As an alternative to #31784.
We should really just always yield each virtual task to a native task.
So that it's 1:1 with native tasks. This affects when microtasks within
each task happens. This brings us closer to native `postTask` semantics
which makes it more seamless to just use that when available.
This still doesn't yield when a task expires to protect against
starvation.
When implementing passive effects we did a pretty massive oversight.
While the passive effect is scheduled into its own scheduler task, the
scheduler doesn't always yield to the browser if it has time left. That
means that if you have a fast commit phase, it might try to squeeze in
the passive effects in the same frame but those then might end being
very heavy.
We had `requestPaint()` for this but that was only implemented for the
`isInputPending` experiment. It wasn't thought we needed it for the
regular scheduler because it yields "every frame" anyway - but it
doesn't yield every task. While the `isInputPending` experiment showed
that it wasn't actually any significant impact, and it was better to
keep shorter yield time anyway. Which is why we deleted the code.
Whatever small win it did see in some cases might have been actually due
to this issue rather than anything to do with `isInputPending` at all.
As you can see in https://github.com/facebook/react/pull/31782 we do
have this implemented in the mock scheduler and a lot of behavior that
we assert assumes that this works.
So this just implements yielding after `requestPaint` is called.
Before:
<img width="1023" alt="Screenshot 2024-12-14 at 3 40 24 PM"
src="https://github.com/user-attachments/assets/d60f4bb2-c8f8-4f91-a402-9ac25b278450"
/>
After:
<img width="1108" alt="Screenshot 2024-12-14 at 3 41 25 PM"
src="https://github.com/user-attachments/assets/170cdb90-a049-436f-9501-be3fb9bc04ca"
/>
Notice how in after the native task is split into two. It might not
always actually paint and the native scheduler might make the same
mistake and think it has enough time left but it's at least less likely
to.
We do have another way to do this. When we yield a continuation we also
yield to the native browser. This is to enable the Suspense Optimization
(currently disabled) to work. We could do the same for passive effects
and, in fact, I have a branch that does but because that requires a lot
more tests to be fixed it's a lot more invasive of a change. The nice
thing about this approach is that this is not even running in tests at
all and the tests we do have assert that this is the behavior already. 😬
## Overview
Updates `eslint-plugin-jest` and enables the recommended rules with some
turned off that are unhelpful.
The main motivations is:
a) we have a few duplicated tests, which this found an I deleted
b) making sure we don't accidentally commit skipped tests
In React 19 React will finally stop publishing UMD builds. This is
motivated primarily by the lack of use of UMD format and the added
complexity of maintaining build infra for these releases. Additionally
with ESM becoming more prevalent in browsers and services like esm.sh
which can host React as an ESM module there are other options for doing
script tag based react loading.
This PR removes all the UMD build configs and forks.
There are some fixtures that still have references to UMD builds however
many of them already do not work (for instance they are using legacy
features like ReactDOM.render) and rather than block the removal on
these fixtures being brought up to date we'll just move forward and fix
or removes fixtures as necessary in the future.
## Summary
`isInputPending` is not in use. This PR cleans up the flags controlling
its gating and parameters to simplify Scheduler.
Makes `frameYieldMs` feature flag static, set to 10ms in www, which we
found built on the wins provided by a broader yield interval via
`isInputPending`. Flag remains set to 5ms in OSS builds.
## How did you test this change?
`yarn test Scheduler`
## Summary
Follow up from #27717 based on feedback to rename the fork module itself
## How did you test this change?
- `yarn build`
- `yarn test
packages/scheduler/src/__tests__/SchedulerUMDBundle-test.internal.js`
Co-authored-by: Jack Pope <jackpope@meta.com>
## Summary
passing both a signal and a priority to `postTask`/`yield` in chrome
causes memory to spike and potentially causes OOMs. a fix for this has
landed in chrome 118, but we can avoid the issue in earlier versions by
setting priority on just the TaskController instead.
https://bugs.chromium.org/p/chromium/issues/detail?id=1469367
## How did you test this change?
```
yarn test SchedulerPostTask
```
## Summary
`scheduler.yield` is entering [Origin Trial soon in Chrome
115](https://chromestatus.com/feature/6266249336586240). This diff adds
it to `SchedulerPostTask` when scheduling continuations to allow Origin
Trial participation for early feedback on the new API.
It seems the difference here versus the current use of `postTask` will
be minor – the intent behind `scheduler.yield` seems to mostly be better
ergonomics for scheduling continuations, but it may be interesting to
see if the follow aspect of it results in any tangible difference in
scheduling (from
[here](https://github.com/WICG/scheduling-apis/blob/main/explainers/yield-and-continuation.md#introduction)):
> To mitigate yielding performance penalty concerns, UAs prioritize
scheduler.yield() continuations over tasks of the same priority or
similar task sources.
## How did you test this change?
```
yarn test SchedulerPostTask
```
To wait for the microtask queue to empty, our internal test helpers
schedule an arbitrary task using `setImmediate`. It doesn't matter what
kind of task it is, only that it's a separate task from the current one,
because by the time it fires, the microtasks for the current event will
have already been processed.
