'Why use a an ML featurestore if you have novel incoming data?
If I have a customer database I fully get the use of having a feature store, since anyone at the company who wants to do any modeling can just grab features from an already calculated pool of features.
But if you have new incoming customers and want to make predictions in realtime you will need to also maintain the code to compute the features at inference time. And at that points, isn't the code itself the feature store? That is, you an reference the same feature code when doing training and inference. If you keep a stored set of features and then derive the features at inference time it means there could version issues with how the features are computed in the featurestore, vs how it's implemented for inference.
It seems in one case featurestore helps, but in the other it introduces points of failure/mismatch. Can someone explain how having a feature store is useful if you have to compute features in realtime on incoming data?
Sources
This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.
Source: Stack Overflow
| Solution | Source |
|---|