Version
v24.15.0
Platform
Microsoft Windows NT 10.0.26200.0 x64
Subsystem
stream, webstreams
What steps will reproduce the bug?
When fs.WriteStream is converted using Writable.toWeb(), the promise returned
by writer.write() can settle before the file stream has consumed the supplied
Uint8Array.
This makes it unsafe to reuse a buffer after awaiting writer.write().
const fs = require('node:fs');
const os = require('node:os');
const path = require('node:path');
const { Writable } = require('node:stream');
(async () => {
const file = path.join(
os.tmpdir(),
`node-writable-to-web-${process.pid}.bin`,
);
const fileStream = fs.createWriteStream(file);
const writer = Writable.toWeb(fileStream).getWriter();
// Reusable scratch buffer.
const storage = new Uint8Array(4);
storage.set([1, 2, 3, 4]);
await writer.write(storage.subarray());
// The previous write promise has settled, so reuse the buffer.
storage.set([5, 6, 7, 8]);
await writer.write(storage.subarray());
await writer.close();
const content = fs.readFileSync(file);
console.log([...content]);
fs.unlinkSync(file);
})().catch(console.error);
Run:
$ node repro.js
[
5, 6, 7, 8,
5, 6, 7, 8
]
How often does it reproduce? Is there a required condition?
It reproduces consistently when:
- A mutable
Uint8Array or subarray() is written.
- The native
writable.write() call returns true.
- The backing buffer is reused after the Web Streams
write() promise settles.
- The native writable has not yet consumed the original bytes.
The problem can depend on chunk size. When the native write() returns false,
the adapter waits for backpressure, so larger chunks may appear to work.
The writable side returned by Duplex.toWeb() is affected by the same adapter
behavior.
What is the expected behavior? Why is that the expected behavior?
The file should contain the values supplied by the two completed writes:
[
1, 2, 3, 4,
5, 6, 7, 8
]
The producer mutates the buffer only after the first writer.write() promise
has settled.
The Web Streams specification advises producers not to mutate a mutable chunk
until the promise returned by write() settles:
https://streams.spec.whatwg.org/#default-writer-write
Following that guidance should ensure that the underlying sink processes the
same bytes that were passed to write().
What do you see instead?
The file contains the second value twice:
[
5, 6, 7, 8,
5, 6, 7, 8
]
The first writer.write() promise settles while fs.WriteStream still retains
a reference to the first subarray(). Reusing the backing buffer therefore also
changes the bytes of the pending file write.
Changing subarray() to slice() avoids the corruption because slice() copies
the bytes, but it requires an allocation and copy for every write.
What do you see instead?
The boolean returned by native writable.write() only represents backpressure.
It does not indicate that the specific chunk has finished being processed.
Node's native writable API provides a per-write callback:
writable.write(chunk, callback);
https://nodejs.org/api/stream.html#writablewritechunk-encoding-callback
A possible fix would be for the Web Streams adapter to:
- Settle the Web
write() promise from the native per-write callback.
- If native
write() returns false, also honor the drain event.
- Reject the Web write promise if the callback reports an error.
This would allow callers to safely reuse mutable buffers after awaiting
writer.write() without requiring slice() copies.
Additional information
No response
Version
v24.15.0
Platform
Subsystem
stream, webstreams
What steps will reproduce the bug?
When
fs.WriteStreamis converted usingWritable.toWeb(), the promise returnedby
writer.write()can settle before the file stream has consumed the suppliedUint8Array.This makes it unsafe to reuse a buffer after awaiting
writer.write().Run:
How often does it reproduce? Is there a required condition?
It reproduces consistently when:
Uint8Arrayorsubarray()is written.writable.write()call returnstrue.write()promise settles.The problem can depend on chunk size. When the native
write()returnsfalse,the adapter waits for backpressure, so larger chunks may appear to work.
The writable side returned by
Duplex.toWeb()is affected by the same adapterbehavior.
What is the expected behavior? Why is that the expected behavior?
The file should contain the values supplied by the two completed writes:
The producer mutates the buffer only after the first
writer.write()promisehas settled.
The Web Streams specification advises producers not to mutate a mutable chunk
until the promise returned by
write()settles:https://streams.spec.whatwg.org/#default-writer-write
Following that guidance should ensure that the underlying sink processes the
same bytes that were passed to
write().What do you see instead?
The file contains the second value twice:
The first
writer.write()promise settles whilefs.WriteStreamstill retainsa reference to the first
subarray(). Reusing the backing buffer therefore alsochanges the bytes of the pending file write.
Changing
subarray()toslice()avoids the corruption becauseslice()copiesthe bytes, but it requires an allocation and copy for every write.
What do you see instead?
The boolean returned by native
writable.write()only represents backpressure.It does not indicate that the specific chunk has finished being processed.
Node's native writable API provides a per-write callback:
https://nodejs.org/api/stream.html#writablewritechunk-encoding-callback
A possible fix would be for the Web Streams adapter to:
write()promise from the native per-write callback.write()returnsfalse, also honor thedrainevent.This would allow callers to safely reuse mutable buffers after awaiting
writer.write()without requiringslice()copies.Additional information
No response