A model trained once on a finite dataset may not keep pace with changing data. Online learning takes a different approach: it updates a model as new data arrives. Flink ML supports both approaches through one API for machine-learning pipelines across batch and streaming data. section 8.4.2
A pipeline can keep feature engineering, preprocessing, training, and inference within Flink. The same code can support offline training on historical data and real-time learning on a stream. This is useful when a model must adapt to evolving patterns rather than rely only on periodic batch training. section 8.4.2
Flink ML offers two ways to supply the learning step:
- Use pre-implemented algorithms for regression, classification, clustering, or feature extraction. section 8.4.2
- Implement a custom algorithm through its interfaces when the supplied options do not fit the task. section 8.4.2
