facebook / react

The library for web and native user interfaces.
https://react.dev
MIT License
230.17k stars 48.16k forks source link

Bypass synthetic event system for Web Component events #7901

Open staltz opened 8 years ago

staltz commented 8 years ago

To use a Web Component in React, you must directly attach an event listener to the React ref for that WC. We could change the implementation so that when React detects a custom element (tag names with dashes, like my-component), it will bypass the synthetic event system (and the whitelist) and just attach the event listener on the element itself.

Why bypass the synthetic event system? Because anyway we already need to bypass it manually when using a Web Component. I'm not so familiar with the React codebase, but this naive approach seems to work. Whoever uses Web Components in React can be responsible for whatever downsides that would cause, maybe in performance, I don't know. They are already having those (supposed) downsides, this issue is just about the convenience of WC usage inside React.

I was about to send a PR for this, but thought of opening an issue. I looked through the issues and didn't see any existing one related to the handling of WC events.

What is the current behavior?

A WC custom event (e.g. flipend) must be handled by attaching the event listener directly to the element in componentDidMount using a ref.

http://jsbin.com/yutocopasu/1/edit?js,output

React v15.1.0

class HelloMessage extends React.Component {
  handleHelloClick() {
    this.refs['foo'].toggle();
  }

  handleFlipend(ev) {
    console.log('Handle flip end');
  }

  componentDidMount() {
    this.refs['foo'].addEventListener('flipend', ev =>
      this.handleFlipend(ev);
    );
  }

  render() {
    return (
      <div>
        <div onClick={ev => this.handleHelloClick()}>
          Hello {this.props.name}, click me!
        </div>
        <brick-flipbox class="demo" ref="foo">
          <div>front</div>
          <div>back</div>
        </brick-flipbox>
      </div>
    );
  }
}

What is the expected behavior?

A WC custom event can be handled with onMyEvent={ev => this.handleMyEvent(ev)} on the ReactElement corresponding to the WC.

class HelloMessage extends React.Component {
  handleHelloClick() {
    this.refs['foo'].toggle();
  }

  handleFlipend(ev) {
    console.log('Handle flip end');
  }

  render() {
    return (
      <div>
        <div onClick={ev => this.handleHelloClick()}>
          Hello {this.props.name}, click me!
        </div>
        <brick-flipbox onFlipend={ev => this.handleFlipend(ev)} class="demo" ref="foo">
          <div>front</div>
          <div>back</div>
        </brick-flipbox>
      </div>
    );
  }
}

PS: this snippet above still has the ref, but for unrelated reasons. Ideally we wouldn't need refs for handling events of WCs.

treshugart commented 8 years ago

https://github.com/webcomponents/react-integration will create a React component from a web component constructor and give you custom event support.

sebmarkbage commented 8 years ago

As long as we only support attributes. I don't see a problem doing this for the heuristic if (typeof props[propName] === 'function') element.addEventListener(propName, props[propName], false). The only concern would be if we should have some heuristic for normalizing the event name. I'm not really a fan of converting things like onXxx into xxx after doing that in MooTools. The other concern is related to #6436 and how we'd handle capture/bubble/passive/active etc.

I'd like it better if we could just pass through all props to element properties but it seems like that ship has sailed since most web components aren't designed to handle properties properly. A massive loss to the community IMO.

treshugart commented 8 years ago

I'd like it better if we could just pass through all props to element properties but it seems like that ship has sailed since most web components aren't designed to handle properties properly. A massive loss to the community IMO.

It feels like that statement is clumping all web components into a single bag of poor design, which in many cases they are, but it doesn't mean you can't optimise for the ones that are designed well. A couple paradigms I'm trying to push in the community (and with SkateJS) are:

Monica Dinculescu mentioned the former and Rob Dodson the latter in their Polymer Summit talks, so I think that's something they're trying to espouse. It's unfortunate the primitives don't make this more obvious, but I think that comes with the nature of most browser built-ins these days.

React not setting props, and not supporting custom events, is the reason we've had to maintain that React integration library I posted above (https://github.com/webcomponents/react-integration). It's worth looking at as a source of some patterns that are definitely working for us in production. It's also worth noting that the patterns employed there are also used in Skate's wrapper around Incremental DOM. We set props for everything we can, falling back to attributes as a last resort.

