Add support for hydrating portals

[marcusdarmstrong]

Do you want to request a feature or report a bug?

Probably bug, but arguably a feature request, I suppose.

What is the current behavior?

I've attempted my best effort at a fiddle that shows off the particular issue. Obviously server side rendering is impossible via JSFiddle, but the markup should be equivalent to having rendered Test into a div with id test-1 during server side render.

https://jsfiddle.net/y8o5n2zg/

As seen in the fiddle, an attempt to ReactDOM.hydrate() a portal results in:

Warning: Expected server HTML to contain a matching text node for "Hello World" in <div>.

Additionally, after failing to hydrate, React renders the component and appends it resulting in a duplicated section of DOM:

<div id="test-1">Hello WorldHello World</div>

What is the expected behavior?

In an ideal world, calling hydrate on a component that has portals would allow those DOM containers to hydrate into the components they were rendered with.

Which versions of React, and which browser / OS are affected by this issue? Did this work in previous versions of React?

I've only tested this in 16.4.1, but I've confirmed the behavior in Chrome and Firefox. Given that I'm really looking at an edge case here I doubt it worked previously.

Why I'm doing this edge-case-y nonsense:

We're currently using multiple React roots on our pages (as some portions of the pages are not rendered by React yet), most of which are server-side rendered. We'd like to be able to hydrate them into a single React root on page, so that we can share contexts between them without difficulty and without repeating those context components in memory (in some cases we can have a good number of roots on the page—20-30, perhaps?).

In searching, I found a few potentially related bugs (#12615, #10713, #11169), but it seemed like these really didn't line up with my (hopefully valid?) use case.

Thanks!

[MaxMotovilov]

Nearly identical use case here. For now, resolving with a kludge in the component rendered into a portal (render self with display:none, detect duplicate, delete, show self).

[gaearon]

When you hydrate, the initial render should match your server render. But portals are currently not supported on the server. Therefore, hydrating a portal doesn't make sense with current limitations.

I think you want to do something like:

state = {mounted: false};

componentDidMount() {
  this.setState({ mounted: true });
}

render() {
  return <div>{this.state.mounted && ReactDOM.createPortal(...)}</div>
}

Does that make sense? Same workaround as you need to use when your client render doesn't match.

[marcusdarmstrong]

Thanks for following up, Dan!

I can't speak for Max here, but in my use case, we have no intention of rendering portals on the server (obviously that concept makes very little sense without a real dom to portal into). We'd like to use a multi-root approach on the server, but then when it comes time to hydrate on the client, declare each of those server-side-rendered roots to a special root component that can hydrate each of those into a Portal on the client, so that we end up with a single React root in-browser.

In the case described by the Fiddle linked above, an initial render does match the server render (both want to have "Hello World" in the test div)—it's just that the mechanism on the server for creating the Portal divs is external to React (just as it is on the client in the case of a Portal).

That is: https://jsfiddle.net/7y3kcnbh/
Rendering these components on the client yields the same thing as my hypothetical server-side-rendered markup in the original fiddle: https://jsfiddle.net/y8o5n2zg/

Rendering the portal on mount doesn't really help here, because it misses out on the opportunity to hydrate the various server-side rendered components that are in their respective DOM roots.

[MaxMotovilov]

Well, my use case is somewhat more convoluted than that. It is not a true hydration: I am loading a "foreign" page (JSP, WordPress -- that sort of thing) that wants to instantiate multiple React sub-applications aware of each other. To smooth out the load experience (and let google see all of content), the page contains a copy of the initial DOM recorded from the browser (call it a poor man's SSR :-) ). Because the final decision on which sub-applications to instantiate at all belongs to the page, they implement a handshake protocol to build a single vDOM tree (and Redux store etc) and decide which of them will go into portals. Thus I am not really re-rendering the container node for the portal at all and have to stick to cleaning up siblings after render.

Edit: now that I read @marcusdarmstrong's comment above, I think our use cases are indeed very similar, except for the top-level embedding mechanism.

[gaearon]

In the case described by the Fiddle linked above, an initial render does match the server render (both want to have "Hello World" in the test div)

I see where you're coming from, I just explain why the current behavior isn't so much a bug but a missing feature. From React's point of view, the initial render does not match the server render because portals are not supported on the server. Therefore, the portal encountered on the client is considered a new thing that needs to be inserted (rather than hydrated).

I agree that hydrating portals could be useful even before React SSR supports it.

[MaxMotovilov]

What if createPortal() could be explicitly told that yes, portal element does already contain pre-rendered DOM we need to diff against? Isn't this sort of what hydrate() does now -- passes in a flag that overrides the check for existing copy of the DOM? I understand it would be a bit of a kludge -- and place the responsibility squarely on user's shoulders as yet another __dangerously__ feature -- but probably much easier to implement than server-side portals in their entirety?

[gaearon]

I think if somebody implements this we can take a look at the PR. It's not a priority for us because:

• it will likely increase the code size
• the current SSR implementation doesn’t support it so this adds more disparity
• you can work around it with multiple hydrate calls (which wouldn’t preserve the context — but you wouldn’t have it on the server either)

There are plans for a different SSR implementation that would support “modern” features like error boundaries and Suspense. I think it would make sense to add full support for portals at the same time, with client and server parity.

