what you mean by the word "interoperate." Curio's preferred mechanism of communication with the external world is a queue. It is possible to communicate between Curio, threads, and other event loops using queues. Curio can also submit work to the asyncio event loop with the provision that it must be running separately in a different thread.

Q: How fast is curio?

A: In rough benchmarking of the simple echo server shown here, curio runs about 90% faster than comparable code using coroutines in asyncio and about 50% faster than similar code written using Trio. This was last measured on Linux using Python 3.7b3. Keep in mind there is a lot more to overall application performance than the performance of a simple echo server so your mileage might vary. See the examples/benchmark directory for various testing programs.

Q: Is curio going to evolve into a framework?

A: No, because evolving into a framework would mean modifying curio to actually do something. If it actually did something, then people would start using it to do things. And then all of those things would have to be documented, tested, and supported. People would start complaining about how all the things related to the various built-in things should have new things added to do some crazy thing. No forget that. Curio remains committed to not doing much of anything the best it can. This includes not implementing HTTP.

Q: What are future plans?

A: Future work on curio will primarily focus on features related to performance, debugging, diagnostics, and reliability. A main goal is to provide a robust environment for running and controlling concurrent tasks. However, it's also supposed to be fun. A lot of time is being spent thinking about the API and how to make it pleasant.

Q: Is there a curio sticker?

A: No. However, you can make a stencil

Q: Is there a curio community?

A: Curio is a lifestyle.

Q: I see various warnings about not using curio. What should I do?

A: Has programming taught you nothing? Warnings are meant to be ignored. Of course you should use curio. However, be aware that the main reason you shouldn't be using curio is that you should be using it.

Q: Can I contribute?

A: Absolutely. Please use the Github page at as the primary point of discussion concerning pull requests, bugs, and feature requests.


Read the official docs here:

Discussion Forum

A discussion forum for curio is available at Please go there to ask questions and find out whats happening with the project.


  • David Beazley
  • Brett Cannon
  • Nathaniel Smith
  • Alexander Zhukov
  • Laura Dickinson


Curio was created by David Beazley (@dabeaz).

It is a young project. All contributions welcome.