Can Feast manage real-time feature engineering for low-latency inference?

All answers to this question.

Yes, you can use On-Demand Feature Views for this. It allows you to define Python-based transformations that execute at the time of the request. This is perfect for features that depend on request-time data, like the distance between a user's current location and a store. However, for rolling averages or complex aggregations, you might still want to use a stream processor like Flink to push pre-calculated windows into the store. On-demand is best for simple logic; for high-frequency streaming, the integration with push sources is usually the way to go in production.

   Answered 2025-02-06 by Pamela Weaver

Are there any significant latency trade-offs when adding multiple Python-based on-demand transformations to a single inference request?

   Answered 2025-02-07 by Jeffrey Payne

• Jeffrey, there is a small overhead because it runs Python code, but for most applications, it's in the millisecond range. If your transformation is vectorized using NumPy or Pandas, it stays very efficient. Just avoid complex loops or external API calls within the transformation logic.

     Commented 2025-02-08 by Larry Henderson

It’s a great bridge between raw streaming data and the model, making the online serving layer much more flexible.

   Answered 2025-02-09 by Michelle Riley

• I agree, Michelle. It really simplifies the architecture by not forcing every single feature to be pre-computed.

     Commented 2025-02-10 by Pamela Weaver