In the integration lib, events have their own special convention, similar to React's. This in particular is something to pay attention to because adding event listeners if prop is a function - without checking the name for something like /^on[A-Z].*/ - will prefer events over properties.You could flip that around and set props if prop in element and leave it up to the web component to addEventListener() when the onClick property is set. This creates a little more work for the component, but means you don't have to do the prefix checking if you're not a fan.

sebmarkbage commented 8 years ago

@treshugart Most people seem to find this counter-intuitive but my preference would be to effectively just call Object.assign(element, props). It sounds like you'd be a fan of that approach as well? Maybe there is still hope to change the Web Components ecosystem to prefer that - and indeed React promoting that style would perhaps help.

@staltz What do you think about that approach?

treshugart commented 8 years ago

@sebmarkbage I like that approach but I think there are a few things to consider (and you're probably already aware):

  1. Props don't work for all elements. SVGElement will error if you set width as a property, for example because it's readonly.
  2. Attributes don't work for all elements. HTMLInputElement needs value set as a property, for example because the attribute does not 2-way sync once the property is set or a user sets a value.
  3. Some web components may have edge-cases where an attribute-only handler is necessary (aria-*, maybe).
  4. Attributes with dashes (my-property) reflected to camel-cased props (myProperty). What is the behaviour here? Probably fine to not worry about this case because the consumer can just use <x-element myProperty={something} />.
  5. Attributes with the same name as the property. What happens if the property is only a getter (no set() or writable: false)? There's no way for the consumer to prefer the attribute. Maybe this is something that is enforced; that the component author must provide a setter or the consumer has to <x-element ref={e => e.setAttribute('some', 'value')} />.
  6. Would this behaviour be applied to custom elements only and how would this be detected if so?

Something like the following might be a little bit more robust:

Object.keys(props).forEach(name => {
  if (name in element) {
    // You might also need to ensure that it's a custom element because of point 1.
    element[name] = props[name];
  } else {
    doWhatReactNormallyDoesWithAttributes();
  }
});

Brainstorming just in case. Overall, I tend to agree with you that it'd be fine for React to just do Object.assign(). It seems like there are ways for the consumer to get around any potential design problems with the component they'd be using.

Maybe there is still hope to change the Web Components ecosystem to prefer that - and indeed React promoting that style would perhaps help.

I sure hope so. I'm trying to and I know the Polymer team is trying to, as well. I think if React did do this, that it'd be a massive help, too.

justinfagnani commented 8 years ago

I'd like it better if we could just pass through all props to element properties but it seems like that ship has sailed since most web components aren't designed to handle properties properly. A massive loss to the community IMO.

I'm not sure what exactly you're referring to, but with Polymer we'd definitely prefer setting properties to setting attributes. Which web components don't handle properties and in what way?

robdodson commented 8 years ago

@sebmarkbage I think we can definitely encourage folks to write components that support the Object.assign approach you mentioned above. As @treshugart mentioned, I spoke about doing this in my Polymer Summit talk based on our previous twitter discussion. Having easy event support in React would also be great.

robdodson commented 8 years ago

Talk is here: https://www.youtube.com/shared?ci=256yfQG-abU

I also mentioned the need to dispatch events for state changes so libraries like React can revert those changes (similar to the way React handles the native input checkbook element)

nhunzaker commented 8 years ago

events up, props down

Yes. I think there would be benefits outside of web components too. I wonder if we'd see libraries rely less on the context API when they could just emit a custom event.

I don't see a problem doing this for the heuristic if (typeof props[propName] === 'function') element.addEventListener(propName, props[propName], false).

@sebmarkbage Unless you know if any current work here, I'd be happen to stencil something out.

sebmarkbage commented 8 years ago

@nhunzaker Well I think that what is being discussed here would be an alternate strategy.

