Data Engineering with Apache Spark, Delta Lake, and Lakehouse

By Manoj Kukreja
    What do you get with a Packt Subscription?

  • Instant access to this title and 7,500+ eBooks & Videos
  • Constantly updated with 100+ new titles each month
  • Breadth and depth in over 1,000+ technologies
  1. Free Chapter
    Chapter 1: The Story of Data Engineering and Analytics
About this book
In the world of ever-changing data and schemas, it is important to build data pipelines that can auto-adjust to changes. This book will help you build scalable data platforms that managers, data scientists, and data analysts can rely on. Starting with an introduction to data engineering, along with its key concepts and architectures, this book will show you how to use Microsoft Azure Cloud services effectively for data engineering. You'll cover data lake design patterns and the different stages through which the data needs to flow in a typical data lake. Once you've explored the main features of Delta Lake to build data lakes with fast performance and governance in mind, you'll advance to implementing the lambda architecture using Delta Lake. Packed with practical examples and code snippets, this book takes you through real-world examples based on production scenarios faced by the author in his 10 years of experience working with big data. Finally, you'll cover data lake deployment strategies that play an important role in provisioning the cloud resources and deploying the data pipelines in a repeatable and continuous way. By the end of this data engineering book, you'll know how to effectively deal with ever-changing data and create scalable data pipelines to streamline data science, ML, and artificial intelligence (AI) tasks.
Publication date:
October 2021


Chapter 1: The Story of Data Engineering and Analytics

Every byte of data has a story to tell. The real question is whether the story is being narrated accurately, securely, and efficiently. In the modern world, data makes a journey of its own—from the point it gets created to the point a user consumes it for their analytical requirements.

But what makes the journey of data today so special and different compared to before? After all, Extract, Transform, Load (ETL) is not something that recently got invented. In fact, I remember collecting and transforming data since the time I joined the world of information technology (IT) just over 25 years ago.

In this chapter, we will discuss some reasons why an effective data engineering practice has a profound impact on data analytics.

In this chapter, we will cover the following topics:

  • The journey of data
  • Exploring the evolution of data analytics
  • The monetary power of data


    the road to effective data analytics leads through effective data engineering.


The journey of data

Data engineering is the vehicle that makes the journey of data possible, secure, durable, and timely. A data engineer is the driver of this vehicle who safely maneuvers the vehicle around various roadblocks along the way without compromising the safety of its passengers. Waiting at the end of the road are data analysts, data scientists, and business intelligence (BI) engineers who are eager to receive this data and start narrating the story of data. You can see this reflected in the following screenshot:

Figure 1.1 – Data's journey to effective data analysis

Figure 1.1 – Data's journey to effective data analysis

Traditionally, the journey of data revolved around the typical ETL process. Unfortunately, the traditional ETL process is simply not enough in the modern era anymore. Due to the immense human dependency on data, there is a greater need than ever to streamline the journey of data by using cutting-edge architectures, frameworks, and tools.

You may also be wondering why the journey of data is even required. Gone are the days where datasets were limited, computing power was scarce, and the scope of data analytics was very limited. We now live in a fast-paced world where decision-making needs to be done at lightning speeds using data that is changing by the second. Let's look at how the evolution of data analytics has impacted data engineering.


Exploring the evolution of data analytics

Data analytics has evolved over time, enabling us to do bigger and better. For many years, the focus of data analytics was limited to descriptive analysis, where the focus was to gain useful business insights from data, in the form of a report. This type of analysis was useful to answer question such as "What happened?". A hypothetical scenario would be that the sales of a company sharply declined within the last quarter.

Very quickly, everyone started to realize that there were several other indicators available for finding out what happened, but it was the why it happened that everyone was after. The core analytics now shifted toward diagnostic analysis, where the focus is to identify anomalies in data to ascertain the reasons for certain outcomes. An example scenario would be that the sales of a company sharply declined in the last quarter because there was a serious drop in inventory levels, arising due to floods in the manufacturing units of the suppliers. This form of analysis further enhances the decision support mechanisms for users, as illustrated in the following diagram:

Figure 1.2 – The evolution of data analytics

Figure 1.2 – The evolution of data analytics

Important note

Both descriptive analysis and diagnostic analysis try to impact the decision-making process using factual data only.

