overreactedby Dan Abramov

Algebraic Effects for the Rest of Us

July 21, 2019

Have you heard about algebraic effects?

My first attempts to figure out what they are or why I should care about them were unsuccessful. I found a few pdfs but they only confused me more. (There’s something about academic pdfs that makes me sleepy.)

But my colleague Sebastian kept referring to them as a mental model for some things we do inside of React. (Sebastian works on the React team and came up with quite a few ideas, including Hooks and Suspense.) At some point, it became a running joke on the React team, with many of our conversations ending with:

"Algebraic Effects" caption on the "Ancient Aliens" guy meme

It turned out that algebraic effects are a cool concept and not as scary as I thought from those pdfs. If you’re just using React, you don’t need to know anything about them — but if you’re feeling curious, like I was, read on.

(Disclaimer: I’m not a programming language researcher, and might have messed something up in my explanation. I am not an authority on this topic so let me know!)

Not Production Ready Yet

Algebraic Effects are a research programming language feature. This means that unlike if, functions, or even async / await, you probably can’t really use them in production yet. They are only supported by a few languages that were created specifically to explore that idea. There is progress on productionizing them in OCaml which is… still ongoing. In other words, Can’t Touch This.

Edit: a few people mentioned that LISP languages do offer something similar, so you can use it in production if you write LISP.

So Why Should I Care?

Imagine that you’re writing code with goto, and somebody shows you if and for statements. Or maybe you’re deep in the callback hell, and somebody shows you async / await. Pretty cool, huh?

If you’re the kind of person who likes to learn about programming ideas several years before they hit the mainstream, it might be a good time to get curious about algebraic effects. Don’t feel like you have to though. It is a bit like thinking about async / await in 1999.

Okay, What Are Algebraic Effects?

The name might be a bit intimidating but the idea is simple. If you’re familiar with try / catch blocks, you’ll figure out algebraic effects very fast.

Let’s recap try / catch first. Say you have a function that throws. Maybe there’s a bunch of functions between it and the catch block:

function getName(user) {
  let name = user.name;
  if (name === null) {
  	throw new Error('A girl has no name');
  }
  return name;
}
 
function makeFriends(user1, user2) {
  user1.friendNames.push(getName(user2));
  user2.friendNames.push(getName(user1));
}
 
const arya = { name: null, friendNames: [] };
const gendry = { name: 'Gendry', friendNames: [] };
try {
  makeFriends(arya, gendry);
} catch (err) {
  console.log("Oops, that didn't work out: ", err);
}

We throw inside getName, but it “bubbles” up right through makeFriends to the closest catch block. This is an important property of try / catch. Things in the middle don’t need to concern themselves with error handling.

Unlike error codes in languages like C, with try / catch, you don’t have to manually pass errors through every intermediate layer in the fear of losing them. They get propagated automatically.

What Does This Have to Do With Algebraic Effects?

In the above example, once we hit an error, we can’t continue. When we end up in the catch block, there’s no way we can continue executing the original code.

We’re done. It’s too late. The best we can do is to recover from a failure and maybe somehow retry what we were doing, but we can’t magically “go back” to where we were, and do something different. But with algebraic effects, we can.

This is an example written in a hypothetical JavaScript dialect (let’s call it ES2025 just for kicks) that lets us recover from a missing user.name:

function getName(user) {
  let name = user.name;
  if (name === null) {
  	name = perform 'ask_name';
  }
  return name;
}
 
function makeFriends(user1, user2) {
  user1.friendNames.push(getName(user2));
  user2.friendNames.push(getName(user1));
}
 
const arya = { name: null, friendNames: [] };
const gendry = { name: 'Gendry', friendNames: [] };
try {
  makeFriends(arya, gendry);
} handle (effect) {
  if (effect === 'ask_name') {
  	resume with 'Arya Stark';
  }
}

(I apologize to all readers from 2025 who search the web for “ES2025” and find this article. If algebraic effects are a part of JavaScript by then, I’d be happy to update it!)

Instead of throw, we use a hypothetical perform keyword. Similarly, instead of try / catch, we use a hypothetical try / handle. The exact syntax doesn’t matter here — I just came up with something to illustrate the idea.

So what’s happening? Let’s take a closer look.