We would not support addEventListener('click', fn). We would instead support element.onclick = fn;

staltz commented 8 years ago

Yeah I think I'm ok with the idea of Object.assign(element, props) and element.onflipend = fn. Just to understand why you proposed this, is it just to avoid translating onFoo to foo? Or are there more reasons?

sebmarkbage commented 8 years ago

A problem with properties, attributes, events and children is that you have to know which one to use so you end up with heuristics or explicit namespaces. I don't think explicit namespaces is very ergonomic.

Properties are more powerful than attributes because:

So if we only had one, I'd prefer it to be properties.

That leaves events. If we use a heuristic, that affects performance negatively since we have to create mappings for it at runtime. It also means that we're claiming a whole namespace. It means that custom elements can't provide an onFoo property if they wanted to. For example, with infinite scrolls there's a pattern like onCreateRow that isn't just a simple a event but a callback that's invoked to create a row for a particular index.

I think that would be unfortunate to claim a whole namespace prefix or type when the precedence of element.onclick = fn is a much simpler model.

Of course, that will still leave us "children" as special but that's kind of a unique property of React that children of components can be any type of value so we'll have to concede that one constrained special case.

robdodson commented 8 years ago

I took a stab at implementing the model discussed above. @sebmarkbage can you let me know if this matches your thinking?

class XCheckbox extends HTMLElement {
  connectedCallback() {
    this.addEventListener('click', this._onclick);
  }
  disconnectedCallback() {
    this.removeEventListener('click', this._onclick);
  }
  _onclick(e) {
    this.checked = !this.checked;
    this.dispatchEvent(new CustomEvent('checkchanged', {
      detail: { checked: this.checked }, bubbles: false
    }));
  }
  set oncheckchanged(fn) {
    this.removeEventListener('checkchanged', this._oncheckchanged);
    this._oncheckchanged = fn;
    this.addEventListener('checkchanged', this._oncheckchanged);
  }
  get oncheckchanged() {
    return this._oncheckchanged;
  }
  set checked(value) {
    this._checked = value;
    value ? this.setAttribute('checked', '') : this.removeAttribute('checked');
  }
  get checked() {
    return this._checked;
  }
}
customElements.define('x-checkbox', XCheckbox);

const props = {
  checked: true,
  oncheckchanged: function(e) {
    console.log('oncheckchanged called with', e);
  }
};
const customCheckbox = document.createElement('x-checkbox');
Object.assign(customCheckbox, props);
document.body.appendChild(customCheckbox);

One concern is that element authors have to opt-in to defining a setter to expose a handler for every event that they dispatch. That may end up bloating the elements, especially if they have a variety of events that they expose. Having React do element.addEventListener(propName, props[propName], false) might make element author's lives a bit easier. Just speaking personally, knowing the trade off between an event name heuristic and having to define setters for everything, I'd choose the heuristic.

justinfagnani commented 8 years ago

@sebmarkbage I get why properties are preferable to attributes, that's why Polymer defaults to setting properties. Since most Web Components these days are Polymer elements, and Polymer automatically supports properties, I think most Web Components handle properties correctly. Most of the other WC libraries I've seen handle properties correctly as well. If you know of a large set of components that don't support properties, let me know and I'd be glad to we see if I can help fix that massive loss.

@robdodson I don't think it's really feasible to have element authors write their own event handler properties. Events work fine with just addEventListener() and dispatchEvent() and event handler properties are an extra burden and won't be consistently implemented, and aren't even sufficient for many uses of addEventListener(). I'd guess that if addEventListener() were in DOM 0, the event handler properties wouldn't even exist - they're a vestige of an earlier time.

Polymer and Angular (and I believe SkateJS as well) have syntax conventions for declaring a binding to a property, attribute or adding an event handler with HTML attribute names. I don't know JSX, but it seems like since it's not HTML and not JavaScript, there's a lot of leeway to invent its own syntax to unambiguously differentiate between properties attributes and events.

sebmarkbage commented 8 years ago

React intentionally went back on that and claims that single event handlers is a better model and the use cases for multiple event listeners is better solved elsewhere since it leads to confusion about where events flow and which order they flow.

Additionally, the string based event system is difficult to type. Both statically for type systems like TypeScript and Flow, and for optimizing adding lots of event subscriptions at runtime (instead of string based hash maps).

So I don't think it's fair to say that event handler properties are strictly worse.

