React Fire: Modernizing React DOM

@nhunzaker

For latest status, see an update from June 5th, 2019: #13525 (comment)

This year, the React team has mostly been focused on fundamental improvements to React.

As this work is getting closer to completion, we're starting to think of what the next major releases of React DOM should look like. There are quite a few known problems, and some of them are hard or impossible to fix without bigger internal changes.

We want to undo past mistakes that caused countless follow-up fixes and created much technical debt. We also want to remove some of the abstraction in the event system which has been virtually untouched since the first days of React, and is a source of much complexity and bundle size.

We're calling this effort "React Fire".

🔥 React Fire

React Fire is an effort to modernize React DOM. Our goal is to make React better aligned with how the DOM works, revisit some controversial past decisions that led to problems, and make React smaller and faster.

We want to ship this set of changes in a future React major release because some of them will unfortunately be breaking. Nevertheless, we think they're worth it. And we have more than 50 thousands components at Facebook to keep us honest about our migration strategy. We can't afford to rewrite product code except a few targeted fixes or automated codemods.

Strategy

There are a few different things that make up our current plan. We might add or remove something but here's the thinking so far:

Stop reflecting input values in the value attribute (Stop syncing value attribute for controlled inputs #11896). This was originally added in React 15.2.0 via Properly set value and defaultValue for input and textarea #6406. It was very commonly requested because people's conceptual model of the DOM is that the value they see in the DOM inspector should match the value JSX attribute. But that's not how the DOM works. When you type into a field, the browser doesn't update the value attribute. React shouldn't do it either. It turned out that this change, while probably helpful for some code relying on CSS selectors, caused a cascade of bugs — some of them still unfixed to this day. Some of the fallout from this change includes: Form submit button has empty values in 15.2.0 #7179, Firefox validation triggers on input component render #8395, IE 11 and Edge no longer prompt to remember password on controlled form #7328, Date input with defaultValue regression in 15.2 #7233, backspace fails to clear values on input type='email' #11881, Bug: Backspace in input type="number" behaves badly in Blink #7253, Remove loose check on non-number controlled inputs. Fix trailing dot issue. #9584, Inputs should not mutate value on type conversion (when they stringify to the same thing) #9806, [#9712] fix <input type="number" /> value '.98' should not be equal to '0.98'. #9714, Use defaultValue instead of setAttribute('value') #11534, Use the same value synchronization function on number blur #11746, Do not assign node.value on input creation if no change will occur #12925. At this point it's clearly not worth it to keep fighting the browser, and we should revert it. The positive part of this journey is that thanks to tireless work from our DOM contributors (@nhunzaker, @aweary, @jquense, and @philipp-spiess) we now have detailed DOM test fixtures that will help us avoid regressions.
Attach events at the React root rather than the document (Attach event per react container root, rather than on the document #2043). Attaching event handlers to the document becomes an issue when embedding React apps into larger systems. The Atom editor was one of the first cases that bumped into this. Any big website also eventually develops very complex edge cases related to stopPropagation interacting with non-React code or across React roots (Native event.stopPropagation outside of React root cuts out React events #8693, Attach event listeners at the root of the tree instead of document #8117, Event listener attached to document will still be called after calling event.stopPropagation() #12518). We will also want to attach events eagerly to every root so that we can do less runtime checks during updates.
Migrate from onChange to onInput and don’t polyfill it for uncontrolled components ([RFC] onChange -> onInput, and don't polyfill onInput for uncontrolled components #9657). See the linked issue for a detailed plan. It has been confusing that React uses a different event name for what's known as input event in the DOM. While we generally avoid making big changes like this without significant benefit, in this case we also want to change the behavior to remove some complexity that's only necessary for edge cases like mutating controlled inputs. So it makes sense to do these two changes together, and use that as an opportunity to make onInput and onChange work exactly how the DOM events do for uncontrolled components.
Drastically simplify the event system (Play Nicely with The DOM Event System (because it's legacy anyway) #4751). The current event system has barely changed since its initial implementation in 2013. It is reused across React DOM and React Native, so it is unnecessarily abstract. Many of the polyfills it provides are unnecessary for modern browsers, and some of them create more issues than they solve. It also accounts for a significant portion of the React DOM bundle size. We don't have a very specific plan here, but we will probably fork the event system completely, and then see how minimal we can make it if we stick closer to what the DOM gives us. It's plausible that we'll get rid of synthetic events altogether. We should stop bubbling events like media events which don’t bubble in the DOM and don’t have a good reason to bubble. We want to retain some React-specific capabilities like bubbling through portals, but we will attempt to do this via simpler means (e.g. re-dispatching the event). Passive events will likely be a part of this.
className → class (Why are attribute names "class" and "for" discouraged? #4331, see also React Fire: Modernizing React DOM #13525 (comment) below). This has been proposed countless times. We're already allowing passing class down to the DOM node in React 16. The confusion this is creating is not worth the syntax limitations it's trying to protect against. We wouldn't do this change by itself, but combined with everything else above it makes sense. Note we can’t just allow both without warnings because this makes it very difficult for a component ecosystem to handle. Each component would need to learn to handle both correctly, and there is a risk of them conflicting. Since many components process className (for example by appending to it), it’s too error-prone.

