I recently contributed to Apache Beam by adding a common pipeline pattern - Cache data using a shared object. Both batch and streaming pipelines are introduced, and they utilise the Shared class of the Python SDK to enrich PCollection elements. This pattern can be more memory-efficient than side inputs, simpler than a stateful DoFn, and more performant than calling an external service, because it does not have to access an external service for every element or bundle of elements. In this post, we discuss this pattern in more details with batch and streaming use cases. For the latter, we configure the cache gets refreshed periodically.

In this post, we develop an Apache Beam pipeline where the input data is augmented by a Remote Procedure Call (RPC) service. Each input element performs an RPC call and the output is enriched by the response. This is not an efficient way of accessing an external service provided that the service can accept more than one element. In the subsequent two posts, we will discuss updated pipelines that make RPC calls more efficiently. We begin with illustrating how to manage development resources followed by demonstrating the RPC service that we use in this series. Finally, we develop a Beam pipeline that accesses the external service to augment the input elements.

In this post, we develop two Apache Beam pipelines that track sport activities of users and output their speed periodically. The first pipeline uses native transforms and Beam SQL is used for the latter. While Beam SQL can be useful in some situations, its features in the Python SDK are not complete compared to the Java SDK. Therefore, we are not able to build the required tracking pipeline using it. We end up discussing potential improvements of Beam SQL so that it can be used for building competitive applications with the Python SDK.

In this post, we develop two Apache Beam pipelines that calculate average word lengths from input texts that are ingested by a Kafka topic. They obtain the statistics in different angles. The first pipeline emits the global average lengths whenever a new input text arrives while the latter triggers those values in a sliding time window.

In this series, we develop Apache Beam Python pipelines. The majority of them are from Building Big Data Pipelines with Apache Beam by Jan Lukavský. Mainly relying on the Java SDK, the book teaches fundamentals of Apache Beam using hands-on tasks, and we convert those tasks using the Python SDK. We focus on streaming pipelines, and they are deployed on a local (or embedded) Apache Flink cluster using the Apache Flink Runner. Beginning with setting up the development environment, we build two pipelines that obtain top K most frequent words and the word that has the longest word length in this post.

In this post, we develop an Apache Beam pipeline using the Python SDK and deploy it on an Apache Flink cluster via the Apache Flink Runner. Same as Part I, we deploy a Kafka cluster using the Strimzi Operator on a minikube cluster as the pipeline uses Apache Kafka topics for its data source and sink. Then, we develop the pipeline as a Python package and add the package to a custom Docker image so that Python user code can be executed externally. For deployment, we create a Flink session cluster via the Flink Kubernetes Operator, and deploy the pipeline using a Kubernetes job. Finally, we check the output of the application by sending messages to the input Kafka topic using a Python producer application.

Flink Kubernetes Operator acts as a control plane to manage the complete deployment lifecycle of Apache Flink applications. With the operator, we can simplify deployment and management of Python stream processing applications. In this series, we discuss how to deploy a PyFlink application and Python Apache Beam pipeline on the Flink Runner on Kubernetes. In Part 1, we first deploy a Kafka cluster on a minikube cluster as the source and sink of the PyFlink application are Kafka topics. Then, the application source is packaged in a custom Docker image and deployed on the minikube cluster using the Flink Kubernetes Operator. Finally, the output of the application is checked by sending messages to the input Kafka topic using a Python producer application.

We developed batch and streaming pipelines in Part 2 and Part 4. Often it is faster and simpler to identify and fix bugs on the pipeline code by performing local unit testing. Moreover, especially when it comes to creating a streaming pipeline, unit testing cases can facilitate development further by using TestStream as it allows us to advance watermarks or processing time according to different scenarios. In this post, we discuss how to perform unit testing of the batch and streaming pipelines that we developed earlier.

In Part 3, we discussed the portability layer of Apache Beam as it helps understand (1) how Python pipelines run on the Flink Runner and (2) how multiple SDKs can be used in a single pipeline, followed by demonstrating local Flink and Kafka cluster creation for developing streaming pipelines. In this post, we develop a streaming pipeline that aggregates page visits by user in a fixed time window of 20 seconds. Two versions of the pipeline are created with/without relying on Beam SQL.

Beam pipelines are portable between batch and streaming semantics but not every Runner is equally capable. The Apache Flink Runner supports Python, and it has good features that allow us to develop streaming pipelines effectively. We first discuss the portability layer of Apache Beam as it helps understand (1) how a pipeline developed by the Python SDK can be executed in the Flink Runner that only understands Java JAR and (2) how multiple SDKs can be used in a single pipeline. Then we move on to how to manage local Flink and Kafka clusters using bash scripts. Finally, we end up illustrating a simple streaming pipeline, which reads and writes website visit logs from and to Kafka topics.