Apache Beam API

Table Of Contents

• Introduction
• Basic Concepts
• Apache Beam - A Samza’s Perspective

Introduction

Apache Beam brings an easy-to-usen but powerful API and model for state-of-art stream and batch data processing with portability across a variety of languages. The Beam API and model has the following characteristics:

• Simple constructs, powerful semantics: the whole beam API can be simply described by a Pipeline object, which captures all your data processing steps from input to output. Beam SDK supports over 20 data IOs, and data transformations from simple Map to complex Combines and Joins.

• Strong consistency via event-time: Beam provides advanced event-time support so you can perform windowing and aggregations based on when the events happen, instead of when they arrive. The event-time mechanism improves the accuracy of processing results, and guarantees repeatability in results when reprocessing the same data set.

• Comprehensive stream processing semantics: Beam supports an up-to-date stream processing model, including tumbling/sliding/session windows, joins and aggregations. It provides triggers based on conditions of early and late firings, and late arrival handling with accumulation mode and allowed lateness.

• Portability with multiple programming languages: Beam supports a consistent API in multiple languages, including Java, Python and Go. This allows you to leverage the rich ecosystem built for different languages, e.g. ML libs for Python.

Basic Concepts

Let’s walk through the WordCount example to illustrate the Beam basic concepts. A Beam program often starts by creating a Pipeline object in your main() function.

// Start by defining the options for the pipeline.
PipelineOptions options = PipelineOptionsFactory.create();

// Then create the pipeline.
Pipeline p = Pipeline.create(options);

The PipelineOptions supported by SamzaRunner is documented in detail here.

Let’s apply the first data transform to read from a text file using TextIO.read():

PCollection<String> lines = p.apply(
  "ReadLines", TextIO.read().from("/path/to/inputData"));

To break down each line into words, you can use a FlatMap:

PCollection<String> words = lines.apply(
    FlatMapElements.into(TypeDescriptors.strings())
        .via((String word) -> Arrays.asList(word.split("\\W+"))));

Beam provides a build-in transform Count.perElement to count the number of elements based on each value. Let’s use it here to count the words:

PCollection<KV<String, Long>> counts = pipeline.apply(Count.perElement());

Finally we format the counts into strings and write to a file using TextIO.write():

counts.apply(ToString.kvs())
      .apply(TextIO.write().to("/path/to/output").withoutSharding());

To run your pipeline and wait for the results, just do:

pipeline.run().waitUntilFinish();

Or you can run your pipeline asynchronously, e.g. when you submit it to a remote cluster:

pipeline.run();

To run this Beam program with Samza, you can simply provide “–runner=SamzaRunner” as a program argument. You can follow our quick start to set up your project and run different examples. For more details on writing the Beam program, please refer the Beam programming guide.

Apache Beam - A Samza’s Perspective

The goal of Samza is to provide large-scale streaming processing capabilities with first-class state support. This does not contradict with Beam. In fact, while Samza lays out a solid foundation for large-scale stateful stream processing, Beam adds the cutting-edge stream processing API and model on top of it. The Beam API and model allows further optimization in the Samza platform, including multi-stage distributed computation and parallel processing on the per-key basis. The performance enhancements from these optimizations will benefit both Samza and its users. Samza can also further improve Beam model by providing various use cases. We firmly believe Samza will benefit from collaborating with Beam.