Batch and stream processing are two types of methods used to process data, which is one of the steps in feature engineering. The processed data can be used in features for generating either real-time or batch predictions.
In Batch processing, historical static data is processed in batches for use in features. Batches might run at scheduled intervals, or run when compute resources are available. With batch processing, heavy feature computations can be performed offline, so that it’s ready for fast inference. However, features do become stale, as changes in the real environment shift over time. It’s important to set up a drift-aware monitoring system for these features that includes a retraining step.
In Stream Processing, predictions are made on real-time inputs with near real-time or streaming features for a given entity.
Predictions made with stream processing and therefore streaming features can improve the quality of predictions by adding more valuable data to the model. For instance, a recommender system with streaming features can make use of a user’s recent website behavior, combined with data like real-time inventory or purchase history.
Stream processing is quite a bit more complicated to embed into an ML workflow, and requires a real-time feature store, mature streaming infrastructure, and an efficient development environment so data scientists and ML engineers can work together to validate new streaming features.