If we don't handle Lazy types specifically in `renderDebugModel`, all of
their properties will be emitted using `renderDebugModel` as well. This
also includes its `_debugInfo` property, if the Lazy comes from the
Flight client. That array might contain objects that are deduped, and
resolving those references in the client can cause runtime errors, e.g.:
```
TypeError: Cannot read properties of undefined (reading '$$typeof')
```
This happened specifically when an "RSC stream" debug info entry, coming
from the Flight client through IO tracking, was emitted and its
`debugTask` property was deduped, which couldn't be resolved in the
client.
To avoid actually initializing a lazy causing a side-effect, we make
some assumptions about the structure of its payload, and only emit
resolved or rejected values, otherwise we emit a halted chunk.
When we report an error we typically log the owner stack of the thing
that caught the error. Similarly we restore the `console.createTask`
scope of the catching component when we call `reportError` or
`console.error`.
We also have a special case if something throws during reconciliation
which uses the Server Component task as far as we got before we threw.
https://github.com/facebook/react/blob/main/packages/react-reconciler/src/ReactChildFiber.js#L1952-L1960
Chrome has since fixed it (on our request) that the Error constructor
snapshots the Task at the time the constructor was created and logs that
in `reportError`. This is a good thing since it means we get a coherent
stack. Unfortunately, it means that the fake Errors that we create in
Flight Client gets a snapshot of the task where they were created so
when they're reported in the console they get the root Task instead of
the Task of the handler of the error.
Ideally we'd transfer the Task from the server and restore it. However,
since we don't instrument the Error object to snapshot the owner and we
can't read the native Task (if it's even enabled on the server) we don't
actually have a correct snapshot to transfer for a Server Component
Error. However, we can use the parent's task for where the error was
observed by Flight Server and then encode that as a pseudo owner of the
Error.
Then we use this owner as the Task which the Error is created within.
Now the client snapshots that Task which is reported by `reportError` so
now we have an async stack for Server Component errors again. (Note that
this owner may differ from the one observed by `captureOwnerStack` which
gets the nearest Server Component from where it was caught. We could
attach the owner to the Error object and use that owner when calling
`onCaughtError`/`onUncaughtError`).
Before:
<img width="911" height="57" alt="Screenshot 2025-09-10 at 10 57 54 AM"
src="https://github.com/user-attachments/assets/0446ef96-fad9-4e17-8a9a-d89c334233ec"
/>
After:
<img width="910" height="128" alt="Screenshot 2025-09-10 at 11 06 20 AM"
src="https://github.com/user-attachments/assets/b30e5892-cf40-4246-a588-0f309575439b"
/>
Similarly, there are Errors and warnings created by ChildFiber itself.
Those execute in the scope of the general render of the parent Fiber.
They used to get the scope of the nearest client component parent (e.g.
div in this case) but that's the parent of the Server Component. It
would be too expensive to run every level of reconciliation in its own
task optimistically, so this does it only when we know that we'll throw
or log an error that needs this context. Unfortunately this doesn't
cover user space errors (such as if an iterable errors).
Before:
<img width="903" height="298" alt="Screenshot 2025-09-10 at 11 31 55 AM"
src="https://github.com/user-attachments/assets/cffc94da-8c14-4d6e-9a5b-bf0833b8b762"
/>
After:
<img width="1216" height="252" alt="Screenshot 2025-09-10 at 11 50
54 AM"
src="https://github.com/user-attachments/assets/f85f93cf-ab73-4046-af3d-dd93b73b3552"
/>
<img width="412" height="115" alt="Screenshot 2025-09-10 at 11 52 46 AM"
src="https://github.com/user-attachments/assets/a76cef7b-b162-4ecf-9b0a-68bf34afc239"
/>
When we emit objects of type `ReactAsyncInfo`, we need to make sure that
their owners are outlined, using `outlineComponentInfo`. Otherwise we
would end up accidentally emitting stashed fields that are not part of
the transport protocol, specifically `debugStack`, `debugTask`, and
`debugLocation`. This would lead to runtime errors in the client, when
for example, the stack for a `debugLocation` is processed in
`buildFakeCallStack`, but the stack was actually omitted from the RSC
payload, because for those fields we don't ensure that the object limit
is increased by the length of the stack, as we do when we're emitting
the `stack` of a `ReactComponentInfo` object in `outlineComponentInfo`.
