search

bytecodealliance / wrpc

Wasm component-native RPC framework
Other
118 stars 19 forks source link

examples/rust: add echo-stream-nats-{client,server} #217

Closed jvatic closed 1 month ago

jvatic commented 1 month ago

This is essentially a copy of the hello-nats-{client,server} example, but with a bi-directional stream (server echos the client) that came out of me wanting to know if this functionality works.

There are a few things I'd like to note:

  1. The streams in the WIT definition are stream<u64> but the bindings are for Vec<u64>.
  2. The output of echo on the client is a tuple with a stream and a future, where the future has to be polled before the stream will work. This feels clunky; could the future be implemented as part of the stream?
  3. wit_bindgen::generate! doesn't implement stream (or future) yet, so I'm not sure if implementing this as a component example is feasible?

I'd like to get this working with a component example, so I'd be happy to help out with any of the above.

rvolosatovs commented 1 month ago

rebased and fixed-up a small merge conflict in Cargo.toml

rvolosatovs commented 1 month ago

1. The streams in the WIT definition are stream<u64> but the bindings are for Vec<u64>.

indeed, that's because wRPC does not handle chunking and it's expected that implementations would handle that instead, I believe https://docs.rs/futures/latest/futures/stream/trait.StreamExt.html#method.ready_chunks should be able to be used out-of-the-box here on sender side

2. The output of echo on the client is a tuple with a stream and a future, where the future has to be polled before the stream will work. This feels clunky; could the future be implemented as part of the stream?

The intention here is to be able to poll both the stream and the I/O future concurrently for most use cases providing for clean separation of data and I/O error handling. I think we need to have both separate for lower-level use cases, however I certainly agree that providing a wrapper, which could take these two and return a single future/stream would be ideal.

3. wit_bindgen::generate! doesn't implement stream (or future) yet, so I'm not sure if implementing this as a component example is feasible?

That's a bit tricky, async support in component model is currently WIP (latest draft at https://github.com/WebAssembly/component-model/blob/main/design/mvp/Async.md, refs https://github.com/WebAssembly/component-model/pull/363). I believe there is a set of patches for Wasmtime, which would allow using async in components, but I have not explored that myself yet. I also don't think we should depend on an unreleased version of Wasmtime in wRPC (we could, however, add an additional runtime using one)

Other than that, I'm thinking to work on a wRPC stream resource and a small reflection API to go along with it, which would allow this functionality to be used by any component. Btw, I'm currently tackling #185 , which is tangentially related.

Would you perhaps be interested in working on (beginnings of) a reflection API? Basically define all WIT kinds in WIT and perhaps have some way to "construct" a record type at runtime?

rvolosatovs commented 1 month ago

Small follow-up, I suggest to take a look at the slightly reworked stream usage at https://github.com/bytecodealliance/wrpc/blob/6ea5b37ec48cc1211723ab129e23c62c2ff358d1/examples/rust/streams-nats-client/src/main.rs#L53-L95 to see examples of batching and I/O future handling

jvatic commented 1 month ago

Small follow-up, I suggest to take a look at the slightly reworked stream usage at

https://github.com/bytecodealliance/wrpc/blob/6ea5b37ec48cc1211723ab129e23c62c2ff358d1/examples/rust/streams-nats-client/src/main.rs#L53-L95 to see examples of batching and I/O future handling

Ah, yes, this certainly makes sense, thanks!

  1. wit_bindgen::generate! doesn't implement stream (or future) yet, so I'm not sure if implementing this as a component example is feasible?

That's a bit tricky, async support in component model is currently WIP (latest draft at https://github.com/WebAssembly/component-model/blob/main/design/mvp/Async.md, refs WebAssembly/component-model#363). I believe there is a set of patches for Wasmtime, which would allow using async in components, but I have not explored that myself yet. I also don't think we should depend on an unreleased version of Wasmtime in wRPC (we could, however, add an additional runtime using one)

Other than that, I'm thinking to work on a wRPC stream resource and a small reflection API to go along with it, which would allow this functionality to be used by any component. Btw, I'm currently tackling #185 , which is tangentially related.

Would you perhaps be interested in working on (beginnings of) a reflection API? Basically define all WIT kinds in WIT and perhaps have some way to "construct" a record type at runtime?

Thanks! I'm new to the WIT ecosystem, and while I'm not entirely sure where to get started on such a reflect API, I'd be interested in working on that given a bit more context. Is the idea to provide non-blocking/blocking APIs within a sync component for working with a stream similar to wasi::io streams that could evolve for async usage?

And would it be possible currently to use wasi::io streams in place of stream<u8> with wRPC? (If so I'm thinking that could be a viable alternative for my use-case for now.)