Overview
Use this skill when the platform must support model training and inference safely. It helps agents design feature generation, point-in-time correctness, serving parity, and operational contracts for ML-focused data products.
When to Use
- building training datasets
- designing feature stores or reusable features
- supporting online and offline feature access
- preventing leakage and training-serving mismatch
- publishing model-ready data products
Do not treat feature pipelines as ordinary marts with different names. ML pipelines have different correctness risks.
Workflow
-
Define the feature contract. Include:
- entity key
- feature meaning
- update cadence
- online or offline use
- freshness expectation
-
Protect point-in-time correctness. Training data must only include information available at prediction time.
-
Align offline and online logic. Reuse definitions and validation wherever possible to prevent training-serving drift.
-
Define feature lifecycle and ownership. Clarify:
- producer
- consumers
- deprecation path
- quality monitoring
-
Validate operational behavior. Models break when stale or missing features silently propagate.
Common Rationalizations
| Rationalization | Reality |
|---|---|
| "We can use the latest value for training." | That often introduces leakage and overstates model performance. |
| "Online parity is a model-team problem." | Feature consistency is a data pipeline responsibility too. |
| "Features are internal, so contracts are unnecessary." | Unclear feature meaning leads to misuse and drift. |
Red Flags
- no point-in-time logic is defined
- offline and online definitions diverge
- stale features are not monitored
- feature ownership is unclear
Verification
- Feature meaning, keys, and freshness are documented
- Point-in-time correctness is protected
- Offline and online parity expectations are explicit
- Monitoring exists for stale, missing, or drifting features