We want to make sure that we can block the reveal of a well designed
complete shell reliably. In the Suspense model, client transitions don't
have any way to implicitly resolve. This means you need to use Suspense
or SuspenseList to explicitly split the document. Relying on implicit
would mean you can't add a Suspense boundary later where needed. So we
highly encourage the use of them around large content.
However, if you have constructed a too large shell (e.g. by not adding
any Suspense boundaries at all) then that might take too long to render
on the client. We shouldn't punish users (or overzealous metrics
tracking tools like search engines) in that scenario.
This opts out of render blocking if the shell ends up too large to be
intentional and too slow to load. Instead it deopts to showing the
content split up in arbitrary ways (browser default). It only does this
for SSR, and not client navs so it's not reliable.
In fact, we issue an error to `onError`. This error is recoverable in
that the document is still produced. It's up to your framework to decide
if this errors the build or just surface it for action later.
What should be the limit though? There's a trade off here. If this limit
is too low then you can't fit a reasonably well built UI within it
without getting errors. If it's too high then things that accidentally
fall below it might take too long to load.
I came up with 512kB of uncompressed shell HTML. See the comment in code
for the rationale for this number. TL;DR: Data and theory indicates that
having this much content inside `rel="expect"` doesn't meaningfully
change metrics. Research of above-the-fold content on various websites
indicate that this can comfortable fit all of them which should be
enough for any intentional initial paint.
Block the view transition on suspensey images Up to 500ms just like the
client.
We can't use `decode()` because a bug in Chrome where those are blocked
on `startViewTransition` finishing we instead rely on sync decoding but
also that the image is live when it's animating in and we assume it
doesn't start visible.
However, we can block the View Transition from starting on the `"load"`
or `"error"` events.
The nice thing about blocking inside `startViewTransition` is that we
have already done the layout so we can only wait on images that are
within the viewport at this point. We might want to do that in Fiber
too. If many image doesn't have fixed size but need to load first, they
can all end up in the viewport. We might consider only doing this for
images that have a fixed size or only a max number that doesn't have a
fixed size.
Alternative to #33421. The difference is that this also adds an
underscore between the "R" and the ID.
The reason we wanted to use special characters is because we use the
full spectrum of A-Z 0-9 in our ID generation so we can basically
collide with any common word (or anyone using a similar algorithm,
base64 or even base16). It's a little less likely that someone would put
`_R_` specifically unless you generate like two IDs separated by
underscore.
fixes https://github.com/facebook/react/issues/32449
This is my first time touching this code. There are multiple systems in
place here and I wouldn't be surprised to learn that this has to be
handled in some other areas too. I have found some other style-related
code areas but I had no time yet to double-check them.
cc @gnoff
I believe that these mean the same thing. We don't have to emit the
attribute if it's `none` for these cases because if there is no matching
scenario we won't apply the animation in this case.
The only case where we have to emit `none` in the attribute is for
`vt-update` because those can block updates from propagating upwards.
When needed.
For the external runtime we always include this wrapper.
For others, we only include it if we have an ViewTransitions affecting.
If we discover the ViewTransitions late, then we can upgrade an already
emitted instruction.
This doesn't yet do anything useful with it, that's coming in a follow
up. This is just the mechanism for how it gets installed.
Stacked on #33194 and #33200.
When Suspense boundaries reveal during streaming, the Fizz runtime will
be responsible for animating the reveal if necessary (not in this PR).
However, for the future runtime to know what to do it needs to know
about the `<ViewTransition>` configuration to apply.
Ofc, these are virtual nodes that disappear from the HTML. We could
model them as comments like we do with other virtual nodes like Suspense
and Activity. However, that doesn't let us target them with
querySelector and CSS (for no-JS transitions). We also don't have to
model every ViewTransition since not every combination can happen using
only the server runtime. So instead this collapses `<ViewTransition>`
and applies the configuration to the inner DOM nodes.
```js
<ViewTransition name="hi">
<div />
<div />
</ViewTransition>
```
Becomes:
```html
<div vt-name="hi" vt-update="auto"></div>
<div vt-name="hi_1" vt-update="auto"></div>
```
I use `vt-` prefix as opposed to `data-` to keep these virtual
attributes away from user specific ones but we're effectively claiming
this namespace.
