Prevent accidental differences between training and serving features#
A training job may compute “purchases in the previous 30 days” differently from the online service. A feature-store architecture can centralize definitions and provide historical, point-in-time-correct training retrieval plus low-latency online values.
| Offline access | Online access |
|---|---|
| Historical scans and training datasets | Low-latency current lookup |
| Point-in-time correctness prevents future leakage | Freshness and availability affect predictions |
This adds pipelines, materialization, ownership, freshness monitoring, and another store. A small ML system with simple features may not need it.
Related practice problems
Use the manifest-backed prompts below to apply this lesson in a complete design.
- Feature store introductionFeast Documentation
Reach this point to mark the lesson as read.