whatwg
A way to stream content into an element
Use-case: I work on a news site and I want to create visual transitions between articles, but I don't want to lose the benefits of streaming. So:
- User clicks on link
- Create a new article element to contain the incoming article
- Start fetching content
- Pipe content into new article element
- Begin the visual transition once elements appear in the new article element
Not only is innerHTML is slower way to do this (due to a lack of streaming), it also introduces a number of behavioural differences. It'd be great to try to limit these, eg allow inline scripts to execute before additional elements are inserted.
Parts of this can be hacked using document.write and <iframe>, but it'd be good to have a non-hacky way.
Code for the above use-case if element.writable provided a way to stream HTML into the element.
const article = document.createElement('article');
const response = await fetch('article.include');
const articleHasContent = new Promise(resolve => {
const observer = new MutationObserver(() => {
observer.disconnect();
resolve();
});
observer.observe(article, {childList: true});
});
response.body
.pipeThrough(new TextDecoder())
.pipeTo(article.writable);
await articleHasContent;
performTransition();
There is also https://w3c.github.io/DOM-Parsing/#idl-def-range-createcontextualfragment(domstring) which does execute scripts (when inserted into a document). As a possible alternative to document.write in the meantime...
No, as an alternative to document.write in your hack. Don't even need an iframe, just a Range instance.
http://software.hixie.ch/utilities/js/live-dom-viewer/saved/4716
@zcorpan ah, that doesn't allow partial trees http://software.hixie.ch/utilities/js/live-dom-viewer/?saved=4717, which you kinda need if you're trying to stream something like an article.
Hello! I'm curious about the possibility to insert scripts.
I'm trying to execute an script and I see the script is executed after insertion.
<!DOCTYPE html>
<body>
<script>
var r = new Range();
// Write some more content - this should be done async:
document.body.appendChild(r.createContextualFragment('<p>hello'));
document.body.appendChild(r.createContextualFragment(' world</p>'));
document.body.appendChild(r.createContextualFragment('<script>console.log("yeap");<\/script>'));
// done!!
</script>
</body>
I've two questions: What about if the inserted script has the flag... async or defer? Those flags will not have effect.
Will you keep this function in the standard? Or are you planning to remove it?
Finally I found (you gave me) the best way to insert content to the document.Thanks a lot!
@rianby64 note that the example above creates two paragraphs rather than one.
What about if the inserted script has the flag... async or defer? Those flags will not have effect.
The scripts will be async, as if you'd created them with document.createElement('script'). For the streaming solution I mentioned in the OP, I'd like the parser to queue DOM modifications while a non-async/defer script downloads and executes, but allow something like a look-ahead parser.
Will you keep this function in the standars?
Which function? createContextualFragment? I don't see why it'd be removed.
In general I think we want to be able to provide ReadableStream or Response objects to APIs that currently take a URL. @jakearchibald, would something that let you assign a ReadableStream or Response (backed by a stream) to an iframe.src satisfy your use case?
The key part here is to not have to use a separate iframe plus adoption of the current parser insertion point into a different document. Instead, we just want to parse into an existing document location.
This means creating an element that has the concept of partially loaded state, right? An iframe already has all of that, but do other html container elements? So wouldn't we need to create something that has all the load event, error event, and other stateful information of an iframe? Or maybe all that exists today. HTML always catches me out.
@wanderview
This means creating an element that has the concept of partially loaded state, right?
I think we can get away without this. If an element has a writable endpoint you'll get locking on that writable for free. However, during streaming you'll be able to modify the children of the element, even set the element's innerHTML. The HTML parser already has to deal with this during page load, so I don't think we need to do anything different.
So wouldn't we need to create something that has all the load event, error event, and other stateful information of an iframe?
We probably don't need this either. htmlStream.pipeTo(div.writable) - since pipeTo already returns a promise you can use that for success/failure info.
How would this interact / compare with the following scenario:
Rather than fetching HTML snippets from the server, I'm much more likely to be able to fetch [we'll assume newline-delimited to enable stream parsing] a minimal JSON encoding of whatever entity I'm trying to display.
Partially, this is just down to the fact that most web servers wrap HTML output in a series of filters, one of which is a base "<html><head/><body>...</body></html>" template. Obviously, that can [easily, from a purely technical perspective] change, but spitting out independent <div>s is going to take some cultural change on the server side. JSON we have today, and won't take any convincing.
... so, assuming that we use JSON, Is there a performance win to being able to render JSON snippets [as they come over the network] to HTML? The trade-off I'd assume we're making is on triggering additional layouts; put another way, is it faster to do:
- [ JSON blob representing n divs ]: JSON -> HTML -> DOM in one step -or-
- [ streaming JSON blob representing n divs ]:
while (nextJSONitem) { JSON -> HTML -> DOM }
My expectation is that the answer is "it depends"; I don't have a sufficiently reliable playground for testing this to any degree of accuracy, but I would expect we'd want to keep the render pipeline as unobtrusive as possible while minimizing network->screen latency for individual items, using the following as trade-offs:
-
time to render & re-compute layout for:
- each items in an array individually vs
- all items in an array as a single DOM manipulation vs
- batches of items
-
total time to:
- fetch all items
- pull a single item out of a streaming JSON blob
... ideally all while minimizing client complexity ("they wrote a lot of code to make things that slow"). Thankfully that part should be hidden in frameworks.