Instead of throwing an error, we perform an effect. Just like we can throw any value, we can pass any value to perform. In this example, I’m passing a string, but it could be an object, or any other data type:

function getName(user) {
  let name = user.name;
  if (name === null) {
  	name = perform 'ask_name';
  }
  return name;
}

When we throw an error, the engine looks for the closest try / catch error handler up the call stack. Similarly, when we perform an effect, the engine would search for the closest try / handle effect handler up the call stack:

try {
  makeFriends(arya, gendry);
} handle (effect) {
  if (effect === 'ask_name') {
  	resume with 'Arya Stark';
  }
}

This effect lets us decide how to handle the case where a name is missing. The novel part here (compared to exceptions) is the hypothetical resume with:

try {
  makeFriends(arya, gendry);
} handle (effect) {
  if (effect === 'ask_name') {
  	resume with 'Arya Stark';
  }
}

This is the part you can’t do with try / catch. It lets us jump back to where we performed the effect, and pass something back to it from the handler. 🤯

function getName(user) {
  let name = user.name;
  if (name === null) {
  	// 1. We perform an effect here
  	name = perform 'ask_name';
  	// 4. ...and end up back here (name is now 'Arya Stark')
  }
  return name;
}
 
// ...
 
try {
  makeFriends(arya, gendry);
} handle (effect) {
  // 2. We jump to the handler (like try/catch)
  if (effect === 'ask_name') {
  	// 3. However, we can resume with a value (unlike try/catch!)
  	resume with 'Arya Stark';
  }
}

This takes a bit of time to get comfortable with, but it’s really not much different conceptually from a “resumable try / catch”.

Note, however, that algebraic effects are much more flexible than try / catch, and recoverable errors are just one of many possible use cases. I started with it only because I found it easiest to wrap my mind around it.

A Function Has No Color

Algebraic effects have interesting implications for asynchronous code.

In languages with an async / await, functions usually have a “color”. For example, in JavaScript we can’t just make getName asynchronous without also “infecting” makeFriends and its callers with being async. This can be a real pain if a piece of code sometimes needs to be sync, and sometimes needs to be async.

// If we want to make this async...
async getName(user) {
  // ...
}
 
// Then this has to be async too...
async function makeFriends(user1, user2) {
  user1.friendNames.push(await getName(user2));
  user2.friendNames.push(await getName(user1));
}
 
// And so on...

JavaScript generators are similar: if you’re working with generators, things in the middle also have to be aware of generators.

So how is that relevant?

For a moment, let’s forget about async / await and get back to our example:

function getName(user) {
  let name = user.name;
  if (name === null) {
  	name = perform 'ask_name';
  }
  return name;
}
 
function makeFriends(user1, user2) {
  user1.friendNames.push(getName(user2));
  user2.friendNames.push(getName(user1));
}
 
const arya = { name: null, friendNames: [] };
const gendry = { name: 'Gendry', friendNames: [] };
try {
  makeFriends(arya, gendry);
} handle (effect) {
  if (effect === 'ask_name') {
  	resume with 'Arya Stark';
  }
}

What if our effect handler didn’t know the “fallback name” synchronously? What if we wanted to fetch it from a database?

It turns out, we can call resume with asynchronously from our effect handler without making any changes to getName or makeFriends:

function getName(user) {
  let name = user.name;
  if (name === null) {
  	name = perform 'ask_name';
  }
  return name;
}
 
function makeFriends(user1, user2) {
  user1.friendNames.push(getName(user2));
  user2.friendNames.push(getName(user1));
}
 
const arya = { name: null, friendNames: [] };
const gendry = { name: 'Gendry', friendNames: [] };
try {
  makeFriends(arya, gendry);
} handle (effect) {
  if (effect === 'ask_name') {
  	setTimeout(() => {
      resume with 'Arya Stark';
  	}, 1000);
  }
}

In this example, we don’t call resume with until a second later. You can think of resume with as a callback which you may only call once. (You can also impress your friends by calling it a “one-shot delimited continuation.”)

Now the mechanics of algebraic effects should be a bit clearer. When we throw an error, the JavaScript engine “unwinds the stack”, destroying local variables in the process. However, when we perform an effect, our hypothetical engine would create a callback with the rest of our function, and resume with calls it.

Again, a reminder: the concrete syntax and specific keywords are made up for this article. They’re not the point, the point is in the mechanics.