Since the advent of time, it has always been a core human desire to look beyond the present and try to forecast the future. If we can predict future outcomes, we can surely make a lot of better decisions, and so the era of predictive analysis dawned, where the focus revolves around "What will happen in the future?". Predictive analysis can be performed using machine learning (ML) algorithms—let the machine learn from existing and future data in a repeated fashion so that it can identify a pattern that enables it to predict future trends accurately.

Now that we are well set up to forecast future outcomes, we must use and optimize the outcomes of this predictive analysis. Based on the results of predictive analysis, the aim of prescriptive analysis is to provide a set of prescribed actions that can help meet business goals.

Important note

Unlike descriptive and diagnostic analysis, predictive and prescriptive analysis try to impact the decision-making process, using both factual and statistical data.

But how can the dreams of modern-day analysis be effectively realized? After all, data analysts and data scientists are not adequately skilled to collect, clean, and transform the vast amount of ever-increasing and changing datasets.

The data engineering practice is commonly referred to as the primary support for modern-day data analytics' needs.

The following are some major reasons as to why a strong data engineering practice is becoming an absolutely unignorable necessity for today's businesses:

  • Core capabilities of compute and storage resources
  • Availability of varying datasets
  • The paradigm shift to distributed computing
  • Adoption of cloud computing
  • Data storytelling

We'll explore each of these in the following subsections.

Important note

Having a strong data engineering practice ensures the needs of modern analytics are met in terms of durability, performance, and scalability.

Core capabilities of storage and compute resources

25 years ago, I had an opportunity to buy a Sun Solaris server—128 megabytes (MB) random-access memory (RAM), 2 gigabytes (GB) storage—for close to $ 25K. The intended use of the server was to run a client/server application over an Oracle database in production. Given the high price of storage and compute resources, I had to enforce strict countermeasures to appropriately balance the demands of online transaction processing (OLTP) and online analytical processing (OLAP) of my users. One such limitation was implementing strict timings for when these programs could be run; otherwise, they ended up using all available power and slowing down everyone else.

Today, you can buy a server with 64 GB RAM and several terabytes (TB) of storage at one-fifth the price. The extra power available can do wonders for us. Multiple storage and compute units can now be procured just for data analytics workloads. The extra power available enables users to run their workloads whenever they like, however they like. In fact, it is very common these days to run analytical workloads on a continuous basis using data streams, also known as stream processing.

The installation, management, and monitoring of multiple compute and storage units requires a well-designed data pipeline, which is often achieved through a data engineering practice.

Availability of varying datasets

A few years ago, the scope of data analytics was extremely limited. Performing data analytics simply meant reading data from databases and/or files, denormalizing the joins, and making it available for descriptive analysis. The structure of data was largely known and rarely varied over time.

We live in a different world now; not only do we produce more data, but the variety of data has increased over time. In addition to collecting the usual data from databases and files, it is common these days to collect data from social networking, website visits, infrastructure logs' media, and so on, as depicted in the following screenshot:

Figure 1.3 – Variety of data increases the accuracy of data analytics

Figure 1.3 – Variety of data increases the accuracy of data analytics

Important note

More variety of data means that data analysts have multiple dimensions to perform descriptive, diagnostic, predictive, or prescriptive analysis.

Naturally, the varying degrees of datasets injects a level of complexity into the data collection and processing process. On the flip side, it hugely impacts the accuracy of the decision-making process as well as the prediction of future trends. A well-designed data engineering practice can easily deal with the given complexity.