There are four triggers `vt-update`, `vt-enter`, `vt-exit` and
`vt-share`. The server resolves which ones might apply to this DOM node.
The value represents the class name (after resolving
view-transition-type mappings) or `"auto"` if no specific class name is
needed but this is still a trigger.
The value can also be `"none"`. This is different from missing because
for example an `vt-update="none"` will block mutations inside it from
triggering the boundary where as a missing `vt-update` would bubble up
to be handled by a parent.
`vt-name` is technically only necessary when `vt-share` is specified to
find a pair. However, since an explicit name can also be used to target
specific CSS selectors, we include it even for other cases.
We want to exclude as many of these annotations as possible.
`vt-enter` can only affect the first DOM node inside a Suspense
boundary's content since the reveal would cause it to enter but nothing
deeper inside. Similarly `vt-exit` can only affect the first DOM node
inside a fallback. So for every other case we can exclude them. (For
future MPA ViewTransitions of the whole document it might also be
something we annotate to children inside the `<body>` as well.) Ideally
we'd only include `vt-enter` for Suspense boundaries that actually
flushed a fallback but since we prepare all that content earlier it's
hard to know.
`vt-share` can be anywhere inside an fallback or content. Technically we
don't have to include it outside the root most Suspense boundary or for
boundaries that are inlined into the root shell. However, this is tricky
to detect. It would also not be correct for future MPA ViewTransitions
because in that case the shared scenario can affect anything in the two
documents so it needs to be in every node everywhere which is
effectively what we do. If a `share` class is specified but it has no
explicit name, we can exclude it since it can't match anything.
`vt-update` is only necessary if something below or a sibling might
update like a Suspense boundary. However, since we don't know when
rendering a segment if it'll later asynchronously add a Suspense
boundary later we have to assume that anywhere might have a child. So
these are always included. We collapse to use the inner most one when
directly nested though since that's the one that ends up winning.
There are some weird edge cases that can't be fully modeled by the lack
of virtual nodes.
Removes the `isFallback` flag on Tasks and tracks it on the
formatContext instead.
Less memory and avoids passing and tracking extra arguments to all the
pushStartInstance branches that doesn't need it.
We'll need to be able to track more Suspense related contexts on this
for View Transitions anyway.
Enabled in experimental channel.
We know this is critical semantics to enforce at the HTML level since if
you don't then you can't add explicit boundaries after the fact.
However, this might have to go in a major release to allow for
upgrading.
Because we now decided whether to outline in the flushing phase, when
we're writing the preamble we don't yet know if we will make that
decision so we don't know if it's safe to omit the external runtime.
However, if you are providing an external runtime it's probably a pretty
safe bet you're streaming something dynamically that's likely to need it
so we can always include it.
The main thing is that this makes it hard to test it because it affects
our tests in ways it wouldn't otherwise so we have to add a bunch of
conditions.
Stacked on #33073.
React semantics is that Suspense boundaries reveal with a throttle
(300ms). That helps avoid flashing reveals when a stream reveals many
individual steps back to back. It can also improve overall performance
by batching the layout and paint work that has to happen at each step.
Unfortunately we never implemented this for SSR streaming - only for
client navigations. This is highly noticeable on very dynamic sites with
lots of Suspense boundaries. It can look good with a client nav but feel
glitchy when you reload the page or initial load.
This fixes the Fizz runtime to be throttled and reveals batched into a
single paint at a time. We do this by first tracking the last paint
after the complete (this will be the first paint if `rel="expect"` is
respected). Then in the `completeBoundary` operation we queue the
operation and then flush it all into a throttled batch.
Another motivation is that View Transitions need to operate as a batch
and individual steps get queued in a sequence so it's extra important to
include as much content as possible in each animated step. This will be
done in a follow up for SSR View Transitions.
Stacked on #33066 and #33068.
Currently we're passing `errorDigest` to `completeBoundary` if there is
a client side error (only CSS loading atm). This only exists because of
`completeBoundaryWithStyles`. Normally if there's a server-side error
we'd emit the `clientRenderBoundary` instruction instead. This adds
unnecessary code to the common case where all styles are in the head.
This is about to get worse with batching because client render shouldn't
be throttled but complete should be.
The first commit moves the client render logic inline into
`completeBoundaryWithStyles` so we only pay for it when styles are used.
