Enable Production ML with Databricks Feature Store
- 데이터 사이언스, 머신 러닝 및 MLOps
- Moscone South | Level 3 | 306
- 35 min
Productionalizing ML models is hard. In fact, very few ML projects make it to production, and one of the hardest problems is data! Most AI platforms are disconnected from the data platform, making it challenging to keep features constantly updated and available in real-time. Offline/online skew prevents models from being used in real-time or, worse, introduces bugs and biases in production. Building systems to enable real-time inference requires valuable production engineering resources. As a result of these challenges, most ML models do not see the light of day.
Learn how you can simplify production ML using Databricks Feature Store, the first feature store built on the data lakehouse. Data sources for features are drawn from a central data lakehouse, and the feature tables themselves are tables in the lakehouse, accessible in Spark and SQL for both machine learning and analytics use cases. Features, data pipelines, source data, and models can all be co-governed in a central platform. Feature Store is seamlessly integrated with Apache Spark™, enabling automatic lineage tracking, and with MLflow, enabling models to look up feature values at inference time automatically. See these capabilities in action and how you can use it for your ML projects.