Incremental read: Why only Uint8Array instead of BufferSource? about fetch

[annevk]

I don't really understand why Web Transport does that. That seems like a bug. We decided that we'd use Uint8Array generally for byte buffer representation.

[saschanaz]

Hmm, it makes sense to return Uint8Array but I don't see the point to accept only Uint8Array?

[saschanaz]

The Encoding APIs also accept (AllowShared)BufferSource via TextDecoder[Stream] and emits Uint8Array via TextEncoder[Stream], which makes sense to me. Would be nice to do some IDL crawl and see how many APIs are currently accepting Uint8Array or BufferSource.

[asutherland]

Would be nice to do some IDL crawl and see how many APIs are currently accepting Uint8Array or BufferSource.

Do the pages https://dontcallmedom.github.io/webidlpedia/names/BufferSource.html and https://dontcallmedom.github.io/webidlpedia/names/Uint8Array.html from https://dontcallmedom.github.io/webidlpedia/names/ accomplish what you want?

[saschanaz]

Thanks Andrew! That clearly shows BufferSource is winning 🥇

[annevk]

Someone will have to figure out why we made this decision this way as there was a reason and I'd rather not flip on it based on popularity. I suspect you can find the rationale either in this repository, Streams, or the "fetch with streams" repository (not sure what the link is, but it should be linked from issues in this repository).

[saschanaz]

The first addition happened in #449, Anne in 2016 said allowing only Uint8Array is the goal in #441 (comment), I don't immediately see the background behind that but I assume maybe that was what implementations did?

[annevk]

No, it was aspirational. Initially we designed it such that (byte) streams would only emit Uint8Array to consumers. And later we adopted that for the other direction as well. Mainly for consistency I think, but also with some hope that would allow optimizations.

Most things in the platform that accept buffers should indeed accept BufferSource, but they're not dealing with streams.

[saschanaz]

Mainly for consistency I think

Emit Uint8Array, so accept only Uint8Array too, you mean? Kinda make sense but still not consistent with other APIs at this point.

but also with some hope that would allow optimizations.

Do you remember in what way it can allow optimizations? Getting bytes from typed arrays should be straightforward both in JS and C++, so I don't really follow here.

[annevk]

I don't think byte streams have to be consistent with other APIs necessarily. They're somewhat unique. Also, consistency within seems more important.

[saschanaz]

I mean, ReadableByteStreamController accepts ArrayBufferView instead of just Uint8Array: https://streams.spec.whatwg.org/#ref-for-rbs-controller-enqueue (Not BufferSource, hmm.)

Edit: Based on that I still think "Emit Uint8Array, receive more broad types" pattern is used across many specs and thus gives consistency.

[saschanaz]

Here is an inconsistency:

// One can enqueue Uint16Array with type: bytes
await new Response(new ReadableStream({
  start(c) {
    c.enqueue(new Uint16Array([0xD55C, 0xAE00]));
    c.close();
  },
  type: "bytes"
})).blob()

// But this fails
await new Response(new ReadableStream({
  start(c) {
    c.enqueue(new Uint16Array([0xD55C, 0xAE00]));
    c.close();
  }
})).blob()