However, the approach I went with in the second commit is to reuse the
`$RX` instruction instead (`clientRenderBoundary`). That way if you have
both it ends up being amortized. However, it does mean we have to emit
the `$RX` (along with the `$RC` helper if any
`completeBoundaryWithStyles` instruction is needed.
When we end up creating an incomplete state in the shell we end up not
flushing anything. As a hack, in this case we need to reset the
ResumableState because some of the ResumableState is still relevant
(e.g. any preloads that went into headers) but some of the
ResumableState needs to be reset since they assume that what we produced
actually flushed.
We didn't reset the instructions state but we haven't actually flushed
any of the instructions so it needs to reset.
The semantics of React is that anything outside of Suspense boundaries
in a transition doesn't display until it has fully unsuspended. With SSR
streaming the intention is to preserve that.
We explicitly don't want to support the mode of document streaming
normally supported by the browser where it can paint content as tags
stream in since that leads to content popping in and thrashing in
unpredictable ways. This should instead be modeled explictly by nested
Suspense boundaries or something like SuspenseList.
After the first shell any nested Suspense boundaries are only revealed,
by script, once they're fully streamed in to the next boundary. So this
is already the case there. However, for the initial shell we have been
at the mercy of browser heuristics for how long it decides to stream
before the first paint.
Chromium now has [an API explicitly for this use
case](https://developer.mozilla.org/en-US/docs/Web/API/View_Transition_API/Using#stabilizing_page_state_to_make_cross-document_transitions_consistent)
that lets us model the semantics that we want. This is always important
but especially so with MPA View Transitions.
After this a simple document looks like this:
```html
<!DOCTYPE html>
<html>
<head>
<link rel="expect" href="#«R»" blocking="render"/>
</head>
<body>
<p>hello world</p>
<script src="bootstrap.js" id="«R»" async=""></script>
...
</body>
</html>
```
The `rel="expect"` tag indicates that we want to wait to paint until we
have streamed far enough to be able to paint the id `"«R»"` which
indicates the shell.
Ideally this `id` would be assigned to the root most HTML element in the
body. However, this is tricky in our implementation because there can be
multiple and we can render them out of order.
So instead, we assign the id to the first bootstrap script if there is
one since these are always added to the end of the shell. If there isn't
a bootstrap script then we emit an empty `<template
id="«R»"></template>` instead as a marker.
Since we currently put as much as possible in the shell if it's loaded
by the time we render, this can have some negative effects for very
large documents. We should instead apply the heuristic where very large
Suspense boundaries get outlined outside the shell even if they're
immediately available. This means that even prerenders can end up with
script tags.
We only emit the `rel="expect"` if you're rendering a whole document.
I.e. if you rendered either a `<html>` or `<head>` tag. If you're
rendering a partial document, then we don't really know where the
streaming parts are anyway and can't provide such guarantees. This does
apply whether you're streaming or not because we still want to block
rendering until the end, but in practice any serialized state that needs
hydrate should still be embedded after the completion id.
This lets us write them early in the render phase.
This should be safe because even if we write them deeply, then they
still can't be wrapped by a element because then they'd no longer be in
the document scope anymore. They end up flat in the body and so when we
search the content we'll discover them.
Uses `&` for Activity as opposed to `$` for Suspense. This will be used
to delimitate which nodes we can skip hydrating.
This isn't used on the client yet. It's just a noop on the client
because it's just an unknown comment. This just adds the SSR parts.
Behind the `enableSrcObject` flag. This is revisiting a variant of what
was discussed in #11163.
Instead of supporting the [`srcObject`
property](https://developer.mozilla.org/en-US/docs/Web/API/HTMLMediaElement/srcObject)
as a separate name, this adds an overload of `src` to allow objects to
be passed. The DOM needs to add separate properties for the object forms
since you read back but it doesn't make sense for React's write-only API
to do that. Similar to how we'll like add an overload for
`popoverTarget` instead of calling it `popoverTargetElement` and how
`style` accepts an object and it's not `styleObject={{...}}`.
There are a number of reason to revisit this.
- It's just way more convenient to have this built-in and it makes
conceptual sense. We typically support declarative APIs and polyfill
them when necessary.
- RSC supports Blobs and by having it built-in you don't need a Client
Component wrapper to render it where as doing it with effects would
require more complex wrappers. By picking Blobs over base64,
client-navigations can use the more optimized binary encoding in the RSC
protocol.