A Note on Purity

It’s worth noting that algebraic effects came out of functional programming research. Some of the problems they solve are unique to pure functional programming. For example, in languages that don’t allow arbitrary side effects (like Haskell), you have to use concepts like Monads to wire effects through your program. If you ever read a Monad tutorial, you know they’re a bit tricky to think about. Algebraic effects help do something similar with less ceremony.

This is why so much discussion about algebraic effects is incomprehensible to me. (I don’t know Haskell and friends.) However, I do think that even in an impure language like JavaScript, algebraic effects can be a very powerful instrument to separate the what from the how in the code.

They let you write code that focuses on what you’re doing:

function enumerateFiles(dir) {
  const contents = perform OpenDirectory(dir);
  perform Log('Enumerating files in ', dir);
  for (let file of contents.files) {
  	perform HandleFile(file);
  }
  perform Log('Enumerating subdirectories in ', dir);
  for (let directory of contents.dir) {
  	// We can use recursion or call other functions with effects
  	enumerateFiles(directory);
  }
  perform Log('Done');
}

And later wrap it with something that specifies how:

let files = [];
try {
  enumerateFiles('C:\\');
} handle (effect) {
  if (effect instanceof Log) {
  	myLoggingLibrary.log(effect.message);
  	resume;
  } else if (effect instanceof OpenDirectory) {
  	myFileSystemImpl.openDir(effect.dirName, (contents) => {
      resume with contents;
  	});
  } else if (effect instanceof HandleFile) {
    files.push(effect.fileName);
    resume;
  }
}
// The `files` array now has all the files

Which means that those pieces can even become librarified:

import { withMyLoggingLibrary } from 'my-log';
import { withMyFileSystem } from 'my-fs';
 
function ourProgram() {
  enumerateFiles('C:\\');
}
 
withMyLoggingLibrary(() => {
  withMyFileSystem(() => {
    ourProgram();
  });
});

Unlike async / await or Generators, algebraic effects don’t require complicating functions “in the middle”. Our enumerateFiles call could be deep within ourProgram, but as long as there’s an effect handler somewhere above for each of the effects it may perform, our code would still work.

Effect handlers let us decouple the program logic from its concrete effect implementations without too much ceremony or boilerplate code. For example, we could completely override the behavior in tests to use a fake filesystem and to snapshot logs instead of outputting them to the console:

import { withFakeFileSystem } from 'fake-fs';
 
function withLogSnapshot(fn) {
  let logs = [];
  try {
  	fn();
  } handle (effect) {
  	if (effect instanceof Log) {
  	  logs.push(effect.message);
  	  resume;
  	}
  }
  // Snapshot emitted logs.
  expect(logs).toMatchSnapshot();
}
 
test('my program', () => {
  const fakeFiles = [/* ... */];
  withFakeFileSystem(fakeFiles, () => {
  	withLogSnapshot(() => {
	  ourProgram();
  	});
  });
});

Because there is no “function color” (code in the middle doesn’t need to be aware of effects) and effect handlers are composable (you can nest them), you can create very expressive abstractions with them.

A Note on Types

Because algebraic effects are coming from statically typed languages, much of the debate about them centers on the ways they can be expressed in types. This is no doubt important but can also make it challenging to grasp the concept. That’s why this article doesn’t talk about types at all. However, I should note that usually the fact that a function can perform an effect would be encoded into its type signature. So you shouldn’t end up in a situation where random effects are happening and you can’t trace where they’re coming from.

You might argue that algebraic effects technically do “give color” to functions in statically typed languages because effects are a part of the type signature. That’s true. However, fixing a type annotation for an intermediate function to include a new effect is not by itself a semantic change — unlike adding async or turning a function into a generator. Inference can also help avoid cascading changes. An important difference is you can “bottle up” an effect by providing a noop or a mock implementation (for example, a sync call for an async effect), which lets you prevent it from reaching the outer code if necessary — or turn it into a different effect.

Should We Add Algebraic Effects to JavaScript?

Honestly, I don’t know. They are very powerful, and you can make an argument that they might be too powerful for a language like JavaScript.

