You don't need promises in Python: just use async/await!

Promises/A+ logo - https://alexn.org/assets/img/2017/then.png

If you're coming from a JavaScript background, it's tempting to try to use the promises that you know and love with Python. That's what I tried to initially too, and I was surprised to see that promises were very rarely used in Python. It turns out that promises are not pythonic: I should have used async/await instead. This post explains why async/await is a better idiom that you can use both in Python and JavaScript.

I found this surprising at first: I thought that promises were very different from Python's async/await. But they're not! The code is written differently (the syntax is different), but what is happening under the hood is actually the same (the semantics are equivalent).

The main takeaway from this post should be that it's a good idea to use the Python idioms everywhere because they are easier to learn and are available in many languages, allowing programmers to switch between languages more easily. In short, I think you should use async/await both in JavaScript and Python.

Motivation

When I first studied promises in JavaScript and async/await in Python, they looked quite different to me. Indeed, promises can use a lot of anonymous functions and they handle errors differently from other JavaScript code. For example:

function getProcessedData(url) {
  downloadData(url)
  .catch(e => downloadFallbackData(url))
  .then(data => processDataInWorker(data))
}

Here's how we would write this in Python with async/await:

async def getProcessedData(url):
    try:
        v = await downloadData(url)
    except IOError:
        v = await downloadFallbackData(url)
    await processDataInWorker(v)

Quite different! The Python code here reads like normal, synchronous code (with an added async and a few await keywords). It's also handling exceptions using a mechanism that is standard in many languages. My goal is to convince you that those two snippets are actually equivalent.

The main reason why they are equivalent is that Python and JavaScript both use an event loop when doing asynchronous I/O. Let's explain what that means.

The goal: not blocking for I/O

First things first, we need to understand what problem such an event loop is trying to solve.

Say you want to write a web server from scratch. One of the first problems you will want to tackle is: how do I accept connections from multiple clients at the same time?

The key insight to solve this issue is that a given request can trigger a lot of input/output (I/O): reading files, accessing databases and actually communicating with the client via the Internet. During I/O, the computer has nothing to do: he can simply take advantage of this to serve the other requests in the meantime.

This idea works in many situations:

  • Web browsers don't wait for one resource to arrive to fetch the next one.
  • A web scraper will not wait to fully receive one page to start fetching the next one.
  • And in a JavaScript web application, different events can take a different amount of time to handle: you don't want to block scrolling events even when loading resources from the server.

All those common use cases are said to be I/O bound, and you can take advantage of this to improve throughput: each request won't be faster, but all of them will finish much sooner.

The idea is simple enough. But how do we do it?

Threads to the rescue?

To execute I/O bound requests concurrently, the solution is usually multithreading. Multithreading actually solves other problems than performing I/O in parallel, because it also allows to run any code (not only I/O operations) in parallel on multiple cores.

Anyway, what's important to us here is that all operating systems will suspend a thread that is waiting for I/O to allow other threads to get work done. This is actually very efficient! This is how projects like Apache and uWSGI work, and it's also supported in nginx. And even though multithreading as a programming model has bad reputation, in those cases it's usually just fine because there is very little actual shared state and your web server takes care of it. (Note that I'm using the word "thread" here but this also applies to processes, especially on Linux where threads are just processes with less overhead.)

Event loop

However, in JavaScript, threads were initially not supported! Fortunately, there are other ways to perform multiple I/O operations in parallel. One way is using an single-threaded event loop.

An event loop is well, a loop! I don't want to go into too much details, but if you're interested, look at this toy reimplementation of an event loop in Python which explains very nicely how we can write an event loop step by step. However, for our purposes you only need to know this:

  • The job of the event loop is to wait for I/O without blocking the rest of the code.
  • The job of the programmer is to tell the loop when it is going perform I/O and what to do when the I/O is over.

What's interesting with event loops is the way the allow programmers to reason about concurrency more easily. Only I/O operations can be executed in parallel, which means that when you're into a block of code between two I/O operations, you know that nothing is going to change under you, which is makes things much easier than with threads!

Okay, but what does it look like in practice? There are different ways to program using an event loop, let's look at them.

Levels of abstraction

The initial mechanism is asking the event loop to call back a function when an event is ready, as is common in graphical interfaces programming. In the early 2000s, when Ajax was all the rage, this is what was happening: when a request was ready, a callback would be called via onreadystatechange.

However, using callbacks every time I/O is needed is not convenient and leads to spaghetti code, also known as callback hell. Gradually, language designers came up with better ways do this. Here they are, from most low-level to most convenient:

  1. low-level state machines (for example with mio in Rust)
  2. callbacks as in the first versions of Node.js or Ajax
  3. Promises that are now widely used in the JavaScript world
  4. async/await as introduced by C# and supported by Python and ES2017

Each new level is a higher level of abstraction that makes the resulting code more readable. Interestingly, the highest level of abstraction can be made as fast as the lowest one, as proven by Rust zero-cost futures and async!/await! macros.

While promises improved the situation greatly, they still require to learn a different control flow. async/await, on the other hand, allows the programmer to reason with the usual tools used in synchronous code.

Python is at the async/await level, JS is mostly at the promises level even though async/await is supported in Node.js since 2016 and is pretty well supported in browsers and compilable away for IE 11.

Moving from Promises to async/await in JavaScript

Okay, as a JavaScript programmer, why would you want to make the switch from promises to async/await? Remember the code above? No, don't scroll! Here it is:

function getProcessedData(url) {
  downloadData(url)
  .catch(e => downloadFallbackData(url))
  .then(data => processDataInWorker(data))
}

You can turn it into this:

async function getProcessedData(url) {
  let v;
  try {
    v = await downloadData(url);
  } catch(e) {
    v = await downloadFallbackData(url);
  }
  await processDataInWorker(v);
}

This has three benefits.

  • You can use existing idioms such as try/catch blocks.
  • Programmers coming from other languages can get up to speed quickly.
  • It's fully compatible with promises, in the sense you can simply await on a promise, not only on an async function.

And it's now identical to the Python equivalent:

async def getProcessedData(url):
    try:
        v = await downloadData(url)
    except Exception:
        v = await downloadFallbackData(url)
    await processDataInWorker(v)

Okay, a few less braces. :)

Conclusion

It's really interesting to see that many languages agree that async/await is the best way to express asynchronous I/O in combination with using an event loop. Other languages that support this idiom are C# (who introduced it) and Rust (who considers it essential to bring futures to developers). Other languages that support this are Dart, Kotlin and Scala: I expect the list to continue to grow.

Using async/await is not that difficult: learn how to do it!

Thanks to Julien Pradet for the detailed and insightful review that helped to massively improve this blog post.

Consider following me on Twitter.

Comments !