tchannel .call api change

gistfile1.py

# raw
response = yield tchannel.call(
    argscheme="raw",
    service="someservice",
    endpoint="something",
    body="RAWRESPONSE",
    headers={
        'X-Source': 'geo'   
    },
    ttl=1000
)

# json
response = yield tchannel.call(
    argscheme = "json"
    service="someservice",
    endpoint="maps",
    body="{
        'lat': 100,
        'lng': 140
    }",
    headers={
        'X-Source': 'geo'   
    },
    ttl=1000
)

# thrift
Foo = get_tchannel_thrift(FooThrift, 'fooservice')

response = yield tchannel.call(
    argscheme="thrift",
    "some_binary",
    headers={
        'X-Source': 'geo'   
    },
    ttl=1000
)
response = yield tchannel.call_thrift(
    Foo.getBar(Foo.BarResponse("hi")),
    headers={
        'X-Source': 'geo'   
    },
    timeout=5000
)
response = yield tchannel.as_thrift().call(
    Foo.getBar(Foo.BarResponse("hi")),
    headers={
        'X-Source': 'geo'   
    },
    timeout=5000
)
response = yield tchannel.thrift.call(
    Foo.getBar(Foo.BarResponse("hi")),
    headers={
        'X-Source': 'geo'   
    },
    timeout=5000
)

future = tchannel_sync.call(
    Foo.baz(True),
    headers={
        'X-Source': 'geo'   
    },
    ttl=1000
)

Finagle's is brutal:

1. create filters
2. add to stack
3. write current stack to header buffer
4. use builder to create client with stack
5. filter stack applies to each request

So with a trace filter they can set zipkin headers for each request.

val tracer = mock[Tracer]
//tracer.sampleTrace(any(classManifest[TraceId])) returns Some(true)
when(tracer.sampleTrace(any(classOf[TraceId]))).thenReturn(Some(true))

val filter = new TTwitterClientFilter("service", true, None, protocolFactory)
val buffer = new OutputBuffer(protocolFactory)
buffer().writeMessageBegin(
  new TMessage(ThriftTracing.CanTraceMethodName, TMessageType.CALL, 0))
val options = new thrift.ConnectionOptions
options.write(buffer())
buffer().writeMessageEnd()

val tracing = new TracingFilter[ThriftClientRequest, Array[Byte]](tracer, "TTwitterClientFilterTest")
val service = mock[Service[ThriftClientRequest, Array[Byte]]]
val _request = ArgumentCaptor.forClass(classOf[ThriftClientRequest])
when(service(_request.capture)).thenReturn(Future(Array[Byte]()))

val stack = tracing andThen filter
stack(new ThriftClientRequest(buffer.toArray, false), service)

val header = new thrift.RequestHeader
InputBuffer.peelMessage(_request.getValue.message, header, protocolFactory)

assert(header.isSampled)

sources:

• https://github.com/twitter/finagle/blob/master/finagle-thrift/src/test/scala/com/twitter/finagle/thrift/TTwitterClientFilterTest.scala
• https://github.com/twitter/finagle/blob/master/finagle-thrift/src/test/scala/com/twitter/finagle/thrift/EndToEndTest.scala
• https://github.com/jghoman/finagle-java-example/blob/master/src/main/java/jghoman/Client.java

Wow. no Finagle.

Based on discussion earlier, this is how we expect the client-side streaming API to look:

---

Response streaming

We'll introduce a stream method on the TChannel object and its argscheme-specific proxies that will accept the same arguments as call. Instead of a standard response, it'll return a streaming response object which provides a .read() method.

response = tchannel.stream(service='foo', endpoint='bar', body='listItems')
headers = response.headers
try:
    while True:
        chunk = yield response.read()
        process(chunk)
except EndOfStream:
    pass

Request streaming

We'll have a body_producer parameter on both call and stream. body_producer may be passed in lieu of the body. It will be a function that accepts a write function and calls it to write to the stream. The function must be a coroutine, or return a function that resolves to None when it finishes writing.

@gen.coroutine
def producer(write):
    for line in some_file:
        yield write(line)

response = yield tchannel.call(endpoint='foo', service='bar', body_producer=producer)

# Bidirectional streaming:

response = yield tchannel.stream(endpoint='foo', service='bar', body_producer=producer)
try:
    while True:
        chunk = yield response.read()
        process(chunk)
except EndOfStream:
    pass

Suggestion: make the streamed response iterable, e.g. for thing in response.

@blampe +1