The paradigm shift to distributed computing

The traditional data processing approach used over the last few years was largely singular in nature. To process data, you had to create a program that collected all required data for processing—typically from a database—followed by processing it in a single thread. This type of processing is also referred to as data-to-code processing. Unfortunately, there are several drawbacks to this approach, as outlined here:

  • Since vast amounts of data travel to the code for processing, at times this causes heavy network congestion. Since a network is a shared resource, users who are currently active may start to complain about network slowness.
  • Being a single-threaded operation means the execution time is directly proportional to the data. Therefore, the growth of data typically means the process will take longer to finish. This could end up significantly impacting and/or delaying the decision-making process, therefore rendering the data analytics useless at times.
  • Something as minor as a network glitch or machine failure requires the entire program cycle to be restarted, as illustrated in the following diagram:
Figure 1.4 – Rise of distributed computing

Figure 1.4 – Rise of distributed computing

The distributed processing approach, which I refer to as the paradigm shift, largely takes care of the previously stated problems. Instead of taking the traditional data-to-code route, the paradigm is reversed to code-to-data.

In a distributed processing approach, several resources collectively work as part of a cluster, all working toward a common goal. In simple terms, this approach can be compared to a team model where every team member takes on a portion of the load and executes it in parallel until completion. If a team member falls sick and is unable to complete their share of the workload, some other member automatically gets assigned their portion of the load.

Distributed processing has several advantages over the traditional processing approach, outlined as follows:

  • The code-to-data paradigm shift ensures the network does not get clogged. The entire idea of distributed processing heavily relies on the assumption that data is stored in a distributed fashion across several machines, also referred to as nodes. At the time of processing, only the code portion (usually a much smaller footprint as compared to actual data) is sent over to each node that stores the portion of the data being processed. This ensures that the processing happens locally on the node where the data is stored.
  • Since several nodes are collectively participating in data processing, the overall completion time is drastically reduced.
  • Program execution is immune to network and node failures. If a node failure is encountered, then a portion of the work is assigned to another available node in the cluster.

    Important note

    Distributed processing is implemented using well-known frameworks such as Hadoop, Spark, and Flink. Modern massively parallel processing (MPP)-style data warehouses such as Amazon Redshift, Azure Synapse, Google BigQuery, and Snowflake also implement a similar concept.

Since distributed processing is a multi-machine technology, it requires sophisticated design, installation, and execution processes. That makes it a compelling reason to establish good data engineering practices within your organization.

Adoption of cloud computing

The vast adoption of cloud computing allows organizations to abstract the complexities of managing their own data centers. Migrating their resources to the cloud offers faster deployments, greater flexibility, and access to a pricing model that, if used correctly, can result in major cost savings.

In the previous section, we talked about distributed processing implemented as a cluster of multiple machines working as a group. For this reason, deploying a distributed processing cluster is expensive.

In the pre-cloud era of distributed processing, clusters were created using hardware deployed inside on-premises data centers. Very careful planning was required before attempting to deploy a cluster (otherwise, the outcomes were less than desired). You might argue why such a level of planning is essential. Let me address this:

  • Since the hardware needs to be deployed in a data center, you need to physically procure it. The real question is how many units you would procure, and that is precisely what makes this process so complex.
  • Order more units than required and you'll end up with unused resources, wasting money.
  • Order fewer units than required and you will have insufficient resources, job failures, and degraded performance.

    To order the right number of machines, you start the planning process by performing benchmarking of the required data processing jobs.

  • The results from the benchmarking process are a good indicator of how many machines will be able to take on the load to finish the processing in the desired time. You now need to start the procurement process from the hardware vendors. Keeping in mind the cycle of procurement and shipping process, this could take weeks to months to complete.
  • Once the hardware arrives at your door, you need to have a team of administrators ready who can hook up servers, install the operating system, configure networking and storage, and finally install the distributed processing cluster software—this requires a lot of steps and a lot of planning.

I hope you may now fully agree that the careful planning I spoke about earlier was perhaps an understatement. The complexities of on-premises deployments do not end after the initial installation of servers is completed. You are still on the hook for regular software maintenance, hardware failures, upgrades, growth, warranties, and more.

This is precisely the reason why the idea of cloud adoption is being very well received. Having resources on the cloud shields an organization from many operational issues. Additionally, the cloud provides the flexibility of automating deployments, scaling on demand, load-balancing resources, and security.

