077_user_defined_parallelism_04

/*
This is example 4 in the series of 12 User Defined Parallelism (UDP) scenarios.
UDP is a great feature to parallelize an entire composite or a particular operator.

This example code is taken from the Streams InfoCenter and added here to benefit the
beginners of the Streams SPL programming model. Many thanks to our Streams colleague
Scott Schneider for coming up with this set of UDP examples. Full credit goes to him.

It is recommended that you run this example in Distributed mode and visualize the
parallel region in the Streams instance graph.
*/
namespace com.acme.test;

// In this example of user-defined parallelism, a logical processing element (PE) contains operators
// from outside the parallel region, a splitter is fused inside the PE, and the merge point is
// not inside the PE. The parallel transformation replicates only the operators and not the PE.
// In this instance, the entire application is a single PE since all the operators are fused.

composite UDP4 {
	graph
		stream<int32 i> MyData = Beacon() {
			param
				iterations: 5000; 
				
			config
				placement: partitionColocation("AB");
		}

		stream<MyData> EnrichedData = Custom(MyData) {
			logic
				state: {
					mutable int32 _i = 0;
				}
				
				onTuple MyData: {
					_i++;
					MyData.i = _i;
					submit(MyData, EnrichedData);
				}
				
			config
				placement: partitionColocation("AB");
		}
		
		@parallel (width=3)
		stream<EnrichedData> TransformedData = Comp3(EnrichedData) {
			config
				placement: partitionColocation("AB");
		}
		
		() as MySink = FileSink(TransformedData) {
			param
				file: "Test1.csv";

			config
				placement: partitionColocation("AB");
		}		
}


composite Comp3(input In; output B) {
	graph
		stream<int32 i> A = Custom(In) {
			logic
				onTuple In: {
					In.i = In.i + 25;
					submit(In, A);
				}
		}
		
		stream<A> B = Custom(A) {
			logic
				onTuple A: {
					A.i = A.i - 4;
					submit(A, B);
				}
		}
}