Fix generators, use text/event-stream protocol

[remorses]

Fix #742
Fix #741

[remorses]

What do you think @SaltyAom is this a breaking change or a bug fix? Previously the response was not following the text/event-stream protocol

[luismanfroni]

Don't know if it's related, but having a better way to catch Aborts would be really great, instead of having to do something like
request.signal.addEventListener("abort", cb) (which is kinda ugly) inside all request generators handlers for cleanups.
Would be cool an "abort" hook, or something like that.

[remorses]

@SaltyAom do you need any help with the review?

I am already using this fork in production and it works great. Another problem of the current approach is that most servers won't flush the body until they find a new line, using the right text/event-stream protocol fixes that

[SaltyAom]

Hey, thanks a lot for your PR.
Would be great if you can help me with these changes and elysiajs/eden#114, after that let's get this merge.

[SaltyAom]

You can remove try catch in here, handleStream function is already wrapped in try-catch that handle onError already

[remorses]

Currently the app will crash if you throw an error in a generator before yield, i added a test that reproduces this here: https://github.com/remorses/elysia/blob/932951679c89aef6bad08779b6eb7fe94c30335d/test/response/stream.test.ts#L75

[remorses]

I fixed this crash removing the try catch when aot is false. In aot mode it still crashes.

the app crashes because the error is throws asynchronously while mapResponse is a not async, i think there are many more of these bugs hidden in the code, the ideal solution would be to always await mapResonse in aot too

[SaltyAom]

Would make sense to check typeof object before using JSON.stringify

[remorses]

If you don't always call JSON.stringify and the user yields a string that contains 2 new lines the response will break the text/event-stream protocol. If you always call JSON.stringify you can be sure that all the new lines are encoded with \n and put in a single line.

The parsing in Eden also becomes simpler because you know messages are always a JSON encoded string instead of using heuristics.

[kravetsone]

Привет, спасибо большое за твой PR. Было бы здорово, если бы ты помог мне с этими изменениями и elysiajs/eden#114 , после этого давайте сделаем это слияние.

Yay! Thanks for the work)

[HK-SHAO]

@SaltyAom @remorses In fact I don't think it should be up to elysia stream to handle exactly how to send the streams of the SSE interface, that should be left to the developer. It should be left to the developer, because exactly how to send each id, event, etc. should be set by the developer.

For example I would like to handle the sent id by itself, it contains some of my rules, for example the id contains a timestamp, server id and a pointer. Or I want messages sent for different events to have different event names. Another example is that I may not need the id and event because ChatGPT's stream response only returns data and does not contain the id and event. there is also a case where I only want to send a single comment. In my opinion, all the above scenarios mean that the generator should only be responsible for sending the data as is, and not have to convert it into an SSE response.

[remorses]

@HK-SHAO if you want customization you will need to implement it from scratch

app.get('/events', async ({ set }) => {
  set.headers['Content-Type'] = 'text/event-stream'
  set.headers['Cache-Control'] = 'no-cache'

  const stream = new ReadableStream({
    async start(controller) {
      for (let count = 0; count < 10; count++) {
        const event = `data: Event ${count}\n\n`
        controller.enqueue(event)
        await Bun.sleep(1000)
      }
      controller.close()
    }
  })

  return new Response(stream)
})

[kravetsone]

@HK-SHAO if you want customization you will need to implement it from scratch

app.get('/events', async ({ set }) => {
  set.headers['Content-Type'] = 'text/event-stream'
  set.headers['Cache-Control'] = 'no-cache'

  const stream = new ReadableStream({
    async start(controller) {
      for (let count = 0; count < 10; count++) {
        const event = `data: Event ${count}\n\n`
        controller.enqueue(event)
        await Bun.sleep(1000)
      }
      controller.close()
    }
  })

  return new Response(stream)
})

Yeah i agree. Elysia should folows specs but if you need sometning custom you can do it

[remorses]

I created a full framework where async generators have first class support and use SSE, it's called Spiceflow

[SaltyAom]

I think I have to added input to this for the record

The problem is there are both use case for both text/event-stream and non text/event-stream
eg. stream a response in which response is not a text eg. image, video or create a proxy from AI provider like OpenAI, Claude, etc which are using direct streaming not text/event-stream, or create a proxy from fetch (Elysia can return fetch(url) and it will just become a proxy which is possible because of direct streaming)
I personally haven't come up with an idiomatic API yet which is why I leave it as it until then because I'm afraid of introducing a breaking changes.