If you’re running Big Data applications, you’re going to want to look at some kind of distributed processing system. Hadoop is one of the best-known clustering systems, but how are you going to process all your data in a reasonable time frame? Apache Spark offers services that go beyond a standard MapReduce cluster.

A choice of job styles
MapReduce has become a standard, perhaps the standard, for distributed file systems. While it’s a great system already, it’s really geared toward batch use, with jobs needing to queue for later output. This can severely hamper your flexibility. What if you want to explore some of your data? If it’s going to take all night, forget about it.

With Apache Spark, you can act on your data in whatever way you want. Want to look for interesting tidbits in your data? You can perform some quick queries. Want to run something you know will take a long time? You can use a batch job. Want to process your data streams in real time? You can do that too.

The biggest advantage of modern programming languages is their use of interactive shells. Sure, Lisp did that back in the ‘60s, but it was a long time before the kind of power to program interactively became available to the average programmer. With Python and Scala you can try out your ideas in real time and develop algorithms iteratively, without the time-consuming write/compile/test/debug cycle.

RDDs
The key to Spark’s flexibility is the Resilient Distributed Datasets, or RDDs. RDDs maintain a lineage of everything that’s done to your data. They’re fine-grained, keeping track of all changes that have been made from other transformations such as map or join. This means that it’s possible to recover from failures by rebuilding from these transformations (which is why they’re called Resilient Distributed Datasets).

RDDs also represent data in memory, which is a lot faster than always pulling data off of disks—even with SSDs making their way into data centers. While having your data in memory might seem like a recipe for slow performance, Spark uses lazy evaluation, only making transformations on data when you specifically ask for the result. This is why you can get queries so quickly even on very large datasets.

You might have recognized the term “lazy evaluation” from functional programming languages like Haskell. RDDs are only loaded when specific actions produce some kind of output; for example, printing to a text file. You can have a complex query over your data, but it won’t actually be evaluated until you ask for it. And the query might only find a specific subset of your data instead of plowing through the whole thing. This lazy evaluation lets you create complex queries on large datasets without incurring a performance penalty.

RDDs are also immutable, which leads to greater protection against data loss even though they’re in memory. In case of an error, Spark can go back to the last part of an RDD’s lineage and recover from there rather than relying on a checkpoint-based system on a disk.

Spark and Hadoop, Not as Different as You Think
Speaking of disks, you might be wondering whether Spark replaces a Hadoop cluster. That’s really a false dichotomy. Hadoop and Spark work together. While Spark provides the processing, Hadoop handles the actual storage and resource management. After all, you can’t store data in your memory forever.

With the combination of Spark and Hadoop in the same cluster, you can cut down on a lot of overhead in maintaining different clusters. This combined cluster will give you unlimited scale for Big Data operations.

Who’s Using Spark?
When you have your Big Data cluster in place, you’ll be able to do lots of interesting things. From genome sequencing analysis, to digital advertising to a major credit card company who uses Spark to match thousands of transactions at once for possible fraud detection. Cisco does something similar with a cloud-based security product to spot possible hacking before it turns into a major data breach. Geneticists use it to match genes to new medicines.

Conclusion
Apache Spark builds on Hadoop and then goes beyond it by adding stream processing capabilities. The MapR distribution is the only one that offers everything you need right out of the box to enable real-time data processing.

For a more in-depth view into how Spark and Hadoop benefit from each other, read chapter four of the free interactive ebook: Getting Started with Apache Spark: From Inception to Production, by James A. Scott.

• Guavus Enhances Its Reflex Operational Intelligence Platform With Apache Spark and Hadoop YARN
• Apache Spark Monitoring in SPM
• Guavus and Databricks Announce Reflex Platform Now a Certified Spark Distribution