... OR am I totally barking up the wrong tree with the idea that JSON is the right delivery mechanism, and we should aim to generate server-side HTML snippets for pretty much anything that can be fetched-with-latency?
@blaine I think I cover what you're asking over at https://jakearchibald.com/2016/fun-hacks-faster-content/
@jakearchibald It kind of feels like there should be a way for code other than the one writing to the element to know if it's complete. The pipeTo promise, while useful, does not seem adequate for that.
For example, code that uses a query selector to get an element and operate on it should have some way to know if the element is in a good state. Seems like that kind of code is usually pretty independent.
response.body
.pipeThrough(new TextDecoder())
.pipeTo(article.writable);
Would indeed be a big win! ❤️
@jakearchibald durr. I'd read that a few days ago and forgotten the second part of your post in this context. Sorry, I blame lack of coffee. ;-)
Re-reading this more carefully, the element.writable pipe makes a ton of sense, and it'd be trivial for a rendering pipeline to make use of it, even in the JSON case. +1
Wait, how would the element.writable getter even work, since a WritableStream usually (bar explicitly passing 'preventClose') can only be pipeTo'd once, after which it becomes closed and can't be written to again?
htmlStream.pipeTo(div.writable).then(() => htmlStream2.pipeTo(div.writable) /* cancels source stream and does nothing? */);
What happens when it's already locked to a previous, still incomplete, still streaming request but you changed your mind/ the user clicked to the next article already?
htmlStream.pipeTo(div.writable); // locked
htmlStream2.pipeTo(div.writable); // doesn't work, stuck waiting?
Would it have to produce a new fresh WritableStream on every access? Then every access would have to instantly invalidate all the previous writable streams so that writing to them does nothing, and only the latest effects the element's contents?
@jakearchibald I'm curious how you respond to @isonmad's comment; it seems like a valid argument against a WritableStream here. And of course the lack of cancelable promises is hurting us here...
Yeah, this seems like a good argument against element.writable and for something like:
htmlStream.pipeTo(div.getWritable());
or
const domStreamer = new DOMStreamer();
div.appendChild(domStreamer);
htmlStream.pipeTo(domStreamer.writable);
What happens when it's already locked to a previous, still incomplete, still streaming request but you changed your mind/ the user clicked to the next article already?
This could be done with domStreamer.abort() or somesuch, but maybe it's a more general problem to solve - how to abort a pipe.
Would it have to produce a new fresh WritableStream on every access? Then every access would have to instantly invalidate all the previous writable streams so that writing to them does nothing, and only the latest effects the element's contents?
Taking the above models, would it be bad to allow two streams to operate within the same element? Sure you could get interleaving, but that's already true with two bits of code calling appendChild asynchronously.
The browser already has to cope with the html parser and appendChild operating on the same element, so it doesn't feel like anything new.
We have discussed this at the last WHATNOT (though not referring to this issue directly).
There is some openness to streaming into an element, either by making the element writable or with something like setHTMLUnsafe(ReadableStream). The main dependency would be to deliver this together with a solution for sanitation, e.g. letting TrustedTypes provide a TransformStream or some such.
See also #11542 for streaming into an element as part of the initial response, which doesn't have the same security traits.
Looking at what it would take to make this proposal aligned with trusted types and sanitation:
- Support streaming in the HTML sanitizer, by letting it receive a stream, buffering internally along the way to ensure element start tags are parsed correctly etc. or whatever the policy requires. This is possibly internal to the sanitizer implementation?
- Enable optional streaming support in trusted types, by allowing the policy to provide a
TransformStreamthat performs custom sanitation. - The
TrustedTypesversion should return something like aTrustedHTMLStreamwhich can also be passed tosetHTMLUnsafe. - Both
setHTMLandsetHTMLUnsafe(or variants with a different name) should accept aReadableStream.
#11669 and #11542 suggest how streaming would work in terms of API, but doesn't define the streaming lifecycle. Since the streaming methods are async, they need to account for conflicts or mismatches. e.g. if multiple streams are applied to the same element, or if a reference node moved during the application of the stream.
To make this behavior consistent across all the methods, suggesting something along the following:
- The stream application starts when the stream is has both a reader and a writer.
- At that point, the different methods are translated to a list of items to remove, a parent node, and a reference node (or
null). The parent+reference node are equivalent to theinsertBeforearguments. - If there is an existing stream with the same parent and reference node, it is aborted.
- The items to be removed are removed at that point.
- The stream feeds the parser. Chunks are inserted into the parent node, before the reference node. If the reference node is not null and is no longer a child of the parent node, the stream is terminated with an error.
The above:
- Produces reasonable results in the happy path
- Allows safe concurrent streaming to multiple locations, including into the same element
- Provides consistent behavior regardless of the method used for streaming.
In https://github.com/whatwg/html/issues/11964 we had consensus to advance this proposal to stage 2, with the specific scope of adding streamHTMLUnsafe() and adding the runScripts option to both streamHTMLUnsafe() and setHTMLUnsafe().
