At Netflix, data scientists provide value to the business in several areas, from the more typical personalisation aspect to helping determine the type of content to produce or the dates at which said content should be released. Providing this business value from a ML model, however, requires the model to be used in some way, for example, through inference or by integrating its results into some other system (a UI, another service, etc.). Furthermore, ML problems come in all sizes, some will produce models that are used very frequently and continuously, while others provide a very point solution to an important problem but are used once or a very small number of times.

This presentation will focus on how, at Netflix, we enable data-scientists to fully own their model, from its inception, training all the way to providing an interface to it by deploying it as a microservice. Our infrastructure allows for the rapid prototyping and iteration between data-scientists and stakeholders as well as the seamless transition between a prototype microservice to a fully scalable and product-quality one if needed. Our infrastructure hits a sweet spot between simplicity and completeness that our data-scientists feel empowered to develop even single-use models due to its simplicity but also confident enough to use it in a more production setting.

In this presentation, we describe our infrastructure which is based on Function as a Service (FaaS), and AWS’ S3 and EC2 infrastructure. We will show how we have hit what we believe is an interesting middle ground providing a rich set of scalable functionality to the data-scientist while still allowing them to feel empowered to own their model from inception to a microservice API.