|By Roger Strukhoff||
|April 19, 2014 01:16 PM EDT||
No discussion of the Red Hat Summit 2014 would be complete without some discussion of Apache Hadoop. The happy elephant has now been pushing data for close to a decade, its distributed file system (HDFS) setting the tone for support of modern-day, highly distributed and very large databases in the cloud.
So I was pleased to have Robert Hodges, CEO of Hadoop-focused Continuent Tungsten, answer a few questions about his company's world.
Roger: What's the scope of the challenge you face in addressing big Hadoop deployments?
Robert: Hadoop is really very powerful as the way to concentrate and analyze information, so the key issue is how the information from existing transactional data stores gets added to Hadoop without implying additional load, application changes, or repetitive dump processes.
From our existing customer deployments, we know that the biggest challenge is getting the information into Hadoop as quickly and timely as possible from multiple different hosts simultaneously. Our customers often have many more transactional hosts running MySQL than they have Hadoop hosts, just because the scale-out and sharding required to support their transactional needs is so high.
Roger: What are the key pain points?
Robert: The key pain points are therefore the extraction of data from the transactional stores without implying additional load on these servers which are running their live customer facing website, while simultaneously loading large quantities of data that needs to be merged and analysed on the Hadoop side.
The replication solution based on Tungsten Replicator provides this very simply by placing a very low-level of load required for extraction of data, while continually streaming the changes over into Hadoop. Because this can be done on a server or cluster basis, it is easy to scale up the replication of data into Hadoop by adding more streams of replication data.
Roger: How critical is the real-time aspect of modern IT? How quickly is it growing?
Robert: It's growing very quickly, and in some cases quicker than some company IT departments and the technology they support are able to cope. Replication has for a long time been the solution for this scale-out process, but the flows of this replication data are changing.
One of the key drivers behind the adoption of Hadoop and Cassandra and similar databases is the ability to parallel process the data to get numbers in real-time. You can see this in a wide range of different markets, from banking, through to social networking and online stores.
As we get access to more information, the services supporting them need to support that an ever faster rate. We all want the lowest rate on my plane ticket purchase, while receiving the absolute best benefits and service, and all those different elements rely on real-time analysis.
Roger: What does IT think of this?
Robert: Of course, this also presents a completely different problem for the IT departments. They must deal with how to get the data into a system so that it can be analyzed quickly. The location for your active transactional dataset is not the same as your analysis tools, and may be based on completely different quantities of raw data.
Transactional databases might be conveniently sharded into 50 or 100 different RDBMS of 100GB each, but analysis needs to process all 10,000GB of data collectively to get meaningful information. That means that the IT infrastructure needs an effective way to combine and transfer this active data.
It's also clear from recent advancements in querying and processing techniques built on top of Hadoop that Hadoop itself is moving into a more real-time tool. Spark, Storm and other query engines provide very fast query and analysis on very large datasets, taking advantage of the distributed nature of Hadoop, and the increasing RAM and CPU power in evolutions of new hardware. Compatibility with Spark and similar live query mechanisms in Hadoop will form a key part of the next evolution of all Hadoop deployments.
Roger: How key is the role of Big Data in developing your solutions? How important is the term Big Data to you?
Robert: Big Data has been a significant requirement for our customers and their needs for some time, but we have definitely seen a shift recently from the scale-out, sharded nature of the typical RDBMS towards concentrating that information for analysis in Big Data stores. As that movement of data moves into the real-time it will be critical to the tools we develop to help make the transfer and management of data replication as easy as possible for our customers.
To us as the provider of the tools that enable our customers to easily share and transfer data, Big Data is therefore as important to us as it is to our customers. Of course, transactional databases are not going away, and we certainly don't expect that to change, but Hadoop and other Big Data solutions are being brought to work alongside these active data stores. Continuent will certainly be looking to expand our different solutions and techniques to bridge the gap between RDBMS and Big Data.