More over, there are other types of first-class event handlers like Observables that would be nice to support using properties.

The most important feature for interop is reflection. Without reflection you can't make automatic wrappers such as providing a first-class Observables as properties for each available event.

This is something that the event listener system doesn't provide. There is no canHandleEvent(type). Of course, just like attributes, this doesn't really make sense since the stringly typed event system can broadcast any string to any element through bubbling.

robdodson commented 8 years ago

@sebmarkbage those are good points. If you have a moment can you take a look at the sample code I posted and let me know if it seems inline with what you're thinking?

sebmarkbage commented 8 years ago

@robdodson Yes, that looks very good to me.

sebmarkbage commented 8 years ago

If the only problem is boilerplate, I think that is solvable. Initially in user space in terms of libraries that make it easy to create best-practice custom elements. Later a standard helper can be batteries included.

robdodson commented 8 years ago

If the only problem is boilerplate, I think that is solvable. Initially in user space in terms of libraries that make it easy to create best-practice custom elements. Later a standard helper can be batteries included.

Yeah that sounds good to me 👍 @treshugart @staltz what do you guys think?

treshugart commented 8 years ago

Looks awesome! I'll reiterate that https://github.com/webcomponents/react-integration currently solves this stuff in userland. That might be a good spot to start collaborating on some of the boilerplate.

pemrouz commented 8 years ago

From experience of using vanilla Web Components in a React-like architecture, a few humble suggestions:

staltz commented 8 years ago

@robdodson Yeah that's aligned with what I think as well.

Just speaking personally, knowing the trade off between an event name heuristic and having to define setters for everything, I'd choose the heuristic.

I'd also choose the same.

What's the action points with this issue? I'm willing to do something but unsure what, and confused whether @robdodson or @sebmarkbage have intentions to act on this.

sebmarkbage commented 8 years ago

To clarify my conclusion: If @robdodson shows that the proposed "best-practice" above is a feasible direction for the Web Components ecosystem including event listeners as properties.

Then React will switch to simply transferring props to the element properties, Object.assign(customElement, reactElement.props) in the next major release.

The next actionable item there would be that someone (maybe Polymer?) provides a way to make it easy to build such Web Components and see if it is viable in the more raw-form Web Components community (as opposed to big libraries).

robdodson commented 8 years ago

@staltz I've been discussing it with some of the platform engineers on Chrome. They raised some issues around how the native on* properties work with builtins and I think it'd be worthwhile to discuss those. I'm working on a follow up that details the points they covered.

robdodson commented 8 years ago

This raises the concern: does it make sense to promote a pattern for CEs that uses the existing on* semantics but doesn't have the same attribute / property parallelism?

I realize for most frameworks/libraries this is not a problem, because they’ll just use properties. But as a general pattern for CEs it’s troubling because it breaks with how all of the other HTML elements work, and could be surprising to anyone assuming that <x-foo onbar=”...”> would behave like other built-in tags.

There are a couple solutions the team proposed:

  1. Spec an API making it easy for CE authors to mimic the native behavior, so they could generate on* properties and attributes that would work exactly as the builtins.
  2. Pursue a properties-only approach, but don’t use the existing on* semantics. Basically use this as an opportunity to say the on* stuff is pretty weird/broken, and instead of encouraging it, let’s pioneer a different system in userland and feed that back to the spec.

I personally lean toward option 2, but am curious what other folks on this thread think?

Because both of the above options will take some time, that leaves the question of what approach should we move forward with in the near term. I think there are a couple options:

1) Similar to @staltz’s original proposal, React could add event listeners for DOM events. Instead of defining a heuristic for translating onFoo to foo, a possible solution would be to express the event name in JSX. e.g.

return (
  <paper-slider
    min={props.min}
    max={props.max}
    value={props.value}
    domEvents={
      click: props.onClick
      'value-changed': props.onValueChanged
    }>
  </paper-slider>

2) OR, element authors continue to use the properties down, events up approach, but we promote @treshugart’s react-integration lib to improve interop with React.