Tradeoffs

We can't make some of these changes if we aim to keep exposing the current private React event system APIs for projects like React Native Web. However, React Native Web will need a different strategy regardless because React Fabric will likely move more of the responder system to the native side.
We may need to drop compatibility with some older browsers, and/or require more standalone polyfills for them. We still care about supporting IE11 but it's possible that we will not attempt to smooth over some of the existing browser differences — which is the stance taken by many modern UI libraries.

Rollout Plan

At this stage, the project is very exploratory. We don't know for sure if all of the above things will pan out. Because the changes are significant, we will need to dogfood them at Facebook, and try them out in a gradual fashion. This means we'll introduce a feature flag, fork some of the code, and keep it enabled at Facebook for a small group of people. The open source 16.x releases will keep the old behavior, but on master you will be able to run it with the feature flag on.

I plan to work on the project myself for the most part, but I would very much appreciate more discussion and contributions from @nhunzaker, @aweary, @jquense, and @philipp-spiess who have been stellar collaborators and have largely steered React DOM while we were working on Fiber. If there's some area you're particularly interested in, please let me know and we'll work it out.

There are likely things that I missed in this plan. I'm very open to feedback, and I hope this writeup is helpful.

I love this. Reducing bundle size and the "class" prop are changes that will be very welcome.

Great work!

🙂

Attention form library authors! 🤣

Great!

className → class is fantastic

What about all the others? Seems weird to still be doing clipPath, htmlFor, tabIndex, etc.

Adopting class is a major breakthrough in making the library more friendly for beginners. Congratulations.

This is awesome. I'm so curious how the move to class is actually going work with props.

Seems like ({ class }) => <div class={class} /> would initially present a reserved keyword problem?

@gaearon

This is fantastic news, thanks @gaearon!

I love every of these points, except the className change. It seems downright contradictory to the goal the other points are pursuing (aligning with the DOM API). React binds to DOM properties, not HTML attributes (this is this even articulated in the first point). The DOM Element property is named className, not class. So why would it be named class in React?

Fantastic! Do you have a goal for bundle size reduction?

👏

What about all the others? Seems weird to still be doing clipPath, htmlFor, tabIndex, etc.

I’m open to discussion but I’d argue these changes aren’t worth it (except for maybe).

I think a re-write of the event system is the most interesting aspect of this. There is significant opportunity to reduce the bundle size and ease community contributions.

Let's do it!

Will "React Flare" come as a part of the default React package or need an additional install considering the amount of API's it will ship with?

We'd like to avoid unused code getting bundled. So the current intention is for it to be opt-in per API. Maybe separate entry points in a single package.

Will this mouse and touch event system leverage PointerEvent? I did not see a mention of this web standard in the previous update, so I just wanted to bring it to your attention.

Pointer events are DOM events that are fired for a pointing device. They are designed to create a single DOM event model to handle pointing input devices such as a mouse, pen/stylus or touch (such as one or more fingers). The pointer is a hardware-agnostic device that can target a specific set of screen coordinates. Having a single event model for pointers can simplify creating Web sites and applications and provide a good user experience regardless of the user's hardware.

And here is a direct link to the current browser compatibility.

