Apache 中的分支和合并 pcollection 列表来自公共输入

Question

I'm building a dataflow pipeline and I'm having some trouble branching and merging outputs. The pipeline I want to build is as follows:

1. Read some input data input_data .
2. A. Extract some metric, metric_1 , on input_data . B. Extract some other metric, metric_2 , on input_data
3. Since these metric extractions are computationally expensive, I want to branch off of the main input_data and merge the outputs afterwards for further calculation. Merge outputs output .

Here's some sample code that encapsulates my actual pipeline

class ReadData(beam.DoFn):
    def process(self, element):
        # read from source
        return [{'input': np.random.rand(100,10)}]


class GetFirstMetric(beam.DoFn):
    def process(self, element):
        # some processing
        return [{'first': np.random.rand(100,4)}]


class GetSecondMetric(beam.DoFn):
    def process(self, element):
        # some processing
        return [{'second': np.random.rand(100,3)}]


def run():
    with beam.Pipeline() as p:
        input_data = (p | 'read sample data' >> beam.ParDo(ReadData()))

        metric_1 = (input_data | 'some metric on input data' >> beam.ParDo(GetFirstMetric()))
        metric_2 = (input_data | 'some aggregate metric' >> beam.ParDo(GetSecondMetric()))

        output = ((metric_1, metric_2) 
                  | beam.Flatten()
                  | beam.combiners.ToList()
                  | beam.Map(print)
        )

When I run this, I get a 'PBegin' object has no attribute 'windowing' error. I've seen some examples and sample code for doing something like this in Java. But I couldn't find the right resources for doing the same in Python. My question is as follows:

1. What's the right way to branch and merge pcollections (especially if the branches came from a common input)?

2. Is there a better pipeline design for accomplishing the same?

Thanks in advance!

Answer 1

In this code, your problem is that you are not 'starting' an initial PCollection. In ReadData.process - what is the value of the variable element ?

Well, the runner can't come up with a value, because there's no input pcollection. You need to create your first PCollection. You'd do something like the following code...

As for making them into a list - perhaps a side input strategy may work. CAn you try the following:

def run():
    with beam.Pipeline() as p:
        starter_pcoll = p | beam.Create(['any'])
        input_data = (starter_pcoll | 'read sample data' >> beam.ParDo(ReadData()))

        metric_1 = (input_data | 'some metric on input data' >> beam.ParDo(GetFirstMetric()))
        metric_2 = (input_data | 'some aggregate metric' >> beam.ParDo(GetSecondMetric()))

        side_in = beam.pvalue.AsList((metric_1, metric_2) 
                                     | beam.Flatten())

        p | beam.Create(['any']) | beam.Map(lambda x, si: print(si),
                                            side_in)

This should make your pipeline run. Happy to clarify about your specific questions further.