Here I lean toward option 1, not that the react-integration lib isn’t awesome, it is(!), but because option 1 lowers the barrier for people to use CEs + React together. In either case, we would continue to encourage folks to treat properties as their source of truth, and we may want to encourage them to only bubble events if they have a good reason to do so. I know that was one of the critiques of the current event system and might be a place where we can agree on a best practice.

Both of these options avoid baking in the on* pattern while we hash out what would be a better long term alternative. I’ll add that I’m very motivated to work on this. I think a primary goal of the extensible web movement and Web Components in general is to take feedback from library and framework authors and use that to improve the platform.

pemrouz commented 8 years ago

+1 for domEvents example, but would prefer just events.

staltz commented 8 years ago

Yes, I agree with everything Rob said. Could still bikeshed on the name for domEvents, but before that we need to know what @sebmarkbage thinks about that.

treshugart commented 8 years ago

+1 for just events

I am curious, though, as to how React will decide how it should set props on an element, as opposed to attributes. This could be troublesome for custom elements needing some things set as attributes, such as aria-.

caridy commented 8 years ago

I'm very supportive of this effort, a property makes sense to accomodate the react's model, but before we go that route, there are other things we should define:

  1. what is the different between events.click prop in a WC and the click listener on a button element?
  2. what is the payload of the event when calling events.click callback?

My main concern is the shape of the event object, and the potential confusions around it. Today, and based on the example from @robdodson, those events will be custom events with a detail object with an arbitrary shape, which will differ from the existing practices and knowledge base that we will have to fight against (e.g.: the click event always comes with a target reference).

I think we will have to bikeshed more around the name of the property. Is it really a collection of events? can we break with the past and not call those event handlers?

dantman commented 8 years ago

I think we will have to bikeshed more around the name of the property. Is it really a collection of events? can we break with the past and not call those event handlers?

Let me put in a -1 for events and a neutral support for domEvents.

events is too generic, we're co-opting a property name that'll apply to every custom element created by anyone with no cooperation with react. A name as generic as events could easily be used by say, a <event-calendar> web component to accept a list of calendar events to display.

treshugart commented 8 years ago

A name as generic as events could easily be used by say, a web component to accept a list of calendar events to display.

This is exactly why we're discussing using a specific syntax (i.e. a $ prefix) for events and attributes in Skate. Though this would be specific to Skate's virtual DOM, it allows the consumer to choose exactly what to do and prevents clashes.

One thing React could do, would be to just Object.assign() like originally proposed. Web components could then expose properties to add / remove event listeners:

class MyComponent extends HTMLElement {
  set onsomeevent (handler) {
    this.addEventListener('someevent', handler);
  }
}

customElements.define('my-component', MyComponent);

const elem = new MyComponent();
elem.onsomeevent = console.log;
elem.dispatchEvent(new Event('someevent'));

The more I think about it, the more I like this proposal for our own purposes instead of using a specific syntax.

dantman commented 8 years ago

One thing React could do, would be to just Object.assign() like originally proposed. Web components could then expose properties to add / remove event listeners:

I don't like this either. Using on{lowercase} properties forces all web components from all frameworks to implement extra code dictated by React in order for their users to be able to use simple event syntax in React; instead of React just having a declarative syntax to interact with the standardized dom events API.

If we want to completely avoid any possibility of property name collision, all we have to do is use a symbol for the property React uses internally.

<paper-slider [ReactDOM.events]={{'value-changed': this.props.onValueChanged}} />

Personally speaking though, if React devs can accept an official interface for 3rd parties to inject custom event implementations (kind of like the current unofficial interface, but better) and 3rd party event modules, using Symbols rather than string property names so there is no collision. I wouldn't mind an alternative where anyone can write a module that will register custom events with React for whatever pattern is used by various Web Component implementations.

import { Event as on } from 'react-polymer';
// ...
<paper-slider [on('value-changed')]={this.props.onValueChanged} />
justinfagnani commented 8 years ago

@treshugart I think putting the burden of writing event handler properties on the custom elements author is the wrong approach. 1) it's extra code that won't always be implemented. 2) it's basically impossible to emulate what the platform does and 3) only using properties doesn't support bubbling or capturing event handlers (Unless you add the property to HTMLElement.prototype, which would be very problematic).

