A classifier labeling Van Gogh drawings as invoices and a chatbot insulting users on Twitter are only two examples of Machine Learning (ML) models, which went wild as soon as they hit production. Although evaluated on a test set, in the face of unseen data ML models oftentimes behave in an unpredictable way. Depending on the application, such a model may lead to decreasing revenue, bad reputation or even a threat to the health of people.

To ensure a stable rollout of new models into production, we have to promote ML models to first class citiziens in the Continuous Delivery pipeline. Kubernetes and Kafka are two great tools to support the rollout of new machine learning models in a (semi-) automated way. I will show a pipeline built with these tools, which will lead to more confidence in your deployments and happier users.