What freshness measures
Feature freshness is how recently a feature value in the online store was updated relative to the real world it represents. A feature that should reflect the last hour of activity but is six hours old is stale.
Why staleness happens
- The pipeline that writes the online store runs on a schedule, so values lag behind reality between runs.
- A streaming job falls behind and the lag grows.
- A failed materialization job leaves old values in place.
The cost of staleness
- For fast moving features like recent clicks or current price, stale values feed the model the wrong picture and predictions degrade.
- The damage scales with how quickly the underlying signal changes. Slow features like a user country tolerate lag well.
Managing freshness
- Define a freshness budget per feature, for example update within five minutes.
- Emit the age of each served feature as a metric and alert when it exceeds the budget.
- Match the materialization cadence to how fast the feature actually changes.
Key idea
Feature freshness tracks how stale online values are, and fast moving features need tight freshness budgets with age monitoring to keep predictions accurate.