Great talk on Spark
For handling huge datasets Spark is evolutionary or revolutionary depending on your point of view. A bit of personal history before I talk specifically about Spark:
In the late 1980s I was an architect and developer on a multinational project to use seismic data from 38 data collection stations to detect atomic bomb tests. All of our data handling software was custom; if we had Spark, or even Hadoop, we would have saved a ton of effort. Similarly, in the 1990s I was tech lead on a fraud detection system that used massive real time telephone records data sets. Modern infrastructure would have saved a lot of time and money.
My first serious use of map reduce was processing large Twitter data sets at Compass Labs. We used Hadoop on Amazon ElasticMapreduce. Later when I worked as a contractor at Google, in addition to using map reduce, I was introduced to realtime interactive tools like Dremel that made it easy to interactively use large data sets.
With Spark, everyone gets to interactively work with massive datasets! I think that Spark is evolutionary in that it builds on and plugs into existing work like the Hadoop File Sytem and supports familiar map reduce style operations. I think that it is revolutionary in the memory based distributed architecture and application programming model. Spark was designed based on limitations of map reduce systems like Hadoop that while providing easy to use programming models, have ineffiencies in data access. With Spark, you have an easy to use programming model, more efficiency, and built in interactivity. I have examples of using Spark in my last book Power Java. You can experiment with Spark on your laptop and only worry about accessing a cluster when you need to scale.