README-gRPC-details.md

gRPC Server Implementation

To build the gRPC server, you should follow this tutorial for a sample python server and client.

You'll need to install the Protobuf tools

sudo apt install protobuf-compiler
sudo pip3 install grpcio-tools

You should be able to do all of your development on the Google Cloud Console or your laptop.

What you need to do for the gRPC section

You need to build a gRPC service that implements the add, rawimage and sum messages and has the same functionality as the REST implementation. Protocol Buffers lets you provide binary data, but just for consistency we're going to have you implement jsonimage as well.

The reference implementation used these types:

message addMsg {
  int32 a = 1;
  int32 b = 2;
}

message rawImageMsg {
    bytes img=1;
}

message dotProductMsg {
    repeated float a = 1;
    repeated float b = 2;
  }
  
message jsonImageMsg {
      string img=1;
  }

message addReply {
    int32 sum = 1;
}

message dotProductReply {
    float dotproduct = 1;
}

message imageReply {
    int32 width=1;
    int32 height=2;
}

The bytes type can be used to transport an array of bytes. The corresponding code to initialize the imageMsg type looks like:

    img = open('Flatirons_Winter_Sunrise_edit_2.jpg', 'rb').read()
    data = lab6_pb2.imageMsg(img=img)

You should implement the dotproduct service first -- that will walk you through the basic steps of creating a protobuf specification and generating the code. After that, you should add the jsonimage endpoint. You can transport the image using the string designation.

You may see warning in the protocol buffer tutorial that protocol buffers (and gRPC) really aren't designed to transport large messages, but it will work fine for this lab.

The remainder of the steps involved following the python tutorial example and adapting the types.

As with the REST exmaple, you should have a client to accept a command-line argument indicating the endpoint to be tested and the number of iterations to test. For example:

python grpc-client.py localhost add 1000

would run the add endpoint 1000 times against the server on the localhost and then report the time taken divided by the number of queries (1000). This gives you a time-per-query, which should be expressed in milliseconds. The python perf_counter routine from the time module makes it easy to conduct such timing tests. We measure multiple queries because each query is fairly short and we want to average over many such queries.