Important note

Many aspects of the cloud particularly scale on demand, and the ability to offer low pricing for unused resources is a game-changer for many organizations. If used correctly, these features may end up saving a significant amount of cost. Having a well-designed cloud infrastructure can work miracles for an organization's data engineering and data analytics practice.

Data storytelling

I started this chapter by stating Every byte of data has a story to tell. Data storytelling is a new alternative for non-technical people to simplify the decision-making process using narrated stories of data. Traditionally, decision makers have heavily relied on visualizations such as bar charts, pie charts, dashboarding, and so on to gain useful business insights. These visualizations are typically created using the end results of data analytics. The problem is that not everyone views and understands data in the same way. Let me give you an example to illustrate this further.

Here is a BI engineer sharing stock information for the last quarter with senior management:

Figure 1.5 – Visualizing data using simple graphics

Figure 1.5 – Visualizing data using simple graphics

And here is the same information being supplied in the form of data storytelling:

Figure 1.6 – Storytelling approach to data visualization

Figure 1.6 – Storytelling approach to data visualization

Important note

Visualizations are effective in communicating why something happened, but the storytelling narrative supports the reasons for it to happen.

Data storytelling tries to communicate the analytic insights to a regular person by providing them with a narration of data in their natural language. This does not mean that data storytelling is only a narrative. It is a combination of narrative data, associated data, and visualizations. With all these combined, an interesting story emerges—a story that everyone can understand.

As data-driven decision-making continues to grow, data storytelling is quickly becoming the standard for communicating key business insights to key stakeholders.

There's another benefit to acquiring and understanding data: financial. Let's look at the monetary power of data next.


The monetary power of data

Modern-day organizations are immensely focused on revenue acceleration. Traditionally, organizations have primarily focused on increasing sales as a method of revenue acceleration… but is there a better method?

Modern-day organizations that are at the forefront of technology have made this possible using revenue diversification. Here are some of the methods used by organizations today, all made possible by the power of data.

Organic growth

During my initial years in data engineering, I was a part of several projects in which the focus of the project was beyond the usual. On several of these projects, the goal was to increase revenue through traditional methods such as increasing sales, streamlining inventory, targeted advertising, and so on. This meant collecting data from various sources, followed by employing the good old descriptive, diagnostic, predictive, or prescriptive analytics techniques.

But what can be done when the limits of sales and marketing have been exhausted? Where does the revenue growth come from?

Some forward-thinking organizations realized that increasing sales is not the only method for revenue diversification. They started to realize that the real wealth of data that has accumulated over several years is largely untapped. Instead of solely focusing their efforts entirely on the growth of sales, why not tap into the power of data and find innovative methods to grow organically?

This innovative thinking led to the revenue diversification method known as organic growth. Subsequently, organizations started to use the power of data to their advantage in several ways. Let's look at several of them.

Customer retention

Data scientists can create prediction models using existing data to predict if certain customers are in danger of terminating their services due to complaints. Based on this list, customer service can run targeted campaigns to retain these customers. By retaining a loyal customer, not only do you make the customer happy, but you also protect your bottom line.

Fraud prevention

Banks and other institutions are now using data analytics to tackle financial fraud. Based on key financial metrics, they have built prediction models that can detect and prevent fraudulent transactions before they happen. These models are integrated within case management systems used for issuing credit cards, mortgages, or loan applications.

Using the same technology, credit card clearing houses continuously monitor live financial traffic and are able to flag and prevent fraudulent transactions before they happen. Detecting and preventing fraud goes a long way in preventing long-term losses.

