Book Image

Big Data Analytics

By : Venkat Ankam
Book Image

Big Data Analytics

By: Venkat Ankam

Overview of this book

Big Data Analytics book aims at providing the fundamentals of Apache Spark and Hadoop. All Spark components – Spark Core, Spark SQL, DataFrames, Data sets, Conventional Streaming, Structured Streaming, MLlib, Graphx and Hadoop core components – HDFS, MapReduce and Yarn are explored in greater depth with implementation examples on Spark + Hadoop clusters. It is moving away from MapReduce to Spark. So, advantages of Spark over MapReduce are explained at great depth to reap benefits of in-memory speeds. DataFrames API, Data Sources API and new Data set API are explained for building Big Data analytical applications. Real-time data analytics using Spark Streaming with Apache Kafka and HBase is covered to help building streaming applications. New Structured streaming concept is explained with an IOT (Internet of Things) use case. Machine learning techniques are covered using MLLib, ML Pipelines and SparkR and Graph Analytics are covered with GraphX and GraphFrames components of Spark. Readers will also get an opportunity to get started with web based notebooks such as Jupyter, Apache Zeppelin and data flow tool Apache NiFi to analyze and visualize data.
Table of Contents (18 chapters)
Big Data Analytics
About the Author
About the Reviewers

Spark resource managers – Standalone, YARN, and Mesos

We have already executed spark applications in the Spark standalone resource manager in other sections of this chapter (within the PySpark shell and applications). Let's try to understand how these cluster resource managers are different from each other and when they should be used.

Local versus cluster mode

Before moving on to cluster resource managers, let's understand how cluster mode is different from local mode.

It is important to understand the scope and life cycle of variables and methods when executing code across a cluster. Let's look at an example with the foreach action:

counter = 0
rdd = sc.parallelize(data)
rdd.foreach(lambda x: counter += x)
print("Counter value: " + counter)

In local mode, the preceding code works fine because the counter variable and RDD are in the same memory space (single JVM).

In cluster mode, the counter value will never change and always remains at 0. In cluster mode, spark computes the closure with...