How did you ensure that AI features remain consistent and reliable during model training and inference, especially with evolving data schemas and streaming data?

I use a feature‑store that separates raw and derived features and keeps a versioned schema registry. During training, the feature store is queried with a specific schema version that matches the training data snapshot, ensuring that the model sees exactly the same feature set it was trained on. For inference, the same version is pinned to the serving layer, so even if the underlying data schema evolves, the model receives the same feature format. Late events are handled by buffering them in a Kafka topic and replaying them to the feature store only after a watermark threshold. I also implement feature lineage tracking so that any change in the feature definition triggers a re‑training alert. This strategy guarantees consistency between training and inference while allowing the data pipeline to evolve.