One thing that can suspend is the downloading of the RSC stream itself.
This tracks an I/O entry for each Promise (`SomeChunk<T>`) that
represents the request to the RSC stream. As the value we use the
`Response` for `createFromFetch` (or the `ReadableStream` for
`createFromReadableStream`). The start time is when you called those.
Since we're not awaiting the whole stream, each I/O entry represents the
part of the stream up until it got unblocked. However, in a production
environment with TLS packets and buffering in practice the chunks
received by the client isn't exactly at the boundary of each row. It's a
bit longer into larger chunks. From testing, it seems like multiples of
16kb or 64kb uncompressed are common. To simulate a production
environment we group into roughly 64kb chunks if they happen in rapid
sequence. Note that this might be too small to give a good idea because
of the throttle many boundaries might be skipped anyway so this might
show too many.
The React DevTools will see each I/O entry as separate but dedupe if an
outer boundary already depends on the same chunk. This deduping makes it
so that small boundaries that are blocked on the same chunk, don't get
treated as having unique suspenders. If you have a boundary with large
content, then that content will likely be in a separate chunk which is
not in the parent and then it gets marked as.
This is all just an approximation. The goal of this is just to highlight
that very large boundaries will very likely suspend even if they don't
suspend on any I/O on the server. In practice, these boundaries can
float around a lot and it's really any Suspense boundary that might
suspend but some are more likely than others which this is meant to
highlight.
It also just lets you inspect how many bytes needs to be transferred
before you can show a particular part of the content, to give you an
idea that it's not just I/O on the server that might suspend.
If you don't use the debug channel it can be misleading since the data
in development mode stream will have a lot more data in it which leads
to more chunking.
Similarly to "client references" these I/O infos don't have an "env"
since it's the client that has the I/O and so those are excluded from
flushing in the Server performance tracks.
Note that currently the same Response can appear many times in the same
Instance of SuspenseNode in DevTools when there are multiple chunks. In
a follow up I'll show only the last one per Response at any given level.
Note that when a separate debugChannel is used it has its own I/O entry
that's on the `_debugInfo` for the debug chunks in that channel.
However, if everything works correctly these should never leak into the
DevTools UI since they should never be propagated from a debug chunk to
the values waited by the runtime. This is easy to break though.
This update was a bit more involved.
- `React$Component` was removed, I replaced it with Flow component
types.
- Flow removed shipping the standard library. This adds the environment
libraries back from `flow-typed` which seemed to have changed slightly
(probably got more precise and less `any`s). Suppresses some new type
errors.
This fixes an edge case where you abort the render while rendering a
component that ends up Suspending. It technically only applied if you
were deep enough to be inside `renderNode` and was not susceptible to
hanging if the abort + suspending component was being tried inside
retryRenderTask/retryReplaytask.
The fix is to preempt the thenable checks in renderNode and check if the
request is aborting and if so just bubble up to the task handler.
The reason this hung before is a new task would get scheduled after we
had aborted every other task (minus the currently rendering one). This
led to a situation where the task count would not hit zero.
This ensures that if the name is set manually after the declaration,
then we get that name when we log the value. For example Node.js
`Response` is declared as `_Response` and then later assigned a new
name.
We should probably really serialize all static enumerable properties but
"name" is non-enumerable so it's still a special case.
Fixes#33534.
`.then` method can be tested when you await a value that's not a
Promise. For regular Client References we have a way to mark those as
"async" and yield a reference to the unwrapped value in case it's a
Promise on the Client.
However, the realization is that we never serialize Promises as opaque
when passed from the client to the server. If a Promise is passed, then
it would've been deserialized as a Promise (while still registered as a
temporary reference) and not one of these Proxy objects.
