Design a low-latency feature serving architecture that avoids training/serving skew

Compute features once in a shared pipeline that writes to both an online store for serving and an offline store for training, use point-in-time correct joins to build training sets, and keep any request-time computation in a library shared by both paths.

Skew arises whenever a feature's value at training time differs from what serving produces for the same entity and moment. The architecture must make the two paths structurally identical, not merely intended to match.

Feature taxonomy — the design falls out of this classification:

3.…