@jonathantneal

@jonathantneal Yes, the new system heavily use Pointer Events – with fallbacks to Mouse/Touch events when there's no support for Pointer Events.

I am concerned that #11347 was not addressed in this issue. React flunks https://custom-elements-everywhere.com.

Please consider shadow root when redesigning the event system - just attaching to the React root wouldn't solve most issues today, only attaching to the element (#9242, #15759, #13713, #11827)

@gaearon

In this update: #13525 (comment) @gaearon mentions:

However, as we started removing parts of the event system that we thought were unnecessary or outdated, we discovered many edge cases where it was being very helpful and prevented bugs — even in modern browsers.

I was curious if a list of these edge cases are documented anywhere?

@gaearon

@gaearon now that Flare has gone out (SCNR), is there an updated plan (regarding the June 5th, 2019 update) how to proceed?

And like @trusktr, I also would like to get #11347 addressed here.

Could be split polyfills into another bundle especially the one not relevant to major evergreen browsers.

@nhunzaker

Hey all, it's been a while and we've tried some of these things on and off.

Let me give an update on each:

Stop reflecting input values in the value attribute (Stop syncing value attribute for controlled inputs #11896). This was originally added in React 15.2.0 via Properly set value and defaultValue for input and textarea #6406. It was very commonly requested because people's conceptual model of the DOM is that the value they see in the DOM inspector should match the value JSX attribute. But that's not how the DOM works. When you type into a field, the browser doesn't update the value attribute. React shouldn't do it either. It turned out that this change, while probably helpful for some code relying on CSS selectors, caused a cascade of bugs — some of them still unfixed to this day. Some of the fallout from this change includes: Form submit button has empty values in 15.2.0 #7179, Firefox validation triggers on input component render #8395, IE 11 and Edge no longer prompt to remember password on controlled form #7328, Date input with defaultValue regression in 15.2 #7233, backspace fails to clear values on input type='email' #11881, Bug: Backspace in input type="number" behaves badly in Blink #7253, Remove loose check on non-number controlled inputs. Fix trailing dot issue. #9584, Inputs should not mutate value on type conversion (when they stringify to the same thing) #9806, [#9712] fix <input type="number" /> value '.98' should not be equal to '0.98'. #9714, Use defaultValue instead of setAttribute('value') #11534, Use the same value synchronization function on number blur #11746, Do not assign node.value on input creation if no change will occur #12925. At this point it's clearly not worth it to keep fighting the browser, and we should revert it. The positive part of this journey is that thanks to tireless work from our DOM contributors (@nhunzaker, @aweary, @jquense, and @philipp-spiess) we now have detailed DOM test fixtures that will help us avoid regressions.

We still want to do this, but we've decided to "reserve" React 17 to have minimal possible breaking changes so that it can focus on the next item on this list. So this change will wait until React 18.

Attach events at the React root rather than the document (Attach event per react container root, rather than on the document #2043). Attaching event handlers to the document becomes an issue when embedding React apps into larger systems. The Atom editor was one of the first cases that bumped into this. Any big website also eventually develops very complex edge cases related to stopPropagation interacting with non-React code or across React roots (Native event.stopPropagation outside of React root cuts out React events #8693, Attach event listeners at the root of the tree instead of document #8117, Event listener attached to document will still be called after calling event.stopPropagation() #12518). We will also want to attach events eagerly to every root so that we can do less runtime checks during updates.

We're doing this in React 17. This turned out to be a huge chunk of work but thankfully it's finished.

Migrate from onChange to onInput and don’t polyfill it for uncontrolled components ([RFC] onChange -> onInput, and don't polyfill onInput for uncontrolled components #9657). See the linked issue for a detailed plan. It has been confusing that React uses a different event name for what's known as input event in the DOM. While we generally avoid making big changes like this without significant benefit, in this case we also want to change the behavior to remove some complexity that's only necessary for edge cases like mutating controlled inputs. So it makes sense to do these two changes together, and use that as an opportunity to make onInput and onChange work exactly how the DOM events do for uncontrolled components.

We'll likely come back to this but it's unclear how much churn is worth doing here. So this is still TBD.

