Apache Hadoop makes it extremely easy to develop parallel programs based on MapReduce programming paradigm by taking care of work decomposition, distribution, assignment, communication, monitoring, and handling intermittent failures. However, developing Hadoop applications that linearly scale to hundreds, or even thousands of nodes, requires extensive understanding of Hadoop architecture and internals, in addition to hundreds of tunable configuration parameters. In this talk, I illustrate common techniques for building scalable Hadoop applications, and pitfalls to avoid. I will explain the seven major causes of sublinear scalability of parallel programs in the context of Hadoop, with real-world examples based on my experiences with hundreds of production applications at Yahoo! and elsewhere. I will conclude with a scalability checklist for Hadoop applications, and a methodical approach to identify and eliminate scalability bottlenecks.