Protobuf.js: Ask clarification about streaming implementation for rpc service

[sulliwane]

protobuf.js version: 6.7.1

I'm trying to implement streaming RPC function, to create services. But looking at the streaming example, I don't understand how I can emit these events:

greeter.on("data", function(response, method) { // data function
    console.log("data in " + method.name + ":", response.message);
});

greeter.on("end", function() { // end function
    console.log("end");
});

greeter.on("error", function(err, method) { // error function
    console.log("error in " + method.name + ":", err);
});

greeter.on("status", function(code, text) { // status function
    console.log("custom status:", code, text);
});

from the rpcImpl function:

// I only have a callback function. It's OK for unary method, but how to use for streaming method (data, end, error, status events)
var greeter = Greeter.create(function myRPCImpl(method, requestData, callback) { 
        responseData = fetch();
        callback(null, responseData);
    };
});

My guess is:

• data event: multiple call to callback(null, data).
• end event: calling callback(null, null).
• error event: calling callback(error).
• status event: ???

If that's correct, does it also mean that rpcImpl will automatically fire events when receiving a streaming method, and fire a promise when receiving a unary method? Or fire both whatever the method?

Thank you for the clarification!

[dcodeIO]

rpcImpl is just the low level method that implements a service request on the network level. It doesn't emit any events but solely uses a callback.

greeter is an instances of rpc.Service here (uses rpcImpl and extends an event emitter). That's where the events come from. MyService.create instantiates an rpc.Service.

The status event is a custom event present in this specific example only, to, well, demonstrate that one can also emit own events.

[jnordberg]

I just came here to create a similar issue (hi @dcodeIO 👋 😄)

I'm also confused by the streaming API. Shouldn't the rpc.Service do something different for streaming methods versus regular ones? Looks to me that a service can only have one method that is streaming and you can call the method only once before having to create a new rpc.Service instance.

Also I don't think the current implementation will work at all with promises, since once resolved the resolve callback is a no-op.

Something like this would make more sense to me:

let stream = myService.streamingMethod()
stream.on('data', (obj) => {
   console.log(obj) // decoded message
})
stream.on('end', () => console.log('ended'))
stream.on('error', (error) => throw error)

Could be done with a simple pseudo-stream interface (like a event emitter with write/end methods) or by using the readable-stream package which is fairly lightweight and very well tested.

[dcodeIO]

Nice to meet you again ;)

I'm also confused by the streaming API. Shouldn't the rpc.Service do something different for streaming methods versus regular ones? Looks to me that a service can only have one method that is streaming and you can call the method only once before having to create a new rpc.Service instance.

Individual methods do not return a stream, but the service instance's data event also references the method a message belongs to:

greeter.on("data", function(response, method) {
    console.log("data in " + method.name + ":", response.message);
});

Likewise, rpcImpl also knows which method a request is coming from: function myRPCImpl(method, requestData, callback).

This is what happens:

1. Call a streaming method
2. rpcImpl is called with method and the request message and sends it to the server (can be anything, but most likely over a websocket or something like that)
3. Server processes the request and responds
4. rpcImpl receives the response from the server and calls its callback
5. Service instance emits a data event with method and response message

Additionally, when rpcImpl's callback is called with a null message, the service instance assumes that the stream has ended and emits the end event.

This should also work for multiple methods. I must admit, though, that the current implementation is built around the desire of not adding another dependency besides a minimal event emitter and it might not be the greatest API of all time to work with.

Edit: I believe that ideally, protobuf.js itself wouldn't offer a service wrapper on its own. An additional package doing solely services might be better suited, and could contain more dependencies because it wouldn't affect the size of the core library.

[jnordberg]

Hmm, I still don't see how the API allows for a proper implementation of streams. Take this example:

service ImageService {
    rpc GetImage (ImageRequest) returns (stream ImageChunk) {}
}

message ImageRequest {
    string name = 1;
}

message ImageChunk {
    bytes data = 1;
}

let service = ImageService(function myRPCImpl(method, request, callback) { 
    let image = sometransport.createReadStream(request.name)
    image.on('data', (chunk) => { callback(null, {chunk}) })
    image.on('end', () => { callback(null) })
    image.on('error', (error) => { callback(error) })
})

let data = await myService.getImage({name: 'bar.jpg'}) // here I have only the first chunk

myService.on('data', (chunk, method) => {
   // here I don't know which image I asked so if I sent multiple getImage requests the chunks would be mixed up
})

// this would throw since myService.rpcImpl is set to null after a stream has finished
// but imagining I create a new service instance for every call I want to make this works:
myService.getImage({name: 'bar.jpg'}, (error, chunk) => {
   // even though calling a callback multiple times is considered bad practice
})

IMHO the rpc.Service class needs a bit of a refactor before working with streams become possible. Unless I'm completely misunderstanding how streaming services should work :)

[dcodeIO]

so if I sent multiple getImage requests the chunks would be mixed up

Ah, you are right, parallel requests using the same method don't work with the current design.

