Training and inference compute a feature differently, which hurts accuracy. What helps?
Why A is the answer
A shared feature store
A feature store keeps feature definitions consistent, which prevents training-serving skew.
In detail
- A shared feature store gives both the same values.
- This mismatch is called training-serving skew.
Use case
An illustrative scenario, not a real incident.
Training computes average spend over 30 days, and serving uses 7. A shared feature store gives both the same definition.
See it in different ways
flowchart TD A["Feature pipeline"] --> B["SageMaker Feature Store"] B --> C["Offline store: training"] B --> D["Online store: low-latency inference"]
sequenceDiagram participant P as Feature pipeline participant O as Offline store participant N as Online store participant T as Training participant I as Inference P->>O: Historical features P->>N: Latest features O->>T: Training data N->>I: Values in milliseconds
mindmap
root(("Feature store"))
n1["Training"]
n2["Same feature values"]
n3["Inference"]
n4["Same feature values"]Key point: One feature definition prevents skew.
Why the other options are wrong
- B. A longer training job with the same feature code
- More training does not fix a mismatch.
- C. A higher learning rate on the next training run
- A learning rate does not fix a mismatch.
- D. A larger endpoint instance for the production model
- Instance size does not change features.