Technically it could be a non-function value on the client which would
be wrong but you're not supposed to dot into it in the first place.
So we can just assume it's `undefined`.
<!--
Thanks for submitting a pull request!
We appreciate you spending the time to work on these changes. Please
provide enough information so that others can review your pull request.
The three fields below are mandatory.
Before submitting a pull request, please make sure the following is
done:
1. Fork [the repository](https://github.com/facebook/react) and create
your branch from `main`.
2. Run `yarn` in the repository root.
3. If you've fixed a bug or added code that should be tested, add tests!
4. Ensure the test suite passes (`yarn test`). Tip: `yarn test --watch
TestName` is helpful in development.
5. Run `yarn test --prod` to test in the production environment. It
supports the same options as `yarn test`.
6. If you need a debugger, run `yarn test --debug --watch TestName`,
open `chrome://inspect`, and press "Inspect".
7. Format your code with
[prettier](https://github.com/prettier/prettier) (`yarn prettier`).
8. Make sure your code lints (`yarn lint`). Tip: `yarn linc` to only
check changed files.
9. Run the [Flow](https://flowtype.org/) type checks (`yarn flow`).
10. If you haven't already, complete the CLA.
Learn more about contributing:
https://reactjs.org/docs/how-to-contribute.html
-->
## Summary
Fixes `await`-ing and returning temporary references in `async`
functions. These two operations invoke `.then()` under the hood if it is
available, which currently results in an "Cannot access then on the
server. You cannot dot into a temporary client reference..." error. This
can easily be reproduced by returning a temporary reference from a
server function.
Fixes#33534
## How did you test this change?
I added a test in a new test file. I wasn't sure where else to put it.
<img width="771" height="138" alt="image"
src="https://github.com/user-attachments/assets/09ffe6eb-271a-4842-a9fe-c68e17b3fb41"
/>
<!--
Demonstrate the code is solid. Example: The exact commands you ran and
their output, screenshots / videos if the pull request changes the user
interface.
How exactly did you verify that your PR solves the issue you wanted to
solve?
If you leave this empty, your PR will very likely be closed.
-->
Stacked on #34058
When tracking how large the shell is we currently only track the bytes
of everything above Suspense boundaries. However since Boundaries that
contribute to the preamble will always be inlined when the shell flushes
they should also be considered as part of the request byteSize since
they always flush alongside the shell. This change adds this tracking
Suspense boundaries that may have contributed to the preamble should not
be outlined due to size because these boundaries are only meant to be in
fallback state if the boundary actually errors. This change excludes any
boundary which has the potential to contribute to the preamble. We could
alternatively track which boundaries actually contributed to the
preamble but in practice there will be very few and I think this is
sufficient.
One problem with this approach is it makes Suspense above body opt out
of the mode where we omit rel="expect" for large shells. In essence
Suspense above body has the semantics of a Shell (it blocks flushing
until resolved) but it doesn't get tracked as request bytes and thus we
will not opt users into the skipped blocking shell for very large
boundaries.
This will be fixed in a followup
There's a lot of overlap between `enableComponentPerformanceTrack` and
`enableAsyncDebugInfo` because they both rely on timing information. The
former is mainly emit timestamps for how long server components and
awaits took. The latter how long I/O took.
`enableAsyncDebugInfo` is currently primarily for the component
performance track but its meta data is useful for other debug tools too.
This promotes that flag to stable.
However, `enableComponentPerformanceTrack` needs more work due to
performance concerns with Chrome DevTools so I need to separate them.
This keeps doing most of the timing tracking on the server but doesn't
emit the per-server component time stamps when
`enableComponentPerformanceTrack` is false.
There is an edge case when prerendering where if you have nothing to
write you can end up in a state where the prerender is in status closed
before you can provide a destination. In this case the destination is
never closed becuase it assumes it already would have been.