[jnordberg]

If you're interested I'd be happy to send a PR with a proposal for a better streaming API

[sulliwane]

this makes sense to me as well:

let stream = myService.streamingMethod(request)
stream.on('data', (obj) => {
   console.log(obj) // decoded message
})
stream.on('end', () => console.log('ended'))
stream.on('error', (error) => throw error)

I'd be happy to test/review any proposal :+1:

[sulliwane]

@jnordberg Did you come up with a solution ?

[jnordberg]

@sulliwane nope, just using non-streaming services for now...

[sulliwane]

Would be nice to have streaming working for rpc.Service :)

Especially since protobuf.js API is so much nicer to work with than google-protobuf

[aep]

any chance to get this going again?

[anjmao]

Hi, is anyone working on this issue? If not I could work on this. Let me know since I need server streaming support. Current RPCImpl is suitable only for simple unary calls. It is even more clear when using with Typescript

Example:

proto

service Debug {
  rpc GetServerStream(StreamReques) returns (stream StreamReply) {}
}

typescript generated method

public getServerStream(request: debug.IStreamRequest): Promise<debug.IStreamReply>;

Instead it should return Observable object or EventEmitter as @sulliwane pointed out.

interface Stream {
    on(eventName: 'data' | 'end' | 'error', cb: (data) => void): void;
}

public getServerStream(request: debug.IStreamRequest): Stream<debug.IStreamReply>;

Also we need separate RCPStreaImp

const rcp: RPCStreamImpl = (method, requestData, stream) => {
    underlingStream.start(method.name, requestData);
    underlingStream.on('data', (data) => stream.emit('data', data));
    underlingStream.on('end', () => stream.emit('end'));
    underlingStream.on('error', (err) => stream.emit('error', err));
};

PS.

RPCImpl doesn't allow to get service and pacakge name. I'm also thinking to change method to return name, service and package names.

[anjmao]

Hi, here is PR for streaming support, let me know what do you think https://github.com/dcodeIO/protobuf.js/pull/1115

[christian-ehmig]

Since I don't know when https://github.com/protobufjs/protobuf.js/pull/1115 will be accepted, here two workarounds, assuming you generated a protobufjs static file and wish to use a bidirectional stream:

Variant 1: inject custom RPC method definition into makeGenericClientConstructor, bidi stream

const grpc = require('grpc');
const messages = require('./pbjs-static.js').my.package; 

const MyService = exports.MyService = {  
  myCall: {
    path: '/my.package/myCall',
    requestStream: true,
    responseStream: true,
    requestType: messages.MyCallRequest,
    responseType: messages.MyCallResponse,
    requestSerialize: arg => messages.MyCallRequest.encode(arg).finish(), 
    responseDeserialize: arg => messages.MyCallResponse.decode(arg)
  }
};

const Client = grpc.makeGenericClientConstructor(MyService)
const client = new Client(
  grpcServerUrl,
  grpc.credentials.createInsecure()
)

// create bidi-stream
var stream = client.myCall();

// listen on incoming data (decoded via protobufjs)
stream.on('data', function(response){      
})

// send data (json is encoded via protobufjs)
stream.write({my: "message"});

Variant 2: directly create a bidi stream from makeBidiStreamRequest

const grpc = require('grpc');
const messages = require('./pbjs-static.js').my.package; 

const Client = grpc.makeGenericClientConstructor()
const client = new Client(
  grpcServerUrl,
  grpc.credentials.createInsecure()
)

var stream = client.makeBidiStreamRequest(
  '/my.package/myCall',      
  arg => messages.MyCallRequest.encode(arg).finish(),
  arg => messages.MyCallResponse.decode(arg)  
); 

// listen on incoming data (decoded via protobufjs)
stream.on('data', function(response){      
})

// send data (json is encoded via protobufjs)
stream.write({my: "message"});