Ensures Flight responses are not loaded as HTML.
## Bug
- [ ] Related issues linked using `fixes #number`
- [ ] Integration tests added
- [ ] Errors have a helpful link attached, see `contributing.md`
## Feature
- [ ] Implements an existing feature request or RFC. Make sure the feature request has been accepted for implementation before opening a PR.
- [ ] Related issues linked using `fixes #number`
- [ ] Integration tests added
- [ ] Documentation added
- [ ] Telemetry added. In case of a feature if it's used or not.
- [ ] Errors have a helpful link attached, see `contributing.md`
## Documentation / Examples
- [ ] Make sure the linting passes by running `pnpm lint`
- [ ] The "examples guidelines" are followed from [our contributing doc](https://github.com/vercel/next.js/blob/canary/contributing/examples/adding-examples.md)
Implements https://github.com/vercel/next.js/issues/30997 with some minor tweaks to the design:
* The hook is moved to Client Components (e.g. `pages/_app` instead of `pages/_document`). This was a silly oversight in the original design: the hook needs to be called during server prerendering.
* `useFlushEffects` instead of `useFlushEffect` as there isn't a particularly safe way to implement the singular semantics as a Client Component hook given the current implementation of server rendering.
---
Fixes#30997
Since we're always using `ReadableStream`, we should just get rid of `ResultPiper`.
This also lets us replace things like `bufferedReadFromReadableStream` with a `TransformStream` that does the same thing, so that it's `TransformStream`s all the way down.
Finally, we can get rid of the one-off call to `renderToReadableStream` and just use `renderToStream` whenever we're rendering a concurrent tree.
Instead of branching rendering based on Node.js and browser/web runtimes, we should just use the web version for now, which can run as-is on versions >=16.5.0 of Node.js, polyfilling `ReadableStream` on older versions when necessary.
There are a few potential downsides to this, as React is less able to optimize flushing and execution. We can revisit that in the future though if desired.
Use `Writable` instead of `Observable` and remove the `zen-observable` dependencies. I initially opted to use `Observable` for simplicity and fast iteration, but we should really just use `Writable` directly (or some other stream in the future).
React's streaming SSR has some [specific requirements](https://github.com/reactwg/react-18/discussions/66#discussioncomment-944266) on the stream API. Rather than trying to also squeeze a `Readable` in here, which might be more standard for node apps, I've just followed React's lead. By limiting ourselves to just `Writable`, it ought to be easier to adopt a different stream type in the future if desired.
The React `pipeToNodeWritable` API requires us to pass a stream immediately, but we don't actually have a `ServerResponse` to give it until `RenderResult.pipe(...)` is called later. For that reason, we pass React a `Writable` that we will simply forward to `res` later. This mechanism of deferring is `NodeWritablePiper`, which is just a function that can be called with `ServerResponse` (or another `Writable`, as we now do to render to string for static results) to have content written to it. `NodeWritablePiper` takes a `next` argument so that we can chain both synchronous and asynchronous pipers together.
Also does some clean up and adds another streaming test for backpressure.
We're no longer currently planning on supporting caching for dynamic responses, so we can do some cleaning & simplification:
* Multiplexing can be removed since we only ever subscribe once (via `RenderResult.pipe`, described below)
* `RenderResult.toUnchunkedString` can become synchronous since static responses are never chunked
* `RenderResult.forEach` can become `RenderResult.pipe` which helps encapsulate some of the details of `RenderResult`