Book Image

Learning Apache Apex

By : Thomas Weise, Ananth Gundabattula, Munagala V. Ramanath, David Yan, Kenneth Knowles
Book Image

Learning Apache Apex

By: Thomas Weise, Ananth Gundabattula, Munagala V. Ramanath, David Yan, Kenneth Knowles

Overview of this book

Apache Apex is a next-generation stream processing framework designed to operate on data at large scale, with minimum latency, maximum reliability, and strict correctness guarantees. Half of the book consists of Apex applications, showing you key aspects of data processing pipelines such as connectors for sources and sinks, and common data transformations. The other half of the book is evenly split into explaining the Apex framework, and tuning, testing, and scaling Apex applications. Much of our economic world depends on growing streams of data, such as social media feeds, financial records, data from mobile devices, sensors and machines (the Internet of Things - IoT). The projects in the book show how to process such streams to gain valuable, timely, and actionable insights. Traditional use cases, such as ETL, that currently consume a significant chunk of data engineering resources are also covered. The final chapter shows you future possibilities emerging in the streaming space, and how Apache Apex can contribute to it.
Table of Contents (17 chapters)
Title Page
Credits
About the Authors
About the Reviewer
www.PacktPub.com
Customer Feedback
Preface

Processing guarantees


The Apex engine by default guarantees that data is processed at-least-once and that state updates within the DAG occur exactly-once. With respect to state mutation through interaction with external systems, the results depend on the connector (refer to Chapter 3, The Apex Library). Connectors that support the exactly-once results include Files, Kafka, JDBC, Cassandra, and all others where the write operations are, or can be made, idempotent. We will look at an example application in the next section.

In distributed systems, a guarantee of the exactly-once processing is not really possible since nodes may go down at any time and when they are restored, some reprocessing of prior data, however minimal, must occur in order to guarantee correctness (or we have to accept data loss, which yields at-most-once processing). So, when we see exactly-once in published feature matrices of stream processors, it really means at-least-once along with the often implicit assumption of...