Hadoop is everywhere. For better or worse, it has become synonymous with big data. In just a few years it has gone from a fringe technology to the de facto standard. Want to be big bata or enterprise analytics or BI-compliant? You better play well with Hadoop.
It’s therefore far from controversial to say that Hadoop is firmly planted in the enterprise as the big data standard and will likely remain firmly entrenched for at least another decade. But, building on some previous discussion, I’m going to go out on a limb and ask, “Is the enterprise buying into a technology whose best day has already passed?”
First, there were Google File System and Google MapReduce
To study this question we need to return to Hadoop’s inspiration – Google’s MapReduce. Confronted with a data explosion, Google engineers Jeff Dean and Sanjay Ghemawat architected (and published!) two seminal systems: the Google File System (GFS) and Google MapReduce (GMR). The former was a brilliantly pragmatic solution to exabyte-scale data management using commodity hardware. The latter was an equally brilliant implementation of a long-standing design pattern applied to massively parallel processing of said data on said commodity machines.
GMR’s brilliance was to make big data processing approachable to Google’s typical user/developer and to make it fast and fault tolerant. Simply put, it boiled data processing at scale down to the bare essentials and took care of everything else. GFS and GMR became the core of the processing engine used to crawl, analyze, and rank web pages into the giant inverted index that we all use daily at google.com. This was clearly a major advantage for Google.
Enter reverse engineering in the open source world, and, voila, Apache Hadoop — comprised of the Hadoop Distributed File System and Hadoop MapReduce — was born in the image of GFS and GMR. Yes, Hadoop is developing into an ecosystem of projects that touch nearly all parts of data management and processing. But, at its core, it is a MapReduce system. Your code is turned into map and reduce jobs, and Hadoop runs those jobs for you.
Then Google evolved. Can Hadoop catch up?
Most interesting to me, however, is that GMR no longer holds such prominence in the Google stack. Just as the enterprise is locking into MapReduce, Google seems to be moving past it. In fact, many of the technologies I’m going to discuss below aren’t even new; they date back the second half of the last decade, mere years after the seminal GMR paper was in print.
Here are technologies that I hope will ultimately seed the post-Hadoop era. While many Apache projects and commercial Hadoop distributions are actively trying to address some of the issues below via technologies and features such as HBase, Hive and Next-Generation MapReduce (aka YARN), it is my opinion that it will require new, non-MapReduce-based architectures that leverage the Hadoop core (HDFS and Zookeeper) to truly compete with Google’s technology. (A more technical exposition with published benchmarks is available at http://www.slideshare.net/mlmilleratmit/gluecon-miller-horizon.)
Percolator for incremental indexing and analysis of frequently changing datasets. Hadoop is a big machine. Once you get it up to speed it’s great at crunching your data. Get the disks spinning forward as fast as you can. However, each time you want to analyze the data (say after adding, modifying or deleting data) you have to stream over the entire dataset. If your dataset is always growing, this means your analysis time also grows without bound.
So, how does Google manage to make its search results increasingly real-time? By displacing GMR in favor of an incremental processing engine called Percolator. By dealing only with new, modified, or deleted documents and using secondary indices to efficiently catalog and query the resulting output, Google was able to dramatically decrease the time to value. As the authors of the Percolator paper write, ”[C]onverting the indexing system to an incremental system … reduced the average document processing latency by a factor of 100.” This means that new content on the Web could be indexed 100 times faster than possible using the MapReduce system!
Coming from the Large Hadron Collider (an ever-growing big data corpus), this topic is near and dear to my heart. Some datasets simply never stop growing. It is why we baked a similar approach deep into the Cloudant data layer service, it is why trigger-based processing is now available in HBase, and it is a primary reason that Twitter Storm is gaining momentum for real-time processing of stream data.
Dremel for ad hoc analytics. Google and the Hadoop ecosystem worked very hard to make MapReduce an approachable tool for ad hoc analyses. From Sawzall through Pig and Hive, many interface layers have been built. Yet, for all of the SQL-like familiarity, they ignore one fundamental reality – MapReduce (and thereby Hadoop) is purpose-built for organized data processing (jobs). It is baked from the core for workflows, not ad hoc exploration.
In stark contrast, many BI/analytics queries are fundamentally ad hoc, interactive, low-latency analyses. Not only is writing map and reduce workflows prohibitive for many analysts, but waiting minutes for jobs to start and hours for workflows to complete is not conducive to the interactive experience. Therefore, Google invented Dremel (now exposed as the BigQuery product) as a purpose-built tool to allow analysts to scan over petabytes of data in seconds to answer ad hoc queries and, presumably, power compelling visualizations.
Google’s Dremel paper says it is “capable of running aggregation queries over trillions of rows in seconds,” and the same paper notes that running identical queries in standard MapReduce is approximately 100 times slower than in Dremel. Most impressive, however, is real world data from production systems at Google, where the vast majority of Dremel queries complete in less than 10 seconds, a time well below the typical latencies of even beginning execution of a MapReduce workflow and its associated jobs.
Interestingly, I’m not aware of any compelling open source alternatives to Dremel at the time of this writing and consider this a fantastic BI/analytics opportunity.
Pregel for analyzing graph data. Google MapReduce was purpose-built for crawling and analyzing the world’s largest graph data structure – the internet. However, certain core assumptions of MapReduce are at fundamental odds with analyzing networks of people, telecommunications equipment, documents and other graph data structures. For example, calculation of the single-source shortest path (SSSP) through a graph requires copying the graph forward to future MapReduce passes, an amazingly inefficient approach and simply untenable at scale.
Therefore, Google built Pregel, a large bulk synchronous processing application for petabyte -scale graph processing on distributed commodity machines. The results are impressive. In contrast to Hadoop, which often causes exponential data amplification in graph processing, Pregel is able to naturally and efficiently execute graph algorithms such as SSSP or PageRank in dramatically shorter time and with significantly less complicated code. Most stunning is the published data demonstrating processing on billions of nodes with trillions of edges in mere minutes, with a near linear scaling of execution time with graph size.
At the time of writing, the only viable option in the open source world is Giraph, an early Apache incubator project that leverages HDFS and Zookeeper. There’s another project called Golden Orb available on GitHub.
In summary, Hadoop is an incredible tool for large-scale data processing on clusters of commodity hardware. But if you’re trying to process dynamic data sets, ad-hoc analytics or graph data structures, Google’s own actions clearly demonstrate better alternatives to the MapReduce paradigm. Percolator, Dremel and Pregel make an impressive trio and comprise the new canon of big data. I would be shocked if they don’t have a similar impact on IT as Google’s original big three of GFS, GMR, and BigTable have had.
Mike Miller (@mlmilleratmit) is chief scientist and co-founder at Cloudant, and Affiliate Professor of Particle Physics at University of Washington.