To really comprehend the history of YARN, you have to start by taking a close look at the evolution of Hadoop itself. Yahoo! adopted Apache Hadoop in 2006 to replace the existing infrastructure that was then driving its WebMap application—the technology that builds a graph of the known web to power its search engine. At that time, the web-graph contained more than 100 billion nodes with roughly 1 trillion edges. The previous infrastructure, named “Dreadnaught,” successfully served its purpose and grew well—starting from a size of just 20 nodes and expanding to 600 cluster nodes—but had reached the limits of its scalability. The software also didn’t perform perfectly in many scenarios, including handling of failures in the clusters’ commodity hardware. A significant shift in its architecture was required to scale out further to match the ever-growing size of the web. The distributed applications running under Dreadnought were very similar to MapReduce programs and needed to span clusters of machines and work at a large scale. This highlights the first requirement that would survive throughout early versions of Hadoop MapReduce, all the way to YARN—[Requirement 1] Scalability.
[Requirement 1] Scalability
The next-generation compute platform should scale horizontally to tens of thousands of nodes and concurrent applications.
For Yahoo!, by adopting a more scalable MapReduce framework, significant parts of the search pipeline could be migrated easily without major refactoring—which, in turn, ignited the initial investment in Apache Hadoop. However, although the original push for Hadoop was for the sake of search infrastructure, other use-cases started taking advantage of Hadoop much faster, even before the migration of the web-graph to Hadoop could be completed. The process of setting up research grids for research teams, data scientists, and the like had hastened the deployment of larger and larger Hadoop clusters. Yahoo! scientists who were optimizing advertising analytics, spam filtering, personalization, and content initially drove Hadoop’s evolution and many of its early requirements. In line with that evolution, the engineering priorities evolved over time, and Hadoop went through many intermediate stages of the compute platform, including ad hoc clusters.