The issue with `setImmediate` is that Jest mocks it. Which can lead to
weird behavior.
I've changed it to instead use a message event, via the MessageChannel
implementation exposed by the `node:worker_threads` module.
We should consider doing this in the public implementation of `act`,
too.
(This only affects our own internal repo; it's not a public API.)
I think most of us agree this is a less confusing name. It's possible
someone will confuse it with `console.log`. If that becomes a problem we
can warn in dev or something.
This converts some of our test suite to use the `waitFor` test pattern,
instead of the `expect(Scheduler).toFlushAndYield` pattern. Most of
these changes are automated with jscodeshift, with some slight manual
cleanup in certain cases.
See #26285 for full context.
This converts some of our test suite to use the `waitFor` test pattern,
instead of the `expect(Scheduler).toFlushAndYield` pattern. Most of
these changes are automated with jscodeshift, with some slight manual
cleanup in certain cases.
See #26285 for full context.
## Summary
- yarn.lock diff +-6249, **small pr**
- use jest-environment-jsdom by default
- uncaught error from jsdom is an error object instead of strings
- abortSignal.reason is read-only in jsdom and node,
https://developer.mozilla.org/en-US/docs/Web/API/AbortSignal/reason
## How did you test this change?
ci green
---------
Co-authored-by: Sebastian Silbermann <silbermann.sebastian@gmail.com>
The old version of prettier we were using didn't support the Flow syntax
to access properties in a type using `SomeType['prop']`. This updates
`prettier` and `rollup-plugin-prettier` to the latest versions.
I added the prettier config `arrowParens: "avoid"` to reduce the diff
size as the default has changed in Prettier 2.0. The largest amount of
changes comes from function expressions now having a space. This doesn't
have an option to preserve the old behavior, so we have to update this.
* Facebook -> Meta in copyright
rg --files | xargs sed -i 's#Copyright (c) Facebook, Inc. and its affiliates.#Copyright (c) Meta Platforms, Inc. and affiliates.#g'
* Manual tweaks
* Yield to main thread if continuation is returned
Instead of using an imperative method `requestYield` to ask Scheduler to
yield to the main thread, we can assume that any time a Scheduler task
returns a continuation callback, it's because it wants to yield to the
main thread. We can assume the task already checked some condition that
caused it to return a continuation, so we don't need to do any
additional checks — we can immediately yield and schedule a new task
for the continuation.
The replaces the `requestYield` API that I added in ca990e9.
* Move unwind after error into main work loop
I need to be able to yield to the main thread in between when an error
is thrown and when the stack is unwound. (This is the motivation behind
the refactor, but it isn't implemented in this commit.) Currently the
unwind is inlined directly into `handleError`.
Instead, I've moved the unwind logic into the main work loop. At the
very beginning of the function, we check to see if the work-in-progress
is in a "suspended" state — that is, whether it needs to be unwound. If
it is, we will enter the unwind phase instead of the begin phase.
We only need to perform this check when we first enter the work loop:
at the beginning of a Scheduler chunk, or after something throws. We
don't need to perform it after every unit of work.
* Yield to main thread whenever a fiber suspends
When a fiber suspends, we should yield to the main thread in case the
data is already cached, to unblock a potential ping event.
By itself, this commit isn't useful because we don't do anything special
in the case where to do receive an immediate ping event. I've split this
out only to demonstrate that it doesn't break any existing behavior.
See the next commit for full context and motivation.
* Resume immediately pinged fiber without unwinding
If a fiber suspends, and is pinged immediately in a microtask (or a
regular task that fires before React resumes rendering), try rendering
the same fiber again without unwinding the stack. This can be super
helpful when working with promises and async-await, because even if the
outermost promise hasn't been cached before, the underlying data may
have been preloaded. In many cases, we can continue rendering
immediately without having to show a fallback.
This optimization should work during any concurrent (time-sliced)
render. It doesn't work during discrete updates because those are
semantically required to finish synchronously — those get the current
behavior.
We need a way to yield control of the main thread and schedule a
continuation in a separate macrotask. (This is related to some Suspense
optimizations we have planned.)
Our solution needs account for how Scheduler is implemented. Scheduler
tasks are not 1:1 with real browser macrotasks — many Scheduler "tasks"
can be executed within a single browser task. If a Scheduler task yields
control and posts a continuation, but there's still time left in the
frame, Scheduler will execute the continuation immediately
(synchronously) without yielding control back to the main thread. That's
not what we want — we want to schedule a new macrotask regardless of
where we are in the browser's render cycle.
There are several ways we could approach this. What I ended up doing was
adding a new Scheduler method `unstable_requestYield`. (It's similar to
the existing `unstable_requestPaint` that we use to yield at the end of
the frame.)