- The timing aspect of coordinating it with Suspensey images and image
decoding is a bit tricky to get right because if you set it in an effect
it's too late because you've already rendered it.
- SSR gets complicated when done in user space because you have to
handle both branches. Likely with `useSyncExternalStore`.
- By having it built-in we could optimize the payloads shared between
RSC payloads embedded in the HTML and data URLs.
This does not support objects for `<source src>` nor `<img srcset>`.
Those don't really have equivalents in the DOM neither. They're mainly
for picking an option when you don't know programmatically. However, for
this use case you're really better off picking a variant before
generating the blobs.
We may support Response objects in the future too as per
https://github.com/whatwg/fetch/issues/49
For the `useId` algorithm we used colon `:` before and after.
https://github.com/facebook/react/pull/23360
This avoids collisions in general by using an unusual characters. It
also avoids collisions when concatenated with some other ID.
Unfortunately, `:` is not a valid character in `view-transition-name`.
This PR swaps the format from:
```
:r123:
```
To the unicode:
```
«r123»
```
Which is valid CSS selectors. This also allows them being used for
`querySelector()` which we didn't really find a legit use for but seems
ok-ish.
That way you can get a view-transition-name that you can manually
reference. E.g. to generate styles:
```js
const id = useId();
return <>
<style>{`
::view-transition-group(${id}) { ... }
::view-transition-old(${id}) { ... }
::view-transition-new(${id}) { ... }
`}</style>
<ViewTransition name={id}>...</ViewTransition>
</>;
```
Link headers are optionally supported for cases where you prefer to send
resource loading hints before you're ready to send the body of a
request. While many resources can be correctly preloaded from a link
header responsive images are currently not supported and end up
preloading the default src rather than the correctly sized image. Until
responsive images are supported React will not allow these images to
preload as headers and will retain them to preload as HTML.
closes: #32437
follow up to https://github.com/facebook/react/pull/32163
This continues the work of making Suspense workable anywhere in a
react-dom tree. See the prior PRs for how we handle server rendering and
client rendering. In this change we update the hydration implementation
to be able to locate expected nodes. In particular this means hydration
understands now that the default hydration context is the document body
when the container is above the body.
One case that is unique to hydration is clearing Suspense boundaries.
When hydration fails or when the server instructs the client to recover
an errored boundary it's possible that the html, head, and body tags in
the initial document were written from a fallback or a different primary
content on the server and need to be replaced by the client render.
However these tags (and in the case of head, their content) won't be
inside the comment nodes that identify the bounds of the Suspense
boundary. And when client rendering you may not even render the same
singletons that were server rendered. So when server rendering a
boudnary which contributes to the preamble (the html, head, and body tag
openings plus the head contents) we emit a special marker comment just
before closing the boundary out. This marker encodes which parts of the
preamble this boundary owned. If we need to clear the suspense boundary
on the client we read this marker and use it to reset the appropriate
singleton state.
Suspense is meant to be composable but there has been a lonstanding
limitation with using Suspense above the `<body>` tag of an HTML
document due to peculiarities of how HTML is parsed. For instance if you
used Suspense to render an entire HTML document and had a fallback that
might flush an alternate Document the comment nodes which describe this
boundary scope won't be where they need to be in the DOM for client
React to properly hydrate them. This is somewhat a problem of our own
making in that we have a concept of a Preamble and we leave the closing
body and html tags behind until streaming has completed which produces a
valid HTML document that also matches the DOM structure that would be
parsed from it. However Preambles as a concept are too important to
features like Float to imagine moving away from this model and so we can
either choose to just accept that you cannot use Suspense anywhere
except inside the `<body>` or we can build special support for Suspense
into react-dom that has a coherent semantic with how HTML documents are
written and parsed.
This change implements Suspense support for react-dom/server by
correctly serializing boundaries during rendering, prerendering, and
resumgin on the server. It does not yet support Suspense everywhere on
the client but this will arrive in a subsequent change. In practice
Suspense cannot be used above the `<body>` tag today so this is not a
breaking change since no programs in the wild could be using this
feature anyway.
