feathr

Feature stores address a persistent pain point in ML systems: features computed for model training are often recomputed from scratch for each new model, not easily shared across teams, and inconsistent between training time and serving timeleading to training-serving skew, duplicated engineering effort, and slow iteration cycles.

Feathr provides a centralized registry and serving layer that makes features a shared organizational asset.

The platform supports both batch and real-time feature computation, with features defined once in a declarative Python SDK and then materialized to offline storage (for training) and online storage (for low-latency inference) through managed pipelines.

Teams can register feature definitions in a central metadata repository, enabling discovery and reuse of features across different model projects without each team needing to understand the underlying computation logic.

Feathr integrates with Apache Spark for large-scale batch computation and with Redis or Azure Cache for online feature serving at millisecond latency.

ML platforms teams at organizations with multiple data science teams producing models that share underlying data signals use Feathr to build a shared feature infrastructure layer.

The feature registry's discovery capabilities reduce the redundant work of different teams independently engineering the same business metrics as model features.