Drastically simplify the event system (Play Nicely with The DOM Event System (because it's legacy anyway) #4751). The current event system has barely changed since its initial implementation in 2013. It is reused across React DOM and React Native, so it is unnecessarily abstract. Many of the polyfills it provides are unnecessary for modern browsers, and some of them create more issues than they solve. It also accounts for a significant portion of the React DOM bundle size. We don't have a very specific plan here, but we will probably fork the event system completely, and then see how minimal we can make it if we stick closer to what the DOM gives us. It's plausible that we'll get rid of synthetic events altogether. We should stop bubbling events like media events which don’t bubble in the DOM and don’t have a good reason to bubble. We want to retain some React-specific capabilities like bubbling through portals, but we will attempt to do this via simpler means (e.g. re-dispatching the event). Passive events will likely be a part of this.

We've tried this early in 2019, and a truly minimal event system did not work out very well in our internal testing. There was quite a bit of cross-browser normalization React is doing that is still useful for people with older browsers, or in more niche areas like rich text input editors using contentEditable. That said, as a part of our work on attaching events to roots, we've removed a lot of abstraction from the event system so it's easier to understand and improve in the future. As a part of React 17, we're removing "event pooling" which has caused a lot of confusion, and we're also no longer bubbling the onScroll event. We will likely follow with stopping bubbling of media events in React 18, bringing React behavior closer to the browser one. We did save some bytes with the new event system, but they were taken by new features we're working on, so it won't lead to a bundle size decrease overall.

className → class (Why are attribute names "class" and "for" discouraged? #4331, see also React Fire: Modernizing React DOM #13525 (comment) below). This has been proposed countless times. We're already allowing passing class down to the DOM node in React 16. The confusion this is creating is not worth the syntax limitations it's trying to protect against. We wouldn't do this change by itself, but combined with everything else above it makes sense. Note we can’t just allow both without warnings because this makes it very difficult for a component ecosystem to handle. Each component would need to learn to handle both correctly, and there is a risk of them conflicting. Since many components process className (for example by appending to it), it’s too error-prone.

This was the most controversial part of the proposal. Since then, we released Hooks, which encourage writing function components. In function components, we generally suggest using destructuring for props, but you can't write { class, ... } because it would be a syntax error. So overall it's not clear that this is ergonomic enough to actually follow through with. I think it's plausible we'll revisit this in the future, or at least make class not warn and let people do what they want. But for now, we'll shelving this idea.

Hi, it's a great article!
Just wanted to know if there is any plan in pipeline to reduce React-DOM prod size? For mobile applications, it is still an overhead as the browser will be parsing 100+ KB of React-DOM JS and then other modules. Then app-specific JS.
For content-rich pages, it is causing greater Blocking and greater TTI.

Any Idea when can we see such changes?

@morevolk-latei

@morevolk-latei In your measurements, how much time is spent parsing 100 KB of ReactDOM?

update navigation component for german language

Navigation Menu

Repository navigation

React Fire: Modernizing React DOM #13525

🔥 React Fire

Strategy

Tradeoffs

Rollout Plan

202 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

React Fire: Modernizing React DOM #13525

Description

🔥 React Fire

Strategy

Tradeoffs

Rollout Plan

Activity

matteing commented on Aug 31, 2018

developit commented on Aug 31, 2018

erikras commented on Aug 31, 2018

LaurenceM10 commented on Aug 31, 2018

ryanflorence commented on Aug 31, 2018

diegohaz commented on Aug 31, 2018

tannerlinsley commented on Aug 31, 2018

solomonhawk commented on Aug 31, 2018

felixfbecker commented on Aug 31, 2018

alexparish commented on Aug 31, 2018

mknepprath commented on Aug 31, 2018

gaearon commented on Aug 31, 2018

nhunzaker commented on Aug 31, 2018

202 remaining items

Technologeek commented on Jul 3, 2019

gaearon commented on Jul 4, 2019

jonathantneal commented on Jul 11, 2019

trueadm commented on Jul 11, 2019

trusktr commented on Sep 26, 2019

felixfbecker commented on Oct 28, 2019

mikesherov commented on Nov 3, 2019

sbusch commented on Jan 20, 2020

marcthe12 commented on Jun 21, 2020

gaearon commented on Aug 11, 2020

morevolk-latei commented on Aug 12, 2020

gaearon commented on Aug 12, 2020

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions