Skip to content

Tutorial 4: Deploy Schema & Feature Store

This tutorial creates the PUDO project schema in Snowflake and registers the feature store entities and feature views.

What you will learn

  • How to deploy a project-specific schema.
  • How entities and feature views are defined and registered.
  • How the feature store connects to the shared data.

Before you start

  • Hub infrastructure is deployed.
  • Shared data is seeded (Tutorial 3).

Step 1: Deploy the project schema

make -C projects/pudo deploy-schema

This creates the project-specific schema (PUDO_DEV or equivalent depending on your environment) in Snowflake. The schema holds:

  • Project-specific tables (inference outputs, evaluation results).
  • Feature store objects.
  • Model registry objects.

Step 2: Deploy the feature store

make -C projects/pudo deploy-feature-store

This registers:

  1. Entities: the primary keys that features are organised around:
  2. PUDO entity, keyed by PUDO location ID.

  3. PUDO_DATE entity, composite key of PUDO ID and date (for temporal features).

  4. Feature views: collections of related features computed from the shared data:

Feature view Features
pudo_geospatial_features Number of nearby competing PUDOs, total nearby capacity.
pudo_historical_features Historical parcel volumes, delivery success rates.
pudo_temporal_features Daily parcel demand, day-of-week patterns.

How feature views work

Each feature view is a Snowflake-managed view that:

  1. Reads from the SHARED_DATA tables.
  2. Computes features using SQL or Snowpark transformations.
  3. Registers the results in the Feature Store with point-in-time correctness.

The feature views are versioned. When you deploy, a new version is created if the feature definitions have changed. Downstream consumers (training, inference) reference specific versions through configuration.

Configuration

Feature store behaviour is controlled by YAML configuration in projects/pudo/config/:

config/
├── feature_view/
│   ├── feature_store/
│   │   ├── base.yaml
│   │   └── dev.override.yaml
│   └── feature_views/
│       ├── base.yaml
│       └── dev.override.yaml

The feature_store/ configures the Feature Store instance (database, schema, versioning). The feature_views/ configures individual feature view parameters (lookback windows, aggregation levels, version pins).

Step 3: Verify the feature store

You can verify the feature store deployment by querying Snowflake:

-- List registered entities
SELECT * FROM PUDO_DEV.FEATURE_STORE.ENTITIES;

-- List feature views and their versions
SELECT * FROM PUDO_DEV.FEATURE_STORE.FEATURE_VIEWS;

What you have now

  • Project schema created in Snowflake.
  • Entities registered (PUDO, PUDO_DATE).
  • Feature views deployed (geospatial, historical, temporal).
  • Feature store ready for dataset generation.

Next step

Continue to Tutorial 5: Deploy & Run Training to train your first model using the feature store data.