It's very easy for a binding library to offer a syntax that differentiates between attributes, properties and events. addEventListener() is the canonical API for adding handlers, IMO libraries should just support it.

treshugart commented 8 years ago

@dantman that's a good point. I like the symbol suggestion.

@justinfagnani I agree it's extra code and that's what I dislike about it. As for 2, that's a given; I don't think any method could emulate that. For 3, setting events (or via symbol) via react, or using a custom syntax wouldn't do this either, right? My reason for proposing that was that it's very much inline with the original proposal.

I actually wouldn't mind having a convention in web components I build to expose properties corresponding to custom events supported by the component as it enables a declarative API without leaving it up to each templating engine to have custom semantics around how it binds events. All the templating engine needs to do is what something like Incremental DOM does by default: if it's a function, set a prop. If it's anything else, set an attribute. I wish it was prop all the time, though.

EDIT

To be clear, I'd prefer using a symbol approach over my proposal.

robdodson commented 8 years ago

@treshugart

I am curious, though, as to how React will decide how it should set props on an element, as opposed to attributes. This could be troublesome for custom elements needing some things set as attributes, such as aria-.

Good question. It may be the case that some attributes still require a special case. Either using special syntax in React to set the attribute, or React could set it using some kind of heuristic ("any property beginning with aria* gets set as an attribute instead). For what it's worth, there's an effort underway to allow you to set accessibility properties in JavaScript. So in the future, that one piece could be a non-issue 😄

@caridy

My main concern is the shape of the event object, and the potential confusions around it. Today, and based on the example from @robdodson, those events will be custom events with a detail object with an arbitrary shape, which will differ from the existing practices and knowledge base that we will have to fight against (e.g.: the click event always comes with a target reference).

As I understand it, CustomEvent inherits from Event and therefore has the same properties. By calling dispatchEvent we implicitly set the target to the element that's doing the event dispatch. The main unique point is that there's a detail object which can hold extra data.

@treshugart

For 3, setting events (or via symbol) via react, or using a custom syntax wouldn't do this either, right?

I think the idea is that events would be calling addEventListener under the hood. So it would support bubbling at least because the component decides if the event its dispatching bubbles or not.

I actually wouldn't mind having a convention in web components I build to expose properties corresponding to custom events supported by the component as it enables a declarative API without leaving it up to each templating engine to have custom semantics around how it binds events.

If there was a standards based way of doing this I'd be cool with it to. For instance, the observedAttributes array is an easy way for Custom Elements to observe attribute mutations. You could imagine something similar like:

static get exposedEvents() {
  return ['foo-changed', 'bar-input, ...];
}

Somehow exposedEvents would automagically generate on* properties

joeldenning commented 7 years ago

@sebmarkbage I see that react@16.0.0-alpha was released today - does that have the new support you talked about for custom elements, where everything is properties instead of attributes?

Also, one thing I have noticed with React@15 that could also possibly be improved is support for the className prop (and maybe other react-specific properties for DOM elements). In React@15 the following code does something unexpected (at least to me as a user):

/* Rendering <Foo /> will result in the actual DOM having a `classname` attribute. For example, the real dom will look like this:
 * <button is="x-button" id="2" classname="custom_styles" />
 */
class Foo extends React.Component {
  render() {
    return <button is="x-button" id="2" className="custom_styles" />
  }
}

Once React starts sending those as properties instead of attributes, then it seems like the id and className properties will work just fine, since they are already implemented as properties of HTMLElement. However, are there any edge cases where that behavior won't work as expected? Not sure if there is, but thought you guys might know.

joeldenning commented 7 years ago

Answering my own question above: React@16.0.0-alpha does not have support for this yet. I have created #8755 to add it to react@16, though

gaearon commented 7 years ago

FYI, 16.0.0-alpha is not representative of what will end up in 16. We only cut the release to test some changes related to React Native.

joeldenning commented 7 years ago

@gaearon I see. Is there somewhere I can read about 16? Has it been decided when it might be released? Is there a roadmap or tentative list of what might be in it?

rnicholus commented 7 years ago

The related issues milestone is filled with cases, but I'm not personally sure if this is fully representative of the final plans for 16.

gaearon commented 7 years ago

There's no definitive list because we're working on a complete rewrite of the reconciler ("Fiber"—see the talk, some info, progress, some more info, more detailed progress), and it is not yet clear if we will ship it in 16 or wait until 17. The work on Fiber is currently prioritized before other issues.

