Implement streaming in the browser

[ariutta]

The current use of chunked transfer encoding is based on xhr.responseText. There are now more efficient methods available for Firefox (moz-chunked-text) and Chrome (ReadableByteStream). Would there be any interest in using those methods when possible?

It would be awesome if superagent supported streaming in the browser along the lines of what this library offers. I'd be willing to contribute if there's any interest.

[kornelski]

I think that would be a good addition.

We currently don't have an API to expose this functionality. The browser version doesn't support streaming at all. OTOH in Node we have .pipe() and node's streams.

Since ReadableByteStream is the standard, I think it should be used in the browser version. Is it possible to emulate ReadableByteStream in browsers that don't support it natively yet?

I wonder how to have first-class support for node's streams, and first-class support for the Streams Standard without confusing and duplicated APIs. Do you have any proposal for the API?

[ariutta]

That's a tough question. I love superagent's clean API and minimal file size. And I also support building on standards. But I'm not sure how to join the two.

Emulating/polyfilling Node streams or ReadableByteStream definitely is possible, but wouldn't this option eliminate superagent's file size advantage?

This issue for stream-http has a comment suggesting it would be possible to create a minimal stream library to replace the Node stream dependency to reduce file size. But, yes, creating yet another stream library would pollute superagent's clean API.

[ariutta]

I ran some quick tests on the minified, bundled sizes of the latest versions of related modules for comparison:

module	size
axios	16K
chunked-request	7.2K
fetch	435K
fetch-stream	91K
whatwg-fetch	7.1K
grab-http	8.1K
hyperquest	134K
nets	32K
request	1.0M
requests	12K
simple-get	90K
stream-http	87K
superagent	13K
xhr	6.1K
xhr-request	9.8K

Here was the bundle command:

browserify -g uglifyify ./size.js --standalone size | uglifyjs -c -o ./size.bundle.js

[ariutta]

Maybe the best option would be to minimally polyfill the stream standard with a notice that the full stream standard will be used when it lands in most major browsers. The minimal polyfill would be limited to event emitters (data, error, end) and basic functions (.write, .end). Polyfilling the entire stream standard, including backpressure, etc., would drastically increase the size of the library.

[kornelski]

Yes, I think a minimal polyfill is a good way to go. With XHR we don't even have ability to do proper backrpessure.

[ariutta]

Hmm... WHAT-WG streams actually remove data events and also split the Node pipe method into pipeTo(writable) and pipeThrough(transform). I'm hesitant to try creating a minimal polyfill of the streams standard when it's not yet in production anywhere (didn't see it on icanuse.com) and is significantly different from Node streams.

[ariutta]

Also, there is the observable standard, which is probably less suited for the purpose of this library, but it does offer yet another way of working with multiple chunks over time. If for some reason the observable standard gets traction and the stream standard doesn't, then it would make sense to use the observable standard here.

[ariutta]

But if we continue with the web streams standard, the web-streams-polyfill could serve as a start. It basically just turns the reference implementation into an npm package. I got it down to a minified size of 47K by using uglifyify and unassert. To reduce size further, we could look at dropping code from readable-stream.js.

[kornelski]

If the polyfill is heavy I see a couple of possible workarounds:

• don't bundle the polyfill with superagent, and inform users that they should polyfill the global ReadableByteStream (and other objects) if they want backwards-compatibility. Users could choose which implementation they want.
  or
• in superagent provide only hooks for plugins for streaming, and leave streams implementation up to plugins.

[ariutta]

I like the second idea. Would the hooks be event emitters and .write/.end methods? If so, pull-through and pull-stream might be useful. 9.8K bundled together and minified.

In this issue, there's some interesting discussion on lightweight transforms for web streams. Would the hooks be easy to use with transducers?

[kornelski]

Currently superagent, in node version, has a rather ad-hoc mechanism for "parsers"

https://github.com/visionmedia/superagent/blob/master/lib/node/parsers/text.js

which boils down to "data" and "end" events. Is that equivalent of .write/.end methods you've had in mind?

[ariutta]

Yes, that sounds perfect.

[XeniaSiskaki]

So, currently if the server is sending the response in chunks, there is no way to receive it accordingly on the browser?

[kornelski]

Depends what you mean by chunks. If you mean chunked HTTP/1.1, then you can, but superagent will wait until the whole response is received.

[XeniaSiskaki]

Well, on the server I subscribe to a stream which is emitting data when they are available (that's what I mean by chunks). But because sending the whole response takes long (fetching one year's data), the client is always throwing a ERR_INCOMPLETE_CHUNKED_ENCODING, even though the server didn't finish sending the data.

[kornelski]

Streaming in the browser will not solve your problem, as the error is due to your web server cutting the connection server-side.

[XeniaSiskaki]

Is there any way to verify that? The server seems to be sending data normally, without any errors.

[kornelski]

Try using raw XmlHTTPRequest or curl to verify whether the server sends data fully.

[XeniaSiskaki]

I tried using an XmlHTTPRequest but still get the error. Even after getting the error on the browser, the server continues to send data and at some point ends (res.end()) with no error. So I'm not sure if the problem is on the server side.

[kornelski]

If you get error in XmlHTTPRequest then there's nothing we can do. The browser would report the same error via streaming API too, and stop streaming early.

It's most likely that it's a problem outside of browser, e.g. in your CDN, your reverse proxy, your web server, or middleware managing node.

[kornelski]

fetch() is a thing now and it works well with streaming, so if you need streams, use fetch().