But again, if somebody sends a PR we can take a look. Here's a few interesting places in code:

• Renderer-independent code that tracks hydration state: https://github.com/facebook/react/blob/21ac62c77ac68c66b2a1f0adcc60f7760b884a0e/packages/react-reconciler/src/ReactFiberHydrationContext.js
• Renderer-independent code that enters hydration code path:

  react/packages/react-reconciler/src/ReactFiberBeginWork.js

  Lines 454 to 458 in 21ac62c

  	if (
  	(current === null || current.child === null) &&
  	root.hydrate &&
  	enterHydrationState(workInProgress)
  	) {

• DOM-specific implementation of hydration traversal

  react/packages/react-dom/src/client/ReactDOMHostConfig.js

  Lines 401 to 495 in 21ac62c

  	export const supportsHydration = true;
  	export function canHydrateInstance(
  	instance: Instance | TextInstance,
  	type: string,
  	props: Props,
  	): null | Instance {
  	if (
  	instance.nodeType !== ELEMENT_NODE ||
  	type.toLowerCase() !== instance.nodeName.toLowerCase()
  	) {
  	return null;
  	}
  	// This has now been refined to an element node.
  	return ((instance: any): Instance);
  	}
  	export function canHydrateTextInstance(
  	instance: Instance | TextInstance,
  	text: string,
  	): null | TextInstance {
  	if (text === '' || instance.nodeType !== TEXT_NODE) {
  	// Empty strings are not parsed by HTML so there won't be a correct match here.
  	return null;
  	}
  	// This has now been refined to a text node.
  	return ((instance: any): TextInstance);
  	}
  	export function getNextHydratableSibling(
  	instance: Instance | TextInstance,
  	): null | Instance | TextInstance {
  	let node = instance.nextSibling;
  	// Skip non-hydratable nodes.
  	while (
  	node &&
  	node.nodeType !== ELEMENT_NODE &&
  	node.nodeType !== TEXT_NODE
  	) {
  	node = node.nextSibling;
  	}
  	return (node: any);
  	}
  	export function getFirstHydratableChild(
  	parentInstance: Container | Instance,
  	): null | Instance | TextInstance {
  	let next = parentInstance.firstChild;
  	// Skip non-hydratable nodes.
  	while (
  	next &&
  	next.nodeType !== ELEMENT_NODE &&
  	next.nodeType !== TEXT_NODE
  	) {
  	next = next.nextSibling;
  	}
  	return (next: any);
  	}
  	export function hydrateInstance(
  	instance: Instance,
  	type: string,
  	props: Props,
  	rootContainerInstance: Container,
  	hostContext: HostContext,
  	internalInstanceHandle: Object,
  	): null | Array<mixed> {
  	precacheFiberNode(internalInstanceHandle, instance);
  	// TODO: Possibly defer this until the commit phase where all the events
  	// get attached.
  	updateFiberProps(instance, props);
  	let parentNamespace: string;
  	if (__DEV__) {
  	const hostContextDev = ((hostContext: any): HostContextDev);
  	parentNamespace = hostContextDev.namespace;
  	} else {
  	parentNamespace = ((hostContext: any): HostContextProd);
  	}
  	return diffHydratedProperties(
  	instance,
  	type,
  	props,
  	parentNamespace,
  	rootContainerInstance,
  	);
  	}
  	export function hydrateTextInstance(
  	textInstance: TextInstance,
  	text: string,
  	internalInstanceHandle: Object,
  	): boolean {
  	precacheFiberNode(internalInstanceHandle, textInstance);
  	return diffHydratedText(textInstance, text);
  	}

[MaxMotovilov]

@gaearon - fair enough, and thanks a lot for the pointers! I will probably stick with the simple kludge I have for now, unless synchronous replacement of the tree (as opposed to true hydration) proves to be an issue.

[gaearon]

You can still hydrate if you don't use portals and instead perform several hydrate calls when mounting the app.

[marcusdarmstrong]

Thanks for the hints! I suspected this might be more of a "by design" type thing. I'll go ahead and add all this to our internal ticket on the subject. It's possible somebody on my team might go ahead and take a look at this. The biggest motivations for us here are that we can use local state rather than module state for our context objects, so it's really an optimization more than anything else, so we'll see how the prioritization goes.

[MaxMotovilov]

You can still hydrate if you don't use portals and instead perform several hydrate calls when mounting the app.

I guess I don't quite understand what will I get in this case. Wouldn't this result in multiple independent vDOMs instead of a single common one I build now? Clearly the contexts will be independent as well, so no common instances of <Provider>, <BrowserRouter> and such; every sub-app would have to be wrapped separately and proper sharing of global resources (location, Redux store et. al.) can only be ensured by these wrapper components. Sounds a bit scary as I can't be sure offhand if this use case is supported by all of the service libs currently in use...

[marcusdarmstrong]

@MaxMotovilov That's what we're doing, for what it's worth. Our on-page "runtime" handles wrapping everything in Providers and whatnot all pointing to the same store instance before hydrating.

[gaearon]

Clearly the contexts will be independent as well, so no common instances of , and such

Right, but don't you have the same problem on the server anyway? Since SSR doesn't support portals.

[marcusdarmstrong]

It's generally less of a problem on the server because (at least our) contexts don't mutate on the server.