React's streaming rendering of HTML doesn't lend itself to replacing the
contents of the documentElement, head, or body of a Document. These are
already special cased in fiber as HostSingletons and similarly for Fizz
the values we render for these tags must never be updated by the Fizz
runtime once written. To accomplish these we redefine the Preamble as
the tags that represent these three singletons plus the contents of the
document.head. If you use Suspense above any part of the Preamble then
nothing will be written to the destination until the boundary is no
longer pending. If the boundary completes then the preamble from within
that boudnary will be output. If the boundary postpones or errors then
the preamble from the fallback will be used instead.
Additionally, by default anything that is not part of the preamble is
implicitly in body scope. This leads to the somewhat counterintuitive
consequence that the comment nodes we use to mark the borders of a
Suspense boundary in Fizz can appear INSIDE the preamble that was
rendered within it.
```typescript
render((
<Suspense>
<html lang="en">
<body>
<div>hello world</div>
</body>
</html>
</Suspense>
))
```
will produce an HTML document like this
```html
<!DOCTYPE html>
<html lang="en">
<head></head>
<body>
<!--$--> <-- this is the comment Node representing the outermost Suspense
<div>hello world</div>
<$--/$-->
</body>
</html>
```
Later when I update Fiber to support Suspense anywhere hydration will
similarly start implicitly in the document body when the root is part of
the preamble (the document or one of it's singletons).
The interstital characters in our link header tracking are not
contributing to the remaining capacity calculation so when a lot of
inditidual links are present in the link header it can allow an
overflowing link header to be included. This change corrects the math so
it properly prevents overflow.
Follow up to #30105.
This supports `renderToMarkup` in a non-RSC environment (not the
`react-server` condition).
This is just a Fizz renderer but it errors at runtime when you use
state, effects or event handlers that would require hydration - like the
RSC version would. (Except RSC can give early errors too.)
To do this I have to move the `react-html` builds to a new `markup`
dimension out of the `dom-legacy` dimension so that we can configure
this differently from `renderToString`/`renderToStaticMarkup`.
Eventually that dimension can go away though if deprecated. That also
helps us avoid dynamic configuration and we can just compile in the
right configuration so the split helps anyway.
One consideration is that if a compiler strips out useEffects or inlines
initial state from useState, then it would not get called an the error
wouldn't happen. Therefore to preserve semantics, a compiler would need
to inject some call that can check the current renderer and whether it
should throw.
There is an argument that it could be useful to not error for these
because it's possible to write components that works with SSR but are
just optionally hydrated. However, there's also an argument that doing
that silently is too easy to lead to mistakes and it's better to error -
especially for the e-mail use case where you can't take it back but you
can replay a queue that had failures. There are other ways to
conditionally branch components intentionally. Besides if you want it to
be silent you can still use renderToString (or better yet
renderToReadableStream).
The primary mechanism is the RSC environment and the client-environment
is really the secondary one that's only there to support legacy
environments. So this also ensures parity with the primary environment.
sanitize javascript: urls for <object> tags
React 19 added sanitization for `javascript:` URLs for `href` properties
on various tags. This PR also adds that sanitization for `<object>` tags
as well that Firefox otherwise executes.
This follows the same principle as in #28611.
We cannot serialize Blobs of a form data into HTML because you can't
initialize a file input to some value. However the serialization of
state in an Action can contain blobs. In this case we do error but
outside the try/catch that recovers to error to client replaying instead
of MPA mode. This errors earlier to ensure that this works.
Testing this is a bit annoying because JSDOM doesn't have any of the
Blob methods but the Blob needs to be compatible with FormData and the
FormData needs to be compatible with `<form>` nodes in these tests. So I
polyfilled those in JSDOM with some hacks.
A possible future enhancement would be to encode these blobs in a base64
mode instead and have some way to receive them on the server. It's just
a matter of layering this. I think the RSC layer's `FORM_DATA`
implementation can pass some flag to encode as base64 and then have
decodeAction include some way to parse them. That way this case would
work in MPA mode too.
stacked on #28870
inline script children have been encoded as HTML for a while now but
this can easily break script parsing so practically if you were
rendering inline scripts you were using dangerouslySetInnerHTML. This is
not great because now there is no escaping at all so you have to be even
more careful. While care should always be taken when rendering untrusted
script content driving users to use dangerous APIs is not the right
approach and in this PR the escaping functionality used for
bootstrapScripts and importMaps is being extended to any inline script.
