At LinkedIn, we use Hadoop as our backbone for big data analytics and machine learning. With an exponentially growing data volume, and the company heavily investing in machine learning and data science, we have been doubling our cluster size year over year to match the compute workload growth. Our largest cluster now has ~10,000 nodes, one of the largest (if not the largest) Hadoop clusters on the planet. Scaling Hadoop YARN has emerged as one of the most challenging tasks for our infrastructure over the years.
In this blog post, we will first discuss the YARN cluster slowdowns we observed as we approached 10,000 nodes and the fixes we developed for these slowdowns. Then, we will share the ways we proactively monitored for future performance degradations, including a now open-sourced tool we wrote called DynoYARN, which reliably forecasts performance for YARN clusters of arbitrary size. Finally, we will describe Robin, an in-house service which enables us to horizontally scale our clusters beyond 10,000 nodes.
Read on to learn about the problems they experienced and how they resolved them.