Ways to bring features to production
Now that we understand the need for features in production, let's look at some traditional ways of bringing features to production. Let's consider two types of pipelines: batch model pipelines and online/transactional model pipelines:
- Batch models: These are models that are run on a schedule, such as hourly, daily, weekly, and so on. Two of the common batch models are forecasting and customer segmentation. Batch inference is easier and less complex than its counterpart since it doesn't have any latency requirements; inference can run for minutes or hours. Batch models can use distributed computational frameworks such as Spark. Also, they can be run with simple infrastructure. Most ML models start as batch models and, over time, depending on the available infrastructure and requirements, they go on to become online/transactional models.
Though batch models' infrastructure is simple to build and manage, these models...