dpa

• Deterministic: always rejects with the same promise for a list of promises where some reject regardless of which one rejects first
• Performant: always resolves or rejects as fast as or faster than Promise.allSettled for the same input
• Familiar: same API as Promise.all including iterables support
• Tiny: ~130 bytes minzipped!

$ npm i dpa

Deterministic promise all. I promise (pun intended!) it's not as esoteric as it sounds!

Suppose you have some asynchronous authorization functions, each resolving to undefined or rejecting, that you'd like to run on all requests to an endpoint. You care about performance! So you run the authorization checks concurrently using Promise.all and display an error to the user on rejection.

Using Remix it might look something like this:

import { useCatch } from 'remix'

export const loader = async ({ request }) => {
  await Promise.all([
    checkThing1(request),
    checkThing2(request),
    checkThing3(request),
  ])

  // Do authorized stuff...
}

// Rendered when an error response is thrown in `loader`
export const CatchBoundary = () => {
  const caught = useCatch()

  // Return some JSX...
}

Everything seems to work great, but what if the promises returned by checkThing1 and checkThing2 both reject? What does the user see? The answer is it depends on which one rejects first!

That's right. Your error page is nondeterministic. The user can visit the same URL with the same authorization state and receive a different page purely based on how quickly each authorization check completes.

Another problematic case in a framework like Remix is redirects. In Remix you can redirect by throwing redirect responses. If you use Promise.all to concurrently run a bunch of functions that may throw redirects, then your redirects are nondeterministic too.

You could use Promise.allSettled to wait for all the promises to either resolve or reject, then if any reject you could reject with the first rejected promise in your list of promises. This way you reject with the same promise regardless of which one rejects first (time-wise). This is what Remix was doing to keep concurrent loader execution deterministic (and maybe still does, but I couldn't figure where that's done at the latest commit).

But can we do better? What if the first promise in your list is the first promise to reject? We could reject right away and still be deterministic! But with Promise.allSettled we're stuck waiting for every promise to resolve in all cases...

More generally, if a promise in your list rejects and every promise before it in the list resolved, then we can immediately reject with that promise. That's what dpa does. It's deterministic while being strictly as fast as or faster than Promise.allSettled!

import dpa from 'dpa'

const delay = ms => new Promise(resolve => setTimeout(resolve, ms))
const test = async fn => {
  const start = Date.now()
  try {
    console.log(await fn())
  } catch (error) {
    console.log(`${error} thrown`)
  } finally {
    const elapsed = Date.now() - start
    console.log(`${Math.trunc(elapsed / 1000)}s elapsed`)
  }
  console.log()
}

await test(() =>
  dpa([
    delay(1000).then(() => 1),
    delay(4000).then(() => 2),
    delay(5000).then(() => 3),
  ]),
)
// => [1, 2, 3]
// => 5s elapsed

await test(() =>
  dpa([
    delay(1000).then(() => 1),
    delay(4000).then(() => {
      throw 2
    }),
    delay(6000).then(() => {
      throw 3
    }),
  ]),
)
// => 2 thrown
// => 4s elapsed

await test(() =>
  dpa([
    dpa([
      delay(1000).then(() => 1),
      delay(6000).then(() => {
        throw 2
      }),
      delay(2000).then(() => {
        throw 3
      }),
    ]),
  ]),
)
// => 2 thrown
// => 6s elapsed

Stars are always welcome!

For bugs and feature requests, please create an issue.

For pull requests, please read the contributing guidelines.

Apache License 2.0

This is not an official Google product.