We present our ongoing work building a next-generation platform for implementing machine learned data products at LinkedIn. While advanced modeling techniques and model fitting algorithms capture the bulk of attention of the research community, applying these techniques in practice reveals challenges that have so far received little attention. We are developing two systems that address these concerns: Metronome, a data product platform that provides a large set of tools for managing training and feature data and automating model fitting, evaluation and deployment, and LASER, which combines a domain-specific language for feature engineering with a highly scalable serving platform. Together, these two systems have had significant impact on developer productivity, system reliability and performance.
Speaker: Jonathan Traupman, Staff Software Engineer
Jon Traupman leads the Relevance Infrastructure team at LinkedIn, which is building the next generation of machine learning systems to power LinkedIn's data products. Previously, he led the design and implementation of LASER, a system for scalable response prediction that underlies revenue optimization of LinkedIn ads and sponsored NUS content. Prior to joining LinkedIn, he did full stack development as an engineer for Choicevendor.com and implemented ad optimization systems for adap.tv. In the now distant past, he hacked compiler tool chains for Hewlett-Packard. Jon received his Ph.D. and M.S. in Computer Science from UC Berkeley and a B.S. In Computer Science from Yale.