the approach is to escape 's' or 'S" with the appropriate unicode code
point if it is inside a <script or </script sequence. This has the nice
benefit of minimally escaping the text for readability while still
preserving full js parsing capabilities. As articulated when we
introduced this escaping for prior use cases this is only safe because
we are escaping the entire script content. It would be unsafe if we were
not escaping the entirety of the script because we would no longer be
able to ensure there are no earlier or later <script sequences that put
the parser in unexpected states.
style text content has historically been escaped as HTML which is
non-sensical and often leads users to using dangerouslySetInnerHTML as a
matter of course. While rendering untrusted style rules is a security
risk React doesn't really provide any special protection here and
forcing users to use a completely unescaped API is if anything worse. So
this PR updates the style escaping rules for Fizz to only escape the
text content to ensure the tag scope cannot be closed early. This is
accomplished by encoding "s" and "S" as hexadecimal unicode
representation "\73 " and "\53 " respectively when found within a
sequence like </style>. We have to be careful to support casing here
just like with the script closing tag regex for bootstrap scripts.
Based on:
- #28804
---
This sets adds a new ReactDOM export called requestFormReset, including
setting up the export and creating a method on the internal ReactDOM
dispatcher. It does not yet add any implementation.
Doing this in its own commit for review purposes.
The API itself will be explained in the next PR.
`<noscript>` scopes should be considered inert from the perspective of
Fizz since we assume they'll only be used in rare and adverse
circumstances. When we added preload support for img tags we did not
include the noscript scope check in the opt-out for preloading. This
change adds it in
fixes: #27910
Stacked on #28771
ReactDOMCurrentDispatcher has longer property names for various methods.
These methods are only ever called internally and don't need to be
represented with as many characters. This change shortens the names and
aligns them with the hint codes we use in Flight. This alignment is
passive since not all dispatcher methods will exist as flight
instructions but where they can line up it seems reasonable to make them
do so
Stacked on #28751
ReactDOMSharedInternals uses properties of considerable length to model
mutuable state. These properties are not mangled during minification and
contribute a not insigificant amount to the uncompressed bundle size and
to a lesser degree compressed bundle size.
This change rewrites the DOMInternals in a way that shortens property
names so we can have smaller builds.
It also treats the entire object as a mutable container rather than
having different mutable sub objects.
The same treatment should be given to ReactSharedInternals
This PR moves `flushSync` out of the reconciler. there is still an
internal implementation that is used when these semantics are needed for
React methods such as `unmount` on roots.
This new isomorphic `flushSync` is only used in builds that no longer
support legacy mode.
Additionally all the internal uses of flushSync in the reconciler have
been replaced with more direct methods. There is a new
`updateContainerSync` method which updates a container but forces it to
the Sync lane and flushes passive effects if necessary. This combined
with flushSyncWork can be used to replace flushSync for all instances of
internal usage.
We still maintain the original flushSync implementation as
`flushSyncFromReconciler` because it will be used as the flushSync
implementation for FB builds. This is because it has special legacy mode
handling that the new isomorphic implementation does not need to
consider. It will be removed from production OSS builds by closure
though
We previously only included the component stack.
Cleaned up the fields in Fizz server that wasn't using consistent hidden
classes in dev vs prod.
Added a prefix to errors serialized from server rendering. It can be a
bit confusing to see where this error came from otherwise since it
didn't come from elsewhere on the client. It's really kind of confusing
with other recoverable errors that happen on the client too.
We've rolled out this flag internally on WWW. This PR removed flag
`enableCustomElementPropertySupport`
Test plan:
-- `yarn test`
Co-authored-by: Ricky Hanlon <rickhanlonii@gmail.com>
This a follow up to #28564. It's alternative to #28609 which takes
#28610 into account.
It used to be possible to return JSX from an action with
`useActionState`.
```js
async function action(errors, payload) {
"use server";
try {
...
} catch (x) {
return <div>Error message</div>;
}
}
```
```js
const [errors, formAction] = useActionState(action);
return <div>{errors}</div>;
```
Returning JSX from an action is itself not anything problematic. It's
that it also has to return the previous state to the action reducer
again that's the problem. When this happens we accidentally could
serialize an Element back to the server.
I fixed this in #28564 so it's now blocked if you don't have a temporary
reference set.