Problem detection

I was part of an internet of things (IoT) project where a company with several manufacturing plants in North America was collecting metrics from electronic sensors fitted on thousands of machinery parts. The sensor metrics from all manufacturing plants were streamed to a common location for further analysis, as illustrated in the following diagram:

Figure 1.7 – IoT is contributing to a major growth of data

Figure 1.7 – IoT is contributing to a major growth of data

These metrics are helpful in pinpointing whether a certain consumable component such as rubber belts have reached or are nearing their end-of-life (EOL) cycle. Collecting these metrics is helpful to a company in several ways, including the following:

  • The data indicates the machinery where the component has reached its EOL and needs to be replaced. Having this data on hand enables a company to schedule preventative maintenance on a machine before a component breaks (causing downtime and delays).
  • The data from machinery where the component is nearing its EOL is important for inventory control of standby components. Before this system is in place, a company must procure inventory based on guesstimates. Buy too few and you may experience delays; buy too many, you waste money. At any given time, a data pipeline is helpful in predicting the inventory of standby components with greater accuracy.

The combined power of IoT and data analytics is reshaping how companies can make timely and intelligent decisions that prevent downtime, reduce delays, and streamline costs.

Data monetization

Innovative minds never stop or give up. They continuously look for innovative methods to deal with their challenges, such as revenue diversification. Organizations quickly realized that if the correct use of their data was so useful to themselves, then the same data could be useful to others as well.

As per Wikipedia, data monetization is the "act of generating measurable economic benefits from available data sources".

The following diagram depicts data monetization using application programming interfaces (APIs):

Figure 1.8 – Monetizing data using APIs is the latest trend

Figure 1.8 – Monetizing data using APIs is the latest trend

In the latest trend, organizations are using the power of data in a fashion that is not only beneficial to themselves but also profitable to others. In a recent project dealing with the health industry, a company created an innovative product to perform medical coding using optical character recognition (OCR) and natural language processing (NLP).

Before the project started, this company made sure that we understood the real reason behind the project—data collected would not only be used internally but would be distributed (for a fee) to others as well. Knowing the requirements beforehand helped us design an event-driven API frontend architecture for internal and external data distribution. At the backend, we created a complex data engineering pipeline using innovative technologies such as Spark, Kubernetes, Docker, and microservices. This is how the pipeline was designed:

  • Several microservices were designed on a self-serve model triggered by requests coming in from internal users as well as from the outside (public).
  • For external distribution, the system was exposed to users with valid paid subscriptions only. Once the subscription was in place, several frontend APIs were exposed that enabled them to use the services on a per-request model.
  • Each microservice was able to interface with a backend analytics function that ended up performing descriptive and predictive analysis and supplying back the results.

The power of data cannot be underestimated, but the monetary power of data cannot be realized until an organization has built a solid foundation that can deliver the right data at the right time. Data engineering plays an extremely vital role in realizing this objective.



In this chapter, we went through several scenarios that highlighted a couple of important points.

Firstly, the importance of data-driven analytics is the latest trend that will continue to grow in the future. Data-driven analytics gives decision makers the power to make key decisions but also to back these decisions up with valid reasons.

Secondly, data engineering is the backbone of all data analytics operations. None of the magic in data analytics could be performed without a well-designed, secure, scalable, highly available, and performance-tuned data repository—a data lake.

In the next few chapters, we will be talking about data lakes in depth. We will start by highlighting the building blocks of effective data—storage and compute. We will also look at some well-known architecture patterns that can help you create an effective data lake—one that effectively handles analytical requirements for varying use cases.

About the Author
  • Manoj Kukreja

    Manoj Kukreja is a Principal Architect at Northbay Solutions who specializes in creating complex Data Lakes and Data Analytics Pipelines for large-scale organizations such as banks, insurance companies, universities, and US/Canadian government agencies. Previously, he worked for Pythian, a large managed service provider where he was leading the MySQL and MongoDB DBA group and supporting large-scale data infrastructure for enterprises across the globe. With over 25 years of IT experience, he has delivered Data Lake solutions using all major cloud providers including AWS, Azure, GCP, and Alibaba Cloud. On weekends, he trains groups of aspiring Data Engineers and Data Scientists on Hadoop, Spark, Kafka and Data Analytics on AWS and Azure Cloud.

    Browse publications by this author
Data Engineering with Apache Spark, Delta Lake, and Lakehouse
Unlock this book and the full library FREE for 7 days
Start now