A model training row represents a prediction made with information available at that instant. Feature transformations, joins and updates need to respect that assumption. For online inference, there is another constraint: can the service retrieve those values within its latency budget?
In this chapter
- Numerical and categorical feature preparationTransformations, missingness, target encoding, fitted statistics and training parity.
- Point-in-time joins and availability leakageChoose historical feature versions using both event and availability timestamps.
- Offline and online feature storesHistorical training, low-latency serving, materialization and consistency checks.