This condition can happen now because of the introduction of the deubg
stream. Before this a request would never entere closed status if there
was no active destination. When a destination was added it would perform
a flush and possibly close the stream. Now, it is possible to flush
without a destination because you might have debug chunks to stream and
you can end up closing the stream independent of an active destination.
There are a number of ways we can solve this but the one that seems to
adhere best to the original design is to only set the status to CLOSED
when a destination is active. This means that if you don't have an
active destination when the pendingChunks count hits zero it will not
enter CLOSED status until you startFlowing.
We need a "value" to represent the I/O that was loaded. We don't
normally actually use the Promise at the callsite that started the I/O
because that's usually deep inside internals. Instead we override the
value of the I/O entry with the Promise that was first awaited in user
space. This means that you could potentially have different values
depending on if multiple things await the same I/O. We just take one of
them. (Maybe we should actually just write the first user space awaited
Promise as the I/O entry? This might instead have other implications
like less deduping.)
When you pass a Promise forward, we may skip the awaits that happened in
earlier components because they're not part of the currently rendering
component. That's mainly for the stack and time stamps though. The value
is still probably conceptually the best value because it represents the
I/O value as far user space is concerned.
This writes the I/O early with the first await we find in user space
even if we're not going to use that particular await for the stack.
If you pass a promise to a client component to be rendered `<Client
promise={promise} />` then there's an internal await inside Flight.
There might also be user space awaits but those awaits may already have
happened before we render this component. Conceptually they were part of
the parent component and not this component. It's tricky to attribute
which await should be used for the stack in this case.
If we can't find an await we can use the JSX callsite as the stack
frame.
However, we don't want to do this for simple cases like if you return a
non-native Promise from a Server Component. Since that would now use the
stack of the thing that rendered the Server Component which is worse
than the stack of the I/O. To fix this, I update the
`debugOwner`/`debugTask`/`debugStack` when we start rendering inside the
Server Component. Conceptually these represent the "parent" component
and is used for errors referring to the parent like when we serialize
client component props the parent is the JSX of the client component.
However, when we're directly inside the Server Component we don't have a
callsite of the parent really. Conceptually it would be the return call
of the Server Component. This might negatively affect other types of
errors but I think this is ok since this feature mainly exists for the
case when you enter the child JSX.
This resolves an outstanding issue where it was possible for debug info
and console logs to become out of order if they up blocked. E.g. by a
future reference or a client reference that hasn't loaded yet. Such as
if you console.log a client reference followed by one that doesn't. This
encodes the order similar to how the stream chunks work.
This also blocks the main chunk from resolving until the last debug info
has fully loaded, including future references and client references.
This also ensures that we could send some of that data in a different
stream, since then it can come out of order.
We already do this with `"new Promise"` and `"Promise.then"`. There are
also many helpers that both create promises and awaits other promises
inside of it like `Promise.all`.
The way this is filtered is different from just filtering out all
anonymous stacks since they're used to determine where the boundary is
between ignore listed and user space.
Ideally we'd cover more wrappers that are internal to Promise libraries.
This lets us pass a writable on the server side and readable on the
client side to send debug info through a separate channel so that it
doesn't interfere with the main payload as much. The main payload refers
to chunks defined in the debug info which means it's still blocked on it
though. This ensures that the debug data has loaded by the time the
value is rendered so that the next step can forward the data.
This will be a bit fragile to race conditions until #33665 lands.
Another follow up needed is the ability to skip the debug channel on the
receiving side. Right now it'll block forever if you don't provide one
since we're blocking on the debug data.
When postponing the root we encode the segment Id into the postponed
state but we should really be reseting it to zero so we can restart the
counter from the beginning when the resume is actually just a re-render.
This also no longer assigns the root segment id based on the postponed
state when resuming the root for the same reason. In the future we may
use the embedded replay segment id if we implement resuming the root
without re-rendering everything but that is not yet implemented or
planned.
This is the same as we do for currently rendering tasks. They get
effectively sync aborted when the listener is invoked.
We potentially miss out on some debug info in that case but that would
only apply to any entries inside the stream which doesn't really have
their own debug info anyway.
