Algorithms and architecture for job recommendations

In this article, we'll describe the evolution of our recommendation engine, from the initial minimum viable product (MVP) built with Apache Mahout, to a hybrid offline online pipeline. We'll explore the impact these changes have had on product metrics and how we've addressed challenges by using incremental modifications to algorithms, system architecture, and model format. To close, we'll review some related lessons in system design that apply to any high-traffic machine learning application. Indeed's production applications run in many data centers around the world. Clickstream data, and other application events from every data center, are replicated into a central HDFS repository, based in our Austin data center.