Mean

You can use Mean transforms to compute the arithmetic mean of the elements in a collection or the mean of the values associated with each key in a collection of key-value pairs. {{if (eq .Sdk “go”)}} Mean() returns a transformation that returns a collection whose content is the average of the elements of the input collection. If there are no elements in the input collection, 0 is returned.

import (
  "github.com/apache/beam/sdks/go/pkg/beam"
  "github.com/apache/beam/sdks/go/pkg/beam/transforms/stats"
)

func ApplyTransform(s beam.Scope, input beam.PCollection) beam.PCollection {
  return stats.Mean(s, input)
}

You can use MeanPerKey() to calculate the mean of the elements associated with each unique key.

import (
  "github.com/apache/beam/sdks/go/pkg/beam"
  "github.com/apache/beam/sdks/go/pkg/beam/transforms/stats"
)

func ApplyTransform(s beam.Scope, input beam.PCollection) beam.PCollection {
  return stats.MeanPerKey(s, input)
}

{{end}} {{if (eq .Sdk “java”)}} Mean.globally() returns a transformation that returns a collection whose content is the average of the elements of the input collection. If there are no elements in the input collection, 0 is returned.

PCollection<Integer> input = pipeline.apply(Create.of(1, 2, 3, 4, 5, 6, 7, 8, 9, 10));
PCollection<Double> mean = input.apply(Mean.globally());

Output

5.5

Mean.perKey() returns a transform that returns a collection containing an output element mapping each distinct key in the input collection to the mean of the values associated with that key in the input collection.

PCollection<KV<String, Integer>> input = pipeline.apply(
    Create.of(KV.of("🥕", 3),
              KV.of("🥕", 2),
              KV.of("🍆", 1),
              KV.of("🍅", 4),
              KV.of("🍅", 5),
              KV.of("🍅", 3)));
PCollection<KV<String, Double>> meanPerKey = input.apply(Mean.perKey());

Output

KV{🍆, 1.0}
KV{🥕, 2.5}
KV{🍅, 4.0}

{{end}} {{if (eq .Sdk “python”)}} {{end}}

Playground exercise

You can find the full code of this example in the playground window, which you can run and experiment with. {{if (eq .Sdk “go”)}} Mean returns the mean from the PCollection. If you replace the integers input with this map input:

input:= beam.ParDo(s, func(_ []byte, emit func(int, int)){
  emit(1,1)
  emit(1,4)
  emit(2,6)
  emit(2,3)
  emit(2,-4)
  emit(3,23)
}, beam.Impulse(s))

And replace stats.Mean with stats.MeanPerKey it will output the mean by key. {{end}} {{if (eq .Sdk “java”)}} Mean.globally returns the mean from the PCollection. If you replace the integers input with this map input:

PCollection<KV<Integer, Integer>> input = pipeline.apply(
    Create.of(KV.of(1, 11),
    KV.of(1, 36),
    KV.of(2, 91),
    KV.of(3, 33),
    KV.of(3, 11),
    KV.of(4, 33)));

And replace Mean.globally with Mean.perKey it will output the means by key. It is also necessary to replace the generic type:

PCollection<KV<Integer, Double>> output = applyTransform(input);

static PCollection<KV<Integer, Double>> applyTransform(PCollection<KV<Integer, Integer>> input) {
        return input.apply(Mean.perKey());
    }

{{end}} {{if (eq .Sdk “python”)}} beam.combiners.Mean.Globally() returns mean from PCCollection. If you replace the integers input with this map input and replace beam.combiners.Mean.Globally() on beam.combiners.Mean.PerKey() it will output the mean by key :

p | beam.Create([(1, 36),(2, 91),(3, 33),(3, 11),(4, 67),]) | beam.combiners.Mean.PerKey()

{{end}}

Have you also noticed the order in which the collection items are displayed in the console? Why is that? You can also run the example several times to see if the output remains the same or changes.