It works by setting the internal start time of the current work loop to
a large negative number, so that when the `shouldYield` call computes
how much time has elapsed, it's guaranteed to exceed the deadline. The
advantage of doing it this way is that there are no additional checks in
the normal hot path of the work loop.
The existing layering between Scheduler and React DOM is not ideal. None
of the APIs are public, so despite the fact that Scheduler is a separate
package, I consider that a private implementation detail, and think of
them as part of the same unit.
So for now, though, I think it makes sense to implement this macrotask
logic directly inside of Scheduler instead of layering it on top.
The rough eventual plan for Scheduler is turn it into a `postTask`
prollyfill. Because `postTask` does not yet have an equivalent for
`shouldYield`, we would split that out into its own layer, perhaps
directly inside the reconciler. In that world, the macrotask logic I've
added in this commit would likely live in that same layer. When the
native `postTask` is available, we may not even need any additional
logic because it uses actual browser tasks.
* Clean up Scheduler forks
* Un-shadow variables
* Use timer globals directly, add a test for overrides
* Remove more window references
* Don't crash for undefined globals + tests
* Update lint config globals
* Fix test by using async act
* Add test fixture
* Delete test fixture
* Move direct port access into a function
* Fork based on presence of setImmediate
* Copy SchedulerDOM-test into another file
* Change the new test to use shimmed setImmediate
* Clarify comment
* Fix test to work with existing feature detection
* Add flags
* Disable OSS flag and skip tests
* Use VARIANT to reenable tests
* lol
`shouldYield` will currently return `true` if there's a higher priority
task in the Scheduler queue.
Since we yield every 5ms anyway, this doesn't really have any practical
benefit. On the contrary, the extra checks on every `shouldYield` call
are wasteful.
Because `postTask` returns a promise, errors inside a `postTask`
callback result in the promise being rejected.
If we don't catch those errors, then the browser will report an
"Unhandled promise rejection" error. This is a confusing message to see
in the console, because the fact that `postTask` is a promise-based API
is an implementation detail from the perspective of the developer.
"Promise rejection" is a red herring.
On the other hand, if we do catch those errors, then we need to report
the error to the user in some other way.
What we really want is the default error reporting behavior that a
normal, non-Promise browser event gets.
So, we'll re-throw inside `setTimeout`.
This updates the experimental Scheduler postTask build to call postTask
directly, instead of managing our own custom queue and work loop.
We still use a deadline 5ms mechanism to implement `shouldYield`.
The main thing that postTask is currently missing is the continuation
feature — when yielding to the main thread, the yielding task is sent
to the back of the queue, instead of maintaining its position.
While this would be nice to have, even without it, postTask may be good
enough to replace our userspace implementation.
We'll run some tests to see.
* Get current time from performance.now in non-DOM environments
* Use local references to native APIs for Date and Performance
* Refactored to read globals directly
* Reduce code to necessities
* Switch to postTask API
* Add SchedulerPostTask tests
* Updates from review
* Fix typo from review
* Generate build of unstable_post_task
Since the Lanes refactor landed, we no longer rely on this anywhere, so
we can remove it.
The `delay` option is still needed by our timer implementation
(setTimeout polyfill). We'll keep the feature, but we'll likely change
how it's exposed once we figure out the proper layering between the
various Scheduler APIs.
* Enable prefer-const rule
Stylistically I don't like this but Closure Compiler takes advantage of
this information.
* Auto-fix lints
* Manually fix the remaining callsites
The mock Scheduler that we use in our tests has its own fake timer
implementation. The `unstable_advanceTime` method advances the timeline.
Currently, a call to `unstable_advanceTime` will also flush any pending
expired work. But that's not how it works in the browser: when a timer
fires, the corresponding task is added to the Scheduler queue. However,
we will still wait until the next message event before flushing it.
This commit changes `unstable_advanceTime` to more closely resemble the
browser behavior, by removing the automatic flushing of expired work.
```js
// Before this commit
Scheduler.unstable_advanceTime(ms);
// Equivalent behavior after this commit
Scheduler.unstable_advanceTime(ms);
Scheduler.unstable_flushExpired();
```
The general principle is to prefer separate APIs for scheduling tasks
and flushing them.
This change does not affect any public APIs. `unstable_advanceTime` is
only used by our own test suite. It is not used by `act`.
However, we may need to update tests in www, like Relay's.
The `performance.now` returns a timestamp in milliseconds as a float.
The browser has the option to adjust the precision of the float, but
it's usually more precise than a millisecond. However, this precision
is lost when the timestamp is logged by the Scheduler profiler, because
we store the numbers in an Int32Array.
This change multiplies the millisecond float value by 1000, giving us
three more degrees of precision.
Sebastian Markbåge
Ricky
Jan Kassens
Josh Story
sweetliquid
Noah Lemen
Andrew Clark
Jack PopeandJack Pope
Ming Ye
Brian Vaughn
Dan Abramov
Emilis Baliukonis
Robert Ferentz
Rick Hanlonandshengxinjing
Sebastian Markbåge
Heaven