Apache Hadoop unquestionably delivers ROI by increasing the flexibility of data management/storage, processing, and analytics abilities. And it’s logical to assume that “commodity” servers translate into cost savings. But the truth is that while Hadoop can and does deliver significant cost reductions and revenue gains, much depends on the actual deployment.
The true total cost of ownership (TCO) of any distributed system is dependent on architecture and best practice IT operations. The cost of purchasing and maintaining hardware aside, performance, reliability and scalability are critical to real TCO—a sluggish, unstable system may cost less for its physical structure, but will drain capital as well as operational expense resources from the company.
Hadoop itself is not a magical solution that makes data management faster, easier, and cheaper. The architectural differences between Hadoop distributions can save companies 20-50% in TCO.
Know Your Hadoop Distribution
Hadoop was created to fill a specific need. To grow and prosper on the web, companies such as Yahoo and Facebook needed a way to efficiently work with very large, dissimilar data sets. Hadoop was the solution.
The open source community further developed Hadoop, and applications for the platform, to meet various data-intensive use cases. As Hadoop gained users, attention turned to ways to tune and refine the platform to reduce costs while improving the performance of the Hadoop parallel processing architecture. For mission-critical business use, Apache Hadoop’s reliance on NameNodes and its lack of a read-write file system created problems in terms of scalability, reliability, and performance. MapR created a file system to make Hadoop enterprise-ready and to remove the single point of failure-the NameNode.
The difference Hadoop can make in TCO when you consider data storage costs alone is impressive. The average cost to store data in a data warehouse is greater than $10,000/terabyte. The cost for a mature enterprise-grade distribution of Hadoop is less than $1,000/terabyte. But how available is that data?
One of Hadoop’s biggest TCO-killers is its use of NameNodes. A NameNode holds metadata about the information stored on all the DataNodes in a cluster. This architecture creates huge data access bottlenecks for businesses, and also creates a single point of failure in which a single job processing 10 million files could make an entire cluster inoperable.
There is a workaround for the latter issue: implementing the journaling NameNode concept removes the single failure threat. Unfortunately setting this up is a complicated procedure which requires an additional investment in hardware, and it doesn’t mitigate the processing bottleneck problem. To solve this problem effortlessly, look for a distribution that forgoes NameNodes in favor of distributed metadata architecture like MapR.
NameNodes also affect Hadoop’s scalability—single clusters can only scale as large as their NameNode, which means a lot of time, resources, and skill are required to ensure the best utilization of blocks. Unmanaged, a 1 MB file could consume a 2 GB block. But when compressed or sequenced to bypass scalability problems, data may not be as readily available for use in analysis.
A mature, enterprise-ready distribution of Hadoop eliminates block size limitations as well as limitations that exist for the total number of files the cluster can manage. It will also utilize a read-write file system to achieve high availability and low latency reliability.
Evaluating For TCO
When evaluating Hadoop distributions for enterprise use, factor in the setup and maintenance time costs along with price of the hardware. Find the sweet spot—the distribution that can deliver higher performance on less hardware. Look for distributions that can perform operations and analytics on the same platform, and that support volumes, making large amounts of data easier to manage.
Look also for features that will enable your Hadoop deployment to continue to deliver strong TCO. For example, a distribution that supports Apache Hadoop’s APIs, as well as multiple versions of key Hadoop components (rather than a single version of Hive, Spark, and Flume, etc.), will help you avoid forced upgrades down the road.
You’ll also want to make sure that the distribution you are evaluating doesn’t make your data a hostage to an exclusive data format. Additionally, since many enterprises utilize applications that work in conjunction with NFS, look for full NFS compatibility to leverage the skills and solutions that you already have in house.
Your Cluster, Your TCO
The information provided above is, by necessity, general in nature. If you’d like to gauge the TCO of Hadoop distributions using your own data, across a number of variables in different scenarios, you can do so with this TCO Calculator for Hadoop. Simply input your estimated volume of data, number of files, and data growth per year, then test different scenarios by adjusting variables across hardware costs, software costs, staffing, and environmental issues.
