Training-Serving Skew
A performance degradation problem in machine learning systems caused by inconsistency between feature definitions in offline training and online serving. It is a common hidden failure source in real-time risk control systems.
Timeline (last 90 days)
A feature store is an intermediate layer between the data pipeline and the model that uniformly manages offline computation, online serving and metadata of features, whose key value lies in eliminating training-serving skew.
The core problem the dual-store pattern solves is training-serving skew.
All Facts (2)
A feature store is an intermediate layer between the data pipeline and the model that uniformly manages offline computation, online serving and metadata of features, whose key value lies in eliminating training-serving skew.
50%UnverifiedThe core problem the dual-store pattern solves is training-serving skew.
50%