HTTP add stream support

[adrieljss]

This PR was meant to add stream support for fetch.

What was changed:

Rust:

• Removed fetch_read_body, because we're already sending response body chunks through the IPC.
• Using reqwest::chunk to send an array of bytes (Vec<u8>) to JavaScript using the IPC Channel

JavaScript:

• return a ReadableStream as InitBody in Response, this should be automatically collected and function as normal even if not used for streaming responses.

Example Usage

const res = await fetch('...')
const rd = res.body.getReader()

while (true) {
  const { done, value } = await rd.read()
  if (done) break
  const txt = new TextDecoder().decode(value)
  console.log(txt)
}

or (TypeScript might not like this)

const res = await fetch('...')

for await (const chunk of res.body) {
  const txt = new TextDecoder().decode(chunk)
  console.log(txt)
}

closes #2140

[FabianLars]

Thanks for opening a PR! Tiny bit more context for reviewers: https://discord.com/channels/616186924390023171/1344666371316908063 This is supposed to be an alternative to #2140

@adrieljss Can you please revert the changes to Cargo.toml, Cargo.lock, package.json, and pnpm-lock.yaml ? Versioning is handled in CI and this would break stuff.

[github-actions]

Package Changes Through 194fdee

There are 4 changes which include http with minor, http-js with minor, log with minor, log-js with minor

Planned Package Versions

The following package releases are the planned based on the context of changes in this pull request.

package	current	next
api-example	2.0.19	2.0.20
api-example-js	2.0.15	2.0.16
http	2.3.0	2.4.0
http-js	2.3.0	2.4.0
log	2.2.3	2.3.0
log-js	2.2.3	2.3.0

Add another change file through the GitHub UI by following this link.

---

Read about change files or the docs at github.com/jbolda/covector

[amrbashir]

This is not performant because of serialization overhead.

To avoid this, you can send a tauri::Response which is just a Vec<u8> and then you could append a single byte at the end that is either 1 or 0, indicating the status of done. Or even an empty array would suffice.

I am not sure if the channel is as performant as invoke (will have to check that later) but would be better to avoid using the channel.

[adrieljss]

Yes you're right, I will change the struct, how about using serde_bytes? Since serde could be serializing the Vec as an array of numbers.

I don't think there's any way to do streaming using invoke(?) so maybe a solution would be having a seperate function for seperate usecases, one for streaming and one for collecting all the contents at the end, but the user has to specify that and that will not be fetch-like.

[adrieljss]

Though, if the server doesn't want to stream, the channel will only receive 1 chunk and that should be (around) the same performance since both uses IPC anyways.

[amrbashir]

this can be done with invoke and an AsyncIterator in JS:

• The removed fetch_read_body instead of using .bytes() it will use .chunk(), effictively returning a single chunk each time it is called
• On JS side, you can create an AsyncIterator that will be used to create the ReadableStream, on each iteration it will call fetch_read_body and append the data to the stream

[adrieljss]

Would that be better though? In channels, it said that is the recommended way for streaming data (like HTTP streaming, which is our case) to the frontend. I think the invoke approach will block operations until it has returned something. While channels is just an event listener.

In my opinion, I feel like it's more intuitive rather than spamming invokes

[amrbashir]

The channel calls invoke under the hood. Channels are more ergonomic but there is room for improvement in their implementation as it uses tauri::State with a Mutex unlike what I am proposing which would use a Mutex only until the request is sent, after that the body could be read without any Mutexes.

For now, let's use the channel approach and I can improve on it later (once I confirm that there is mutex overhead involved).

My main blocker for this PR is the use of serialization which we can avoid even with channels.

[adrieljss]

Yes okay, I will fix that 👍

[adrieljss]

@amrbashir can you check the new commit I did? I can't use tauri::Response inside the channel because it doesn't impl Clone and the compiler doesn't allow me to. I used serde_bytes to try and make it as efficient as possible, instead of just sending an array of numbers.

[amrbashir]

That clone bound shouldn't be needed, I will open a PR to fix it later.
Edit: here it is tauri-apps/tauri#12876

For now, you can use InvokeResponseBody::Raw which is Clone and implements IpcResponse trait and so can be used with channels.

[adrieljss]

That clone bound shouldn't be needed, I will open a PR to fix it later.

For now, you can use InvokeResponseBody::Raw which is Clone and implements IpcResponse trait and so can be used with channels.

done! 👍

[amrbashir]

let's remove, no need to keep it around anymore

[amrbashir]

this should be an ArrayBuffer iirc, and sometimes in the case where IPC fails to use the borwser fetch, it fallsback to Json serialization so this would be Array<number>

So we need to do some checks here, similar as what was done before

[adrieljss]

Yes, you're right. I'll update that.

[adrieljss]

Happy to say the stream that was implemented worked:

Something is wrong when i do content processing (like .json(), .text(), etc) so i'm trying to figure it out.

[adrieljss]

@amrbashir I added a new commit, can you check? I also fixed the content conversion bug said above, by converting the ArrayBuffer | number[] to Uint8Array (the content reader/consumer on Response doesn't like non-Uint8Arrays). So far it works fine when I tested it.

[adrieljss]

I also removed ReqwestResponse completely acafdd8, it's never read on Rust and is not needed in guest-js.

[amrbashir]

LGTM, now you just need to run cargo fmt and pnpm format, and a change file in .changes directory and we should be good to go.

[adrieljss]

Done!

just a heads up,

Channel might not be supported on the browser (http://localhost:xxxx), but when ran as an application it was fine.

Maybe should just fallback to fetch when in a browser setting?

[amrbashir]

that's out of the plugin scope users will need to add check and fallback if they wish so.

[amrbashir]

looks like you need a final pnpm build to build the JS API

[DresAaron]

Really appreciate the great work! Any plans to release this feature soon?

[letscagefdn]

good job guys, when do you think this will be released ?

[lucasfernog]

@amrbashir i'll approve and merge this one, it works well
i think any further change should come after a change in tauri core too right?