However, you can't have that for the progressive enhancement case. The
reply is eagerly encoded as part of the SSR render. Typically you
wouldn't have these in the initial state so the common case is that they
show up after the first POST back that yields an error and it's only in
the no-JS case where this happens so it's hard to discover.
As noted in #28609 there's a security implication with allowing elements
to be sent across this kind of payload, so we can't just make it work.
When an error happens during SSR our general policy is to try to recover
on the client instead. After all, SSR is mainly a perf optimization in
React terms and it's not primarily intended for a no JS solution.
This PR takes the approach that if we fail to generate the progressive
enhancement payload. I.e. if the serialization of previous state /
closures throw. Then we fallback to the replaying semantics just client
actions instead which will succeed.
The effect of this is that this pattern mostly just works:
- First render in the typical case doesn't have any JSX in it so it just
renders a progressive enhanced form.
- If JS fails to hydrate or you click early we do a form POST. If that
hits an error and it tries to render it using JSX, then the new page
will render successfully - however this time with a Replaying form
instead.
- If you try to submit the form again it'll have to be using JS.
Meaning if you use JSX as the error return value of form state and you
make a first attempt that fails, then no JS won't work because either
the first or second attempt has to hydrate.
We have ideas for potentially optimizing away serializing unused
arguments like if you don't actually use previous state which would also
solve it but it wouldn't cover all cases such as if it was deeply nested
in complex state.
Another approach that I considered was to poison the prev state if you
passed an element back but let it through to the action but if you try
to render the poisoned value, it wouldn't work. The downside of this is
when to error. Because in the progressive enhancement case it wouldn't
error early but when you actually try to invoke it at which point it
would be too late to fallback to client replaying. It would probably
have to always error even on the client which is unfortunate since this
mostly just works as long as it hydrates.
A while back we implemented a heuristic that if a chunk was large it was
assumed to be produced by the render and thus was safe to stream which
results in transferring the underlying object memory. Later we ran into
an issue where a precomputed chunk grew large enough to trigger this
hueristic and it started causing renders to fail because once a second
render had occurred the precomputed chunk would not have an underlying
buffer of bytes to send and these bytes would be omitted from the
stream. We implemented a technique to detect large precomputed chunks
and we enforced that these always be cloned before writing.
Unfortunately our test coverage was not perfect and there has been for a
very long time now a usage pattern where if you complete a boundary in
one flush and then complete a boundary that has stylehsheet dependencies
in another flush you can get a large precomputed chunk that was not
being cloned to be sent twice causing streaming errors.
I've thought about why we even went with this solution in the first
place and I think it was a mistake. It relies on a dev only check to
catch paired with potentially version specific order of operations on
the streaming side. This is too unreliable. Additionally the low limit
of view size for Edge is not used in Node.js but there is not real
justification for this.
In this change I updated the view size for edge streaming to match Node
at 2048 bytes which is still relatively small and we have no data one
way or another to preference 512 over this. Then I updated the assertion
logic to error anytime a precomputed chunk exceeds the size. This
eliminates the need to clone these chunks by just making sure our view
size is always larger than the largest precomputed chunk we can possibly
write. I'm generally in favor of this for a few reasons.
First, we'll always know during testing whether we've violated the limit
as long as we exercise each stream config because the precomputed chunks
are created in module scope. Second, we can always split up large chunks
so making sure the precomptued chunk is smaller than whatever view size
we actually desire is relatively trivial.
The idea here is that host dispatchers are not bound to renders so we
need to be able to dispatch to them at any time. This updates the
implementation to chain these dispatchers so that each renderer can
respond to the dispatch. Semantically we don't always want every
renderer to do this for instance if Fizz handles a float method we don't
want Fiber to as well so each dispatcher implementation can decide if it
makes sense to forward the call or not. For float methods server
disaptchers will handle the call if they can resolve a Request otherwise
they will forward. For client dispatchers they will handle the call and
always forward. The choice needs to be made for each dispatcher method
and may have implications on correct renderer import order. For now we
just live with the restriction that if you want to use server and client
together (such as renderToString in the browser) you need to import the
server renderer after the client renderer.
Sebastian Markbåge
Mateusz Burzyński
Josh Story
Ricky
Brooke
Jan Kassens
Andrew Clark
Sebastian Silbermann
Jack Pope
Andrey LunyovandRicky Hanlon
Justin Dhillon
