Stream: git-wasmtime

Topic: wasmtime / PR #13418 Fix: cross component stream copy

Wasmtime GitHub notifications bot (May 20 2026 at 14:43):

Aditya1404Sal opened PR #13418 from Aditya1404Sal:fix-cross-component-stream-copy to bytecodealliance:main:

Summary

Fix cross-component async stream / future copy logic so writer-side payload metadata and memory are resolved from the writer component instance, while reader-side payload metadata and memory are resolved from the reader component instance.

Previously, Instance::copy resolved both write_ty and read_ty against self.component().types(). That works when both endpoints are inside one composed component graph, but breaks when the stream crosses dynamically linked component instances with distinct ComponentTypes tables.

Problem

When a guest-created stream<T> is returned by one component and passed into another dynamically linked component, the stream handle can be valid while the payload copy uses the wrong component type table or memory/options.

The same issue can also surface as internal panics such as:

BUG: expected write payload type to be present

Fix

Instance::copy now receives both endpoint instances:

• write_instance for writer payload ABI, writer memory, and LiftContext
• read_instance for reader payload ABI, reader memory, and LowerContext

The guest write/read rendezvous call sites now pass the correct endpoint instances through to copy.

This also fixes partial stream writes by preserving the original write_instance when re-storing a pending WriteState::GuestReady; otherwise subsequent reads of the remaining write buffer could accidentally use the reader instance.

Validation

Only Validated locally with wasmCloud’s examples/patch-stream a setup (which I will promptly create for wasmtime), where:

• a producer component returns stream<u8>
• a consumer component forwards that stream to a third component
• the sink component drains and logs each NDJSON patch line

Before this fix, the sink saw corrupted binary-looking data. After this fix, it receives all expected patch lines and the stream closes cleanly:

Also ran:

cargo check -p wasmtime --features component-model,component-model-async

Wasmtime GitHub notifications bot (May 20 2026 at 14:56):

Aditya1404Sal edited PR #13418:

Summary

Fix cross-component async stream / future copy logic so writer-side payload metadata and memory are resolved from the writer component instance, while reader-side payload metadata and memory are resolved from the reader component instance.

Previously, Instance::copy resolved both write_ty and read_ty against self.component().types(). That works when both endpoints are inside one composed component graph, but breaks when the stream crosses dynamically linked component instances with distinct ComponentTypes tables.

Problem

When a guest-created stream<T> is returned by one component and passed into another dynamically linked component, the stream handle can be valid while the payload copy uses the wrong component type table or memory/options.

The same issue can also surface as internal panics such as:

BUG: expected write payload type to be present

Fix

Instance::copy now receives both endpoint instances:

• write_instance for writer payload ABI, writer memory, and LiftContext
• read_instance for reader payload ABI, reader memory, and LowerContext

The guest write/read rendezvous call sites now pass the correct endpoint instances through to copy.

This also fixes partial stream writes by preserving the original write_instance when re-storing a pending WriteState::GuestReady; otherwise subsequent reads of the remaining write buffer could accidentally use the reader instance.

Validation

Only Validated locally with wasmCloud’s examples/patch-stream a setup (which I will promptly create for wasmtime), where:

• a producer component returns stream<u8>
• a consumer component forwards that stream to a third component
• the sink component drains and logs each NDJSON patch line

Before this fix, the sink saw corrupted binary-looking data. After this fix, it receives all expected patch lines and the stream closes cleanly:

Also ran:

cargo check -p wasmtime --features component-model,component-model-async

Wasmtime GitHub notifications bot (May 20 2026 at 15:02):

Aditya1404Sal has marked PR #13418 as ready for review.

Wasmtime GitHub notifications bot (May 20 2026 at 15:02):

Aditya1404Sal requested pchickey for a review on PR #13418.

Wasmtime GitHub notifications bot (May 20 2026 at 15:02):

Aditya1404Sal requested wasmtime-core-reviewers for a review on PR #13418.

Wasmtime GitHub notifications bot (May 20 2026 at 17:42):

github-actions[bot] added the label wasmtime:api on PR #13418.

Wasmtime GitHub notifications bot (May 20 2026 at 18:20):

pchickey unassigned pchickey from PR #13418 Fix: cross component stream copy.

Wasmtime GitHub notifications bot (May 20 2026 at 18:20):

pchickey requested alexcrichton for a review on PR #13418.

Wasmtime GitHub notifications bot (May 20 2026 at 18:21):

:memo: pchickey submitted PR review:

I dont have the context at the moment to be able to deeply review this. @alexcrichton is traveling this week for a conference, so his review may take a bit, but I've re-assigned to him.

Wasmtime GitHub notifications bot (May 20 2026 at 22:14):

:memo: dicej submitted PR review:

Thanks, @Aditya1404Sal. I'll admit I hadn't considered the possibility of passing the read end of a stream or future to a different top-level component instance than the write end, but it does make sense.

Would you be up for adding a test for this scenario, e.g. a new test in crates/misc/component-async-tests/tests/scenario/streams.rs?

Wasmtime GitHub notifications bot (May 20 2026 at 22:14):

:speech_balloon: dicej created PR review comment:

Does it still make sense to take a self parameter here? It doesn't seem to be used anymore, and would presumably be a mistake to use instead of write_instance and read_instance.

Wasmtime GitHub notifications bot (May 20 2026 at 22:14):

:speech_balloon: dicej edited PR review comment.

Wasmtime GitHub notifications bot (May 21 2026 at 19:51):

Aditya1404Sal updated PR #13418.

Wasmtime GitHub notifications bot (May 21 2026 at 20:00):

Aditya1404Sal commented on PR #13418:

@dicej thanks for the review! I've added a regression test, lmk if there's anything else.

Wasmtime GitHub notifications bot (May 21 2026 at 20:01):

Aditya1404Sal edited a comment on PR #13418:

Thanks for the review @dicej -- I've added a regression test, lmk if there's anything else.

Wasmtime GitHub notifications bot (May 21 2026 at 20:23):

:thumbs_up: dicej submitted PR review:

Great work, thanks!

Wasmtime GitHub notifications bot (May 21 2026 at 20:23):

dicej added PR #13418 Fix: cross component stream copy to the merge queue.

Wasmtime GitHub notifications bot (May 21 2026 at 20:47):

:check: dicej merged PR #13418.

Wasmtime GitHub notifications bot (May 21 2026 at 20:47):

dicej removed PR #13418 Fix: cross component stream copy from the merge queue.

Last updated: Jun 01 2026 at 09:49 UTC