The rise of streaming in IoT, AdTech, and other applications is increasing the demand of real time inference on streaming data sets. Unfortunately, there are more possible solutions to this problem on K8s alone than there are stars in the sky. Kubeflow is a modular ML Platform that allows users to bring virtually any tool for data prep, modeling, and deployments of models, as well as real time model quality evaluation, real time model updates with no down time, and automated creation of new models based on updated data.


In this talk, we’ll discuss the challenges of implementing an ML Framework for a streaming-first organization. We will further present how we addressed these challenges, and share a modular reference architecture featuring Apache Flink on K8s as the streaming component, and models being served and updated without taking the stream offline.