B4 - feature retrieval specifications

• training-serving skew is possible if the training and serving feature logic differ

• eg: training uses 24 hour mean and age in years, but production scoring uses 12 hour mean and age in months

• here, the model learned $f(X_{\text{traiing}})$, where $X$ has a particular meaning, but $X_{\text{prod}}\ne X_{\text{expected}}$

• it might be something like the same measurement, but in different units

• a feature store helps manage reusable ML features so that teams don't independently recreate the same feature logic everywhere

RAW TELEMETRY
      ↓
feature transformations
      ↓
FEATURE STORE
│
├── temperature_24h_mean
├── voltage_7d_std
├── detector_age_days
└── noise_6h_max
      │
      ├────────► TRAINING
      │
      └────────► INFERENCE

• this helps with consistent feature definitions across training and serving

• feature retrieval system informs downstream deployment of the features required and how to retrieve them

• a signature simply states the shape/schema

• it should be packaged with the model so the lineage includes the feature dependency