Before I dive into more complex methods to analyze your data later in the book, I would like to stop at basic data exploratory tasks on which almost all data scientists spend at least 80-90% of their productive time. The data preparation, cleansing, transforming, and joining the data alone is estimated to be a $44 billion/year industry alone (Data Preparation in the Big Data Era by Federico Castanedo and Best Practices for Data Integration, O'Reilly Media, 2015). Given this fact, it is surprising that people only recently started spending more time on the science of developing best practices and establishing good habits, documentation, and teaching materials for the whole process of data preparation (Beautiful Data: The Stories Behind Elegant Data Solutions, edited by Toby Segaran and Jeff Hammerbacher, O'Reilly Media, 2009 and Advanced Analytics with Spark: Patterns for Learning from Data at Scale by Sandy Ryza et al., O'Reilly Media, 2015).

Few data scientists would agree on specific tools and techniques—and there are multiple ways to perform the exploratory data analysis, ranging from Unix command line to using very popular open source and commercial ETL and visualization tools. The focus of this chapter is how to use Scala and a laptop-based environment to benefit from techniques that are commonly referred as a functional paradigm of programming. As I will discuss, these techniques can be transferred to exploratory analysis over distributed system of machines using Hadoop/Spark.

What has functional programming to do with it? Spark was developed in Scala for a good reason. Many basic principles that lie at the foundation of functional programming, such as lazy evaluation, immutability, absence of side effects, list comprehensions, and monads go really well with processing data in distributed environments, specifically, when performing the data preparation and transformation tasks on big data. Thanks to abstractions, these techniques work well on a local workstation or a laptop. As mentioned earlier, this does not preclude us from processing very large datasets up to dozens of TBs on modern laptops connected to distributed clusters of storage/processing nodes. We can do it one topic or focus area at the time, but often we even do not have to sample or filter the dataset with proper partitioning. We will use Scala as our primary tool, but will resort to other tools if required.

While Scala is complete in the sense that everything that can be implemented in other languages can be implemented in Scala, Scala is fundamentally a high-level, or even a scripting, language. One does not have to deal with low-level details of data structures and algorithm implementations that in their majority have already been tested by a plethora of applications and time, in, say, Java or C++—even though Scala has its own collections and even some basic algorithm implementations today. Specifically, in this chapter, I'll be focusing on using Scala/Spark only for high-level tasks.

In this chapter, we will cover the following topics: