Cloudera floats commercial Hadoop distro
Open source for the Google wannabe
Face it: You want to launch your own Google and get your hands on some of that (easy?) internet money. Well, now's your chance to take a stab at it.
Today, a startup called Cloudera is launching a commercial distribution of the Google-inspired open source Hadoop software underpinning Yahoo, Facebook, and a number of other hot-shot web companies.
The Cloudera team includes four founders, all of which bring different things to the Hadoop table: Christophe Bisciglia, who led the partnership between Google, IBM, and the National Science Foundation to create Hadoop grids for academics to play around with; Amr Awadallah, a former Yahoo vice president of engineering that led the data warehousing and analytics effort behind that company's mail, search, finance, and news services; - Mike Olson, formerly the chief executive officer of open source database maker Sleepycat Software (now owned by Oracle); and Jeff Hammerbacher, formerly of social networking giant Facebook and the manager who created the Hive project, which is a data warehousing layer that works in conjunction with Hadoop and that Facebook uses to do data analysis on its many petabytes of information stored in its user data warehouse.
Hammerbacker is an entrepreneur in residence at venture capitalist Accel Partners, and back in October, Accel kicked in $5m in Series A funding for Cloudera. The startup has also tapped Hadoop creators Doug Cutting and Mike Cafarella as advisors as well as Diane Green (founder and former CEO at virtualization specialist VMware) and Marten Mikos (the former CEO of MySQL before Sun Microsystems bought it). These and a handful of other tech luminaries are not just advisors, but investors in Cloudera.
According to Christophe, the Hadoop stack that Cloudera is supporting is based on the latest stable releases of the code that is available through the Apache Foundation, where the open source version of Hadoop lives. This includes Hadoop 0.18.3, which has the Hadoop Distributed File System - as the name suggests, a distributed and fault-tolerant file system - and the MapReduce application parallelization and execution environment that works in conjunction with HDFS.
The Cloudera Hadoop distro will also include the Hive client library associated with Hadoop (and also available through Apache), but according to Christophe, it doesn't really have version numbers yet. The important thing is that Cloudera found a set of Hive code that works with Hadoop 0.18.3 and that Hive includes a query language called HQL, which allows Hadoop data sets to be queried in a manner that is similar to SQL queries against a relational database.
Olson says that Cloudera was founded last summer, and the company is clearly ramping quickly if it has already secured so much financial and technical backing. And the reason is simple: People want to figure out how to use Hadoop in their own IT operations, but it is a pain in the neck to get it all set up and working.
"Adoption of Hadoop has been slow in mainstream computing because it is still hard to install, build, and maintain a Hadoop cluster," explains Olson. "We are convinced that normal companies are going to be coping with terabytes and petabytes of data, and Hadoop is the most interesting thing to come along in a decade for dealing with large data sets. We want to be the Hadoop company that enterprises come to when they want to crunch those big data sets."
One of the things that Cloudera started doing with its beta customers when it started alpha trials of its services last fall was get everyone on the same release of Hadoop and Hive. And standardization means not being on the bleeding edge, by the way. Hadoop 0.19 is out, and according to Christophe, it has many needed features. But "much-needed features have come with bugs." These may be shaken out in Hadoop 0.20, but commercial companies that are basing their business on this software don't want to mess around with code.
The analogy with Linux is plain enough: They want something akin to the hardened and slow-changing Red Hat Enterprise Linux, not the Fedora development release.
Hadoop is written in Java, which means it can run on any Java-enabled platform, but Christophe says that 90 per cent of companies deploy it on a Linux operating system and most deploy it on x64 iron. The Cloudera Hadoop distro is packaged up in Red Hat-style RPMs, and the Hadoop functions are available as Linux services, just like a Web server is, for instance. The Cloudera package, technically known as the Cloudera Distribution for Hadoop, is also available as an Amazon EC2 image. Given that all of the Hadoop code is open source, the Cloudera packages are all available for free thanks to open source Apache 2 licenses governing the code.
Cloudera plans to make money selling consulting, training, and support, just like Linux distros do. Pricing has not been announced yet, and Olson was pretty stubborn about the need to keep pricing private until Cloudera gets some more business under its belt. The current pricing metrics, he did say, were based on the size of the Hadoop cluster, including the number of servers and the size of the data sets. ®