You can read our weekly meeting notes to keep track of our progress and planning. I expect that we will talk about beginning to plan for 16 and establishing its scope soon, but we have not yet done that.

justinfagnani commented 7 years ago

@joeldenning I took a quick look at #8755 and it seemed like it didn't provide any way to set attributes on custom elements. There really needs to be a way to set both, as some thing like aria attribute can't simply be set via properties, and plenty of styling scenarios might require attributes that the element doesn't support as reflected properties.

Also, some attribute/property pairs are very different to set. Attribute data-foo="bar" needs to be set as a nested property e.dataset.foo = 'bar', style becomes style.cssText. This difference would mean that you need data structured differently depending on whether it's being set on a custom element or not.

justinfagnani commented 7 years ago

This is related to #7249 :

I really think the best way here is explicit control over whether a prop is set as an attribute or property or as an event handler via addEventListener. These API surfaces of an element are distinct. Even though they're often correlated, they needn't be and sometimes aren't even on built-in elements.

In #7249 there's a suggestion to set a property if an element has a property of that name, and an attribute otherwise. This is problematic for custom element upgrades. An element may be not-upgraded, have an attribute set that should have been a property, and then be upgraded and not be able to read the correct value from the attribute.

joeldenning commented 7 years ago

@justinfagnani those are good points. I have a question about the upgrading scenario that you brought up, though: In general, don't elements (both native and custom) look at attributes when they are constructed/connected? And then they look at properties after that? In other words, isn't the general convention for both native and custom elements that properties are the source of truth, and that attributes are just for initial configuration? If a custom element follows that convention (of looking at attributes in connectedCallback and properties after that), doesn't the upgrade problem you talked about disappear?

I've been thinking that the goal is to make it easy for React to interoperate with custom elements that follow a convention like the one I described above. It will of course be possible to interoperate with custom elements that break the convention (via refs), but it will be easy to interoperate with the ones that do.

Do you agree with that? Or do you lean more towards not establishing such a convention?

justinfagnani commented 7 years ago

Elements can certainly look at attributes when they boot up, but only for values that are string serializable.

joeldenning commented 7 years ago

Ah that's a good point. So if an element isn't upgraded yet, you may still need to set properties on it because setting it as an attribute won't work if it's not a string. So

if (propName in obj) {
  obj[propName] = value;
} else {
  obj.setAttribute(propName, value);
}

is probably not the right thing to do. Should React consider doing a customElements.whenDefined() and then execute the above code once we're sure that the custom element is defined and things are upgraded?

treshugart commented 7 years ago

@joeldenning customElements.whenDefined() accepts a name argument for the custom element you're checking for, so React would not only have to know the name of it, but also that it is indeed a custom element with a corresponding definition that will definitely be loaded.

Giving the user full control over if a property, attribute or event is applied to the element means that there's less checks and heuristics that React needs to maintain. It also means the API consumer gets expected behaviour and complete DOM integration. I say DOM because this is really a DOM integration problem, not just custom elements. Integration with the DOM means custom elements inherently work.

caridy commented 7 years ago

update: after a couple of months of research, and some prototyping, it seems to us that React will have to do nothing new to support web-components interoperability in both ways:

@sebmarkbage we can chat next week at TC39, and eventually share the examples, and the docs if it is sufficient in your opinion.

Slotting remains an open question though!

staltz commented 7 years ago

it seems to us that React will have to do nothing new to support web-components interoperability in both ways

Well, that's not anything new. As most frameworks, nothing needs to be done to support web components. This issue started out as an improvement suggestion on how to avoid addEventListener in the first place.

staltz commented 7 years ago

As a reminder, could we reconsider onFoo => foo conversion? Before we went deep into props vs attrs, Sebastian said:

The only concern would be if we should have some heuristic for normalizing the event name. I'm not really a fan of converting things like onXxx into xxx after doing that in MooTools.

Would be better to have more solid arguments for or against that.