Imagine that you are training and deploying hungry machine learning models. They are models of two types: batch scoring models and real time models. Your models are hungry for loads of data but also are picky and want fresh data. They are hard working though, and handle thousands of requests per second in a global distributed setup. How do you compute and serve within milliseconds the features that feed those hungry models? What about deploying the models themselves?
In this talk we will share our journey, starting from scale needs and use cases and then head into the world of feature stores, tech choices, open source, what worked, what didn’t and explore why a feature platform was a better choice also in providing autonomy to data scientists.