Follow up to #33736.
If we need to save on CPU/memory pressure, we can instead just pray and
hope that a Promise doesn't get garbage collected before we need to read
it.
This can cause fragile access to the Promise value in devtools
especially if it's a slow and pressured render.
Basically, you'd have to hope that GC doesn't run after the inner await
finishes its microtask callback and before the resolution of the
component being rendered is invoked.
If we have the ability to lazy load Promise values, i.e. if we have a
debug channel, then we should always use it for Promises that aren't
already resolved and instrumented.
There's little downside to this since they're async anyway.
This also lets us avoid adding `.then()` listeners too early. E.g. if
adding the listener would have side-effect. This avoids covering up
"unhandled rejection" errors. Since if we listen to a promise eagerly,
including reject listeners, we'd have marked that Promise's rejection as
handled where as maybe it wouldn't have been otherwise.
In this mode we can also indefinitely wait for the Promise to resolve
instead of just waiting a microtask for it to resolve.
We use the stack of a Promise as the start of the I/O instead of the
actual I/O since that can symbolize the start of the operation even if
the actual I/O is batched, deduped or pooled. It can also group multiple
I/O operations into one.
We want the deepest possible Promise since otherwise it would just be
the Component's Promise.
However, we don't really need deeper than the boundary between first
party and third party. We can't just take the outer most that has third
party things on the stack though because third party can have callbacks
into first party and then we want the inner one. So we take the inner
most Promise that depends on I/O that has a first party stack on it.
The realization is that for the purposes of determining whether we have
a first party stack we need to ignore async stack frames. They can
appear on the stack when we resume third party code inside a resumption
frame of a first party stack.
<img width="832" alt="Screenshot 2025-07-08 at 6 34 25 PM"
src="https://github.com/user-attachments/assets/1636f980-be4c-4340-ad49-8d2b31953436"
/>
---------
Co-authored-by: Sebastian Sebbie Silbermann <sebastian.silbermann@vercel.com>
We don't really need to retain a reference to whatever Promise another
Promise was created in. Only awaits need to retain both their trigger
and their previous context.
When we know that the object that we pass in is immediately parsed, then
we know it couldn't have been reified into a unstructured stack yet. In
this path we assume that we'll trigger `Error.prepareStackTrace`.
Since we know that nobody else will read the stack after us, we can skip
generating a string stack and just return empty. We can also skip
caching.
If we're about to defer an object, then we shouldn't store a reference
to it because then we can end up deduping by referring to the deferred
string. If in a different context, we should still be able to emit the
object.
This is a compromise because there can be a lot of Promise instances
created. They're useful because they generally provide a better stack
when batching/pooled connections are used.
This restores stack collection for I/O nodes so we have something to
fallback on if there's no owner.
That way we can at least get a name or something out of I/O that was
spawned outside a render but mostly avoids collecting starting I/O
outside of render.
Because the object limit is unfortunately depth first due to limitations
of JSON stringify, we need to ensure that things we really don't want
outlined are first in the enumeration order.
We add the stack length to the object limit to ensure that the stack
frames aren't outlined. In console all the user space arguments are at
the end of the args. In server component props, the props are at the end
of the properties of the element.
For the `value` of I/O we had it before the stack so it could steal the
limit from the stack. The fix is to put it at the end.
We unnecessarily render the preamble in a task. This updates the
implementation to perform this render inline.
Testing this is tricky because one of the only ways you could assert
this was even happening is based on how things error if you abort while
rendering the root.
While adding a test for this I discovered that not all abortable tasks
report errors when aborted during a normal render. I've asserted the
current behavior and will address the other issue at another time and
updated the assertion later as necessary
When a debug channel is available, we now allow objects to be lazily
requested though the debug channel and only then will the server send
it.
The client will actually eagerly ask for the next level of objects once
it parses its payload. That way those objects have likely loaded by the
time you actually expand that deep e.g. in the console repl. This is
needed since the console repl is synchronous when you ask it to invoke
getters.