I think they could be a great fit for a language where mutation is uncommon, and where the standard library fully embraced effects. If you primarily do perform Timeout(1000), perform Fetch('http://google.com'), and perform ReadFile('file.txt'), and your language has pattern matching and static typing for effects, it might be a very nice programming environment.

Maybe that language could even compile to JavaScript!

How Is All of This Relevant to React?

Not that much. You can even say it’s a stretch.

If you watched my talk about Time Slicing and Suspense, the second part involves components reading data from a cache:

function MovieDetails({ id }) {
  // What if it's still being fetched?
  const movie = movieCache.read(id);
}

(The talk uses a slightly different API but that’s not the point.)

This builds on a React feature called “Suspense”, which is in active development for the data fetching use case. The interesting part, of course, is that the data might not yet be in the movieCache — in which case we need to do something because we can’t proceed below. Technically, in that case the read() call throws a Promise (yes, throws a Promise — let that sink in). This “suspends” the execution. React catches that Promise, and remembers to retry rendering the component tree after the thrown Promise resolves.

This isn’t an algebraic effect per se, even though this trick was inspired by them. But it achieves the same goal: some code below in the call stack yields to something above in the call stack (React, in this case) without all the intermediate functions necessarily knowing about it or being “poisoned” by async or generators. Of course, we can’t really resume execution in JavaScript later, but from React’s point of view, re-rendering a component tree when the Promise resolves is pretty much the same thing. You can cheat when your programming model assumes idempotence!

Hooks are another example that might remind you of algebraic effects. One of the first questions that people ask is: how can a useState call possibly know which component it refers to?

function LikeButton() {
  // How does useState know which component it's in?
  const [isLiked, setIsLiked] = useState(false);
}

I already explained the answer near the end of this article: there is a “current dispatcher” mutable state on the React object which points to the implementation you’re using right now (such as the one in react-dom). There is similarly a “current component” property that points to our LikeButton’s internal data structure. That’s how useState knows what to do.

Before people get used to it, they often think it’s a bit “dirty” for an obvious reason. It doesn’t “feel right” to rely on shared mutable state. (Side note: how do you think try / catch is implemented in a JavaScript engine?)

However, conceptually you can think of useState() as of being a perform State() effect which is handled by React when executing your component. That would “explain” why React (the thing calling your component) can provide state to it (it’s above in the call stack, so it can provide the effect handler). Indeed, implementing state is one of the most common examples in the algebraic effect tutorials I’ve encountered.

Again, of course, that’s not how React actually works because we don’t have algebraic effects in JavaScript. Instead, there is a hidden field where we keep the current component, as well as a field that points to the current “dispatcher” with the useState implementation. As a performance optimization, there are even separate useState implementations for mounts and updates. But if you squint at this code very hard, you might see them as essentially effect handlers.

To sum up, in JavaScript, throwing can serve as a crude approximation for IO effects (as long as it’s safe to re-execute the code later, and as long as it’s not CPU-bound), and having a mutable “dispatcher” field that’s restored in try / finally can serve as a crude approximation for synchronous effect handlers.

You can also get a much higher fidelity effect implementation with generators but that means you’ll have to give up on the “transparent” nature of JavaScript functions and you’ll have to make everything a generator. Which is… yeah.

Learn More

Personally, I was surprised by how much algebraic effects made sense to me. I always struggled understanding abstract concepts like Monads, but Algebraic Effects just “clicked”. I hope this article will help them “click” for you too.

I don’t know if they’re ever going to reach mainstream adoption. I think I’ll be disappointed if they don’t catch on in any mainstream language by 2025. Remind me to check back in five years!

I’m sure there’s so much more you can do with them — but it’s really difficult to get a sense of their power without actually writing code this way. If this post made you curious, here’s a few more resources you might want to check out:

  • https://github.com/ocamllabs/ocaml-effects-tutorial

  • https://www.janestreet.com/tech-talks/effective-programming/

  • https://www.youtube.com/watch?v=hrBq8R_kxI0

Many people also pointed out that if you omit the typing aspects (as I did in this article), you can find much earlier prior art for this in the condition system in Common Lisp. You might also enjoy reading James Long’s post on continuations that explains how the call/cc primitive can also serve as a foundation for building resumable exceptions in userland.

If you find other useful resources on algebraic effects for people with JavaScript background, please let me know on Twitter!


Discuss on 𝕏  ·  Edit on GitHub