Pydoc Pydoc

Separates elements in a collection into multiple output collections. The partitioning function contains the logic that determines how to separate the elements of the input collection into each resulting partition output collection.

The number of partitions must be determined at graph construction time. You cannot determine the number of partitions in mid-pipeline

See more information in the Beam Programming Guide.


In the following examples, we create a pipeline with a PCollection of produce with their icon, name, and duration. Then, we apply Partition in multiple ways to split the PCollection into multiple PCollections.

Partition accepts a function that receives the number of partitions, and returns the index of the desired partition for the element. The number of partitions passed must be a positive integer, and it must return an integer in the range 0 to num_partitions-1.

Example 1: Partition with a function

In the following example, we have a known list of durations. We partition the PCollection into one PCollection for every duration type.

Example 2: Partition with a lambda function

We can also use lambda functions to simplify Example 1.

Example 3: Partition with multiple arguments

You can pass functions with multiple arguments to Partition. They are passed as additional positional arguments or keyword arguments to the function.

In machine learning, it is a common task to split data into training and a testing datasets. Typically, 80% of the data is used for training a model and 20% is used for testing.

In this example, we split a PCollection dataset into training and testing datasets. We define split_dataset, which takes the plant element, num_partitions, and an additional argument ratio. The ratio is a list of numbers which represents the ratio of how many items will go into each partition. num_partitions is used by Partitions as a positional argument, while plant and ratio are passed to split_dataset.

If we want an 80%/20% split, we can specify a ratio of [8, 2], which means that for every 10 elements, 8 go into the first partition and 2 go into the second. In order to determine which partition to send each element, we have different buckets. For our case [8, 2] has 10 buckets, where the first 8 buckets represent the first partition and the last 2 buckets represent the second partition.

First, we check that the ratio list’s length corresponds to the num_partitions we pass. We then get a bucket index for each element, in the range from 0 to 9 (num_buckets-1). We could do hash(element) % len(ratio), but instead we sum all the ASCII characters of the JSON representation to make it deterministic. Finally, we loop through all the elements in the ratio and have a running total to identify the partition index to which that bucket corresponds.

This split_dataset function is generic enough to support any number of partitions by any ratio. You might want to adapt the bucket assignment to use a more appropriate or randomized hash for your dataset.

Pydoc Pydoc