Each level is lazily parsed which means that we don't parse the next
level even though we eagerly loaded it. We parse it once the getter is
invoked (in Chrome DevTools you have to click a little `(...)` to invoke
the getter). When the getter is invoked, the chunk is initialized and
parsed. This then causes the next level to be asked for through the
debug channel. Ensuring that if you expand one more level you can do so
synchronously.
Currently debug chunks are eagerly parsed, which means that if you have
things like server component props that are lazy they can end up being
immediately asked for, but I'm trying to move to make the debug chunks
lazy.
We need to optimize the collection of debug info for dev mode. This is
an incredibly hot path since it instruments all I/O and Promises in the
app.
These optimizations focus primarily on the collection of stack traces.
They are expensive to collect because we need to eagerly collect the
stacks since they can otherwise cause memory leaks. We also need to do
some of the processing of them up front. We also end up only using a few
of them in the end but we don't know which ones we'll use.
The first compromise here is that I now only collect the stacks of
"awaits" if they were in a specific request's render. In some cases it's
useful to collect them even outside of this if they're part of a
sequence that started early. I still collect stacks for the created
Promises outside of this though which can still provide some context.
The other optimization to awaits, is that since we'll only use the inner
most one that had an await directly in userspace, we can stop collecting
stacks on a chain of awaits after we find one. This requires a quick
filter on a single callsite to determine. Since we now only collect
stacks from awaits that belongs to a specific Request we can use that
request's specific filter option. Technically this might not be quite
correct if that same thing ends up deduped across Requests but that's an
edge case.
Additionally, I now stop collecting stack for I/O nodes. They're almost
always superseded by the Promise that wraps them anyway. Even if you
write mostly Promise free code, you'll likely end up with a Promise at
the root of the component eventually anyway and then you end up using
its stack anyway. You have to really contort the code to end up with
zero Promises at which point it's not very useful anyway. At best it's
maybe mostly useful for giving a name to the I/O when the rest is just
stuff like `new Promise`.
However, a possible alternative optimization could be to *only* collect
the stack of spawned I/O and not the stack of Promises. The issue with
Promises (not awaits) is that we never know what will end up resolving
them in the end when they're created so we have to always eagerly
collect stacks. This could be an issue when you have a lot of
abstractions that end up not actually be related to I/O at all. The
issue with collecting stacks only for I/O is that the actual I/O can be
pooled or batched so you end up not having the stack when the conceptual
start of each operation within the batch started. Which is why I decided
to keep the Promise stack.
Content in Suspense fallbacks are really not considered part of the
Suspense but since it does have some behavior it should be marked
somehow separately from the Suspense content.
A follow up would be to do the same in Fiber.
Same as #33716 but without the separate close signal.
We'll need the ref count for separate debug channel anyway but I'm not
sure we'll need the separate close signal.
If I/O is not awaited in user space in a "previous" path we used to just
drop it on the floor. There's a few strategies we could apply here. My
first commit just emits it without an await but that would mean we don't
have an await stack when there's no I/O in a follow up.
I went with a strategy where the "previous" I/O is used only if the
"next" didn't have I/O. This may still drop I/O on the floor if there's
two back to back within internals for example. It would only log the
first one even though the outer await may have started earlier.
It may also log deeper in the "next" path if that had user space stacks
and then the outer await will appear as if it awaited after.
So it's not perfect.
When a `.then()` callback returns another Promise, there's effectively
another "await" on that Promise that happens in the internals but that
was not modeled. In effect the Promise returned by `.then()` is blocked
on both the original Promise AND the promise returned by the callback.
This models that by cloning the original node and treat that as the
await on the original Promise. Then we use the existing Node to await
the new Promise but its "previous" points to the clone. That way we have
a forked node that awaits both.
---------
Co-authored-by: Sebastian Sebbie Silbermann <sebastian.silbermann@vercel.com>
Hendrik Liebau
Sebastian Markbåge
Jan Kassens
Sebastian "Sebbie" Silbermann
Josh Story
Dennis Kats
Hiroshi Ogawa