Productionizing custom ML models with MLOps

ML systems have unique characteristics that differentiate them from traditional software. They require the testing and validation of both code and data, they have unique ways of measuring quality and evaluating performance, and deployed ML models typically degrade over time if they don't continuously evolve. Moreover, observability becomes difficult since systems can underperform without throwing errors or showing signs of it. Therefore, managing and operating ML models can be challenging.

In Chapter 9, Jumping on the DevOps Bandwagon with Site Reliability Engineering (SRE), we've discussed DevOps principles and how they can help improve the reliability of systems and shorten development cycles. As data science and ML became crucially important capabilities for modern enterprises, applying a similar set of principles to ML systems has become a priority for many. Hence, the Machine Learning Operations (MLOps) paradigm emerged...