Book Image

Distributed Data Systems with Azure Databricks

By : Alan Bernardo Palacio
Book Image

Distributed Data Systems with Azure Databricks

By: Alan Bernardo Palacio

Overview of this book

Microsoft Azure Databricks helps you to harness the power of distributed computing and apply it to create robust data pipelines, along with training and deploying machine learning and deep learning models. Databricks' advanced features enable developers to process, transform, and explore data. Distributed Data Systems with Azure Databricks will help you to put your knowledge of Databricks to work to create big data pipelines. The book provides a hands-on approach to implementing Azure Databricks and its associated methodologies that will make you productive in no time. Complete with detailed explanations of essential concepts, practical examples, and self-assessment questions, you’ll begin with a quick introduction to Databricks core functionalities, before performing distributed model training and inference using TensorFlow and Spark MLlib. As you advance, you’ll explore MLflow Model Serving on Azure Databricks and implement distributed training pipelines using HorovodRunner in Databricks. Finally, you’ll discover how to transform, use, and obtain insights from massive amounts of data to train predictive models and create entire fully working data pipelines. By the end of this MS Azure book, you’ll have gained a solid understanding of how to work with Databricks to create and manage an entire big data pipeline.
Table of Contents (17 chapters)
Section 1: Introducing Databricks
Section 2: Data Pipelines with Databricks
Section 3: Machine and Deep Learning with Databricks

Using the Horovod distributed learning library in Azure Databricks

horovod is a library for distributed deep learning training. It supports commonly used frameworks such as TensorFlow, Keras, and PyTorch. As mentioned before, it is based on the tensorflow-allreduce library and implements the ring allreduce algorithm in order to ease the migration from single-graphics processing unit (GPU) training to parallel-GPU distributed training.

In order to do this, we adapt a single-GPU training script of a deep learning model to use the horovod library during the training process. Once we have adapted the script, it can run on single or multiple GPUs without changes to the code.

The horovod library uses a data parallelization strategy by allowing efficient distribution of the training to multiple GPUs in parallel in an optimized way, by implementing the ring allreduce algorithm to overcome communication limitations.

It is implemented in a way that each GPU gets a mini-batch of data...