Cloud-based tools, including large-scale data-intensive computing as offered by Hadoop, are key to the rise and rise of cloud computing. In this wide-ranging Exclusive Q&A with SYS-CON's Cloud Computing Journal, the Director of Grid Services at Yahoo! - Rob Weltman - explains to Jeremy Geelan, Conference Chair of SYS-CON's 1st International Cloud Computing Conference & Expo held last week in San Jose, CA, how analyzing and learning from ever-growing volumes of business data is essential to continuously refining and improving service offerings.

Cloud Computing Journal: Yahoo! has been the largest contributor to the Hadoop project and uses Hadoop extensively in its Web search and advertising businesses. Can you explain a little of the background to that?
Rob Weltman: Yahoo! Search (and before it Inktomi) was a pioneer in using large clusters of commodity computers to speed up the crawling and indexing of Web sites. While working on the architecture and design of the next generation of Web Search crawling and indexing, we came in touch with Doug Cutting and the open source Lucene project for text indexing/search. Lucene contained a distributed file system with integrated computation using the map-reduce paradigm. It looked very promising and appropriate for many data-intensive applications. Hadoop was then split out as its own project. Yahoo! supported Hadoop in a big way, both in contributing to its development as an open source project and in applying it to solve many large-scale data/computation problems in the company.

Hadoop has matured at an amazingly fast pace. From a 20-node cluster two years ago, to many 2,000-node clusters today; from a somewhat embarrassing terasort (a benchmark) performance to the terasort leader; from a no-access control to user- and group-owned files and directories. There is now a high-level language - Pig - that allows you to express complex operations on data in an intuitive way and have them translated into Hadoop map-reduce jobs.

In 2007, Hadoop at Yahoo! was used primarily for research - analyzing enormous volumes of data to find the best algorithms and parameters for selecting search results or ads to present to users. Now it is also a central component in many production operations, including Web Search, ad serving, and personalization.

Cloud Computing Journal: Are cloud-based tools like Hadoop the most important kinds of tools for the future, do you think?
RW: Being able to add capacity as needed without major software or infrastructure changes is clearly important for many organizations. Sharing resources and dynamically allocating more or less to various functions on demand is highly attractive as companies strive to control costs while the computing needs grow and shift. Analyzing and learning from ever-growing volumes of business data is essential to continuously refining and improving service offerings. The ability to quickly explore new algorithms and put them into production will be a competitive advantage for those with the resources to apply them. All of these speak to the importance of Cloud Computing,

Cloud Computing Journal: How important a role does Java play in the project? Is that because of the need to scale horizontally (and massively)?
RW: Hadoop supports programming and scripting in many languages. Hadoop, itself, is written in Java. The language provides strong support for the central infrastructure needs of system and network programming. There is a large body of experience in developing robust, performance-optimized, scalable platforms in Java.

Java provides portability to many hardware and software environments however Hadoop's horizontal scalability is not a result of the choice of language but rather of a design that is strongly focused on fault-tolerance and distribution.

Cloud Computing Journal: Is the Yahoo! Search Webmap still the world's largest Hadoop production application so far as you are aware? Can you share some size data about Webmap with us?
RW: Yes, as far as I know, the Yahoo! WebMap is the largest Hadoop application in production. It uses 2,000+ computers and is still continuously growing. It produces 300TB of data per run, including 1.2 trillion links.

Cloud Computing Journal: How important are Hadoop clusters to Yahoo! Overall? Do your Web search queries depend on them?
RW: Hadoop isn't directly involved in responding to queries typed in by users, but it is responsible for much of the backend work that produces the indexes used to service those queries. If the Hadoop clusters were down, the quality of search results would quickly degrade as the indexes became stale.

Cloud Computing Journal: Who else besides Yahoo! uses Hadoop to run large distributed computations?
RW: Many of the major Hadoop users are listed at http://wiki.apache.org/hadoop/PoweredBy. Facebook has several hundred nodes in a cluster for backend processing and analysis. Quantcast has several thousand cores in a very large cluster. Many companies, including AOL, A9 (Amazon), and IBM have deployed somewhat smaller clusters. It's likely that almost all of the uses involve large quantities of data.

Cloud Computing Journal: Can Hadoop be run on Amazon EC2?
RW: Absolutely! There is a ready-to-run AMI (virtual machine definition for EC2) for Hadoop. Among many others, Powerset (now owned by Microsoft) runs on EC2.

Cloud Computing Journal: What about Sun's Grid Engine - can it also be run on that?
RW: Yes, Hadoop works with Sun's Grid Engine but you lose the benefit of data locality (putting the computation of each piece of a distributed job near the data needed by that piece).

Cloud Computing Journal: Does the Hadoop team have any kind of a blog or forum?
RW: We have a blog at http://developer.yahoo.net/blogs/hadoop/. The team is also heavily engaged in the user and developer Hadoop mailing lists at hadoop.apache.org.

Cloud Computing Journal: Doug Cutting named it after his child's stuffed elephant. Is there any downside to an Enterprise IT tool having the name of a stuffed elephant?
RW: I did get some ribbing during the election period when I wore my Hadoop Summit t-shirt with the elephant on it, but I was able to clarify Hadoop's open source and non-partisan nature.

Cloud Computing Journal: What else have you and your team developed at Yahoo!, in terms of data-analytics applications for example?
RW: The Grid Computing development team at Yahoo! works on the Hadoop core software, the Pig high-level language, the ZooKeeper distributed coordination service, and the Chukwa monitoring and metric analysis system. In addition, it provides various Hadoop add-ons and tools to e.g. facilitate joining of very large data sets or to understand and improve the performance and efficiency of Hadoop jobs. We provide consulting to application teams that develop large-scale Hadoop programs (often involving feature extraction, modeling, optimization, and index creation) but do not produce them ourselves.