You are here:  Home  >  Data Education  >  BI / Data Science News, Articles, & Education  >  BI / Data Science Blogs  >  Current Article

Are Spark and Hadoop Allies, or Are They Enemies?

By   /  March 2, 2016  /  No Comments

Click here to learn more about Ariel Amster.

Any discussion about big data and the merits of analytics will in all likelihood eventually involve the topics of Hadoop and Spark. This discussion more than often becomes a debate over which is better. After all, both share similarities in that they are both big data frameworks. Many businesses feel like they have to choose one or the other in order to effectively utilize the big data they collect. While there is certainly merit for the ongoing Spark vs. Hadoop conversation, others have taken the more diplomatic approach, stating that while the two share similarities, there are enough differences as to make the need to choose between them unnecessary. It’s not a question of figuring out which one is better but rather which one is best for specific situations. In a sense, this debate comes down to a fundamental question: are Spark and Hadoop allies, or are they enemies? Answering this question can help businesses make a more informed decision over what to do with their big data.

The very debate of Apache Spark and Apache Hadoop arose out of the differences between the two. Hadoop at the most basic level is a distributed data infrastructure. It takes the data organizations collect and distributes them across different nodes located within different servers. Considering the amount of data regularly gathered these days, Hadoop has proven instrumental in getting a handle on big data. Spark, on the other hand, acts as a processing package that works on the data collections found on Hadoop clusters. Spark does not have a system it uses for organizing distributed files, instead running on top of the distributed files provided by Hadoop. Hadoop meanwhile can support Spark, though it usually works with MapReduce. This is where the “choose one or the other” mentality originates from. Spark also carries with it a number of attributes that appear to work in its advantage. Some of them include being able to process extremely large amounts of data, showcasing resiliency, performing machine learning at a higher level, and most of all, being really really fast.

On the surface, Spark and Hadoop appear to be competing for the same purpose, thus necessitating a decision in favor of only one, but a closer examination shows that both Spark and Hadoop work best at different tasks. Yes, Spark may perform data processing more quickly than MapReduce, but it isn’t always the better option. Spark is best when a business is streaming data and needs it processed in real time. Some of those use cases can include fraud detection, cybersecurity analytics, online product recommendations for customers, and networking monitoring, among many others. But some businesses might not need to stream their data. Perhaps they don’t have a need to distribute their data across multiple different nodes and servers. If data is stored on a disk, for example, Hadoop and MapReduce can handle the task just fine. If a business doesn’t need real time analysis, like performing their own analytics once or twice each day, Hadoop can get the job done, in some cases better than Spark can.

It would simply be incorrect to look at Spark and Hadoop as competing interests. Spark is excellent for some tasks, while Hadoop is good for others. If anything, businesses should view Spark as an extra feature that can be added to the Hadoop infrastructure when the need arises. When speed is needed for data science applications, Spark can kick in and perform the processes needed to reach valuable insights. If businesses only need to perform a limited number of things with their data, Hadoop will run fine on its own. In other cases, such as with the Internet of Things (IoT), Spark and Hadoop actually work well together, acting as complementary pieces to make the work more efficient while performing fast analytics.

It’s probably best to think of Spark and Hadoop as players on the same team. They each have their specialties and excel in different areas, but they’re both striving toward the same goal. They are most certainly not mutually exclusive, and one won’t likely push out the other in terms of capabilities. If businesses use them in tandem, they’ll get the most out of the big data collected from multiple sources.

About the author

Ariel Amster, Director of Sales Channels for Qubole: Ariel has over twenty years of business development, sales, consulting, product management and business analysis experience in Technology and Finance related businesses and has held various management, consulting and product management positions at companies such as Qubole, First Derivatives, Citi, CheckFree (Fiserv) and marchFirst (USWeb). He is currently the Director of Sales Channels for Qubole, a Big Data company that makes it easy for data scientists and data engineers to manage analytics at scale on the public cloud. Previously, he has provided consulting services and held positions in Sales & Trading divisions of Tier 1 investment banks as well as being a principle and entrepreneur in a technology start-up, Experient. His experience includes, technology product development, business strategy, account management, change management, and project management within institutional finance (prime brokerage), consumer finance (credit card/payments/unsecured lending), and technology companies (Big Data, portfolio accounting, content management). Ariel holds a BA from UCLA with honors.

You might also like...

Machine Learning Moves to the Hybrid Cloud

Read More →