Feature Store

Feast-based feature store with offline (Parquet) and online (Redis) stores. Single source of truth for all vehicle features used in training and serving.

Stores

graph LR
    MP[Materialization Pipeline] -->|fleet vehicles| PQ
    MP -->|new arrivals| RD
    FM[FeatureMaterializer] -->|on vehicle save| RD

    PQ[(Offline Store<br/><i>Parquet — fleet only</i>)]
    RD[(Online Store<br/><i>Redis — new arrivals only</i>)]

    PQ -->|read| TRAIN[Training]
    PQ -->|read| OFR[OfflineFeatureReader<br/>fleet display]
    RD -->|read| FL[FeatureLookup<br/>real-time inference]

Offline store (Parquet) — contains fleet vehicles only (those with an observed num_reservations, including 0). Populated by the materialization pipeline (daily via Airflow). Read by the training pipeline (as training data) and by the serving layer's OfflineFeatureReader for fleet vehicle display.

Online store (Redis) — contains only new arrivals (vehicles with no observed reservations). Populated two ways: - Real-time: FeatureMaterializer Ray actor computes features and writes to Redis on vehicle save (via Redis pub/sub) - Batch backfill: materialization pipeline writes new arrivals via store.write_to_online_store() on each daily run

Read by FeatureLookup for real-time inference on new arrivals.

Feature View

Property	Value
Name	vehicle_features
Entity	vehicle (key: vehicle_id)
TTL	365 days
Source	FileSource (Parquet)

Feature Schema

5 model features — raw prices are vehicle attributes used to compute price_diff but are not model inputs (they are collinear with the derived feature).

Feature	Type	Source	Description
technology	Int64	Raw	Instant-bookable tech package (0/1)
num_images	Int64	Raw	Number of listing photos (1–5)
street_parked	Int64	Raw	Street parked flag (0/1)
description	Int64	Raw	Character count of listing description
price_diff	Float64	Derived	actual_price - recommended_price

Label (not a model input):

Field	Type	Description
num_reservations	Int64 (nullable)	Observed reservation count. NULL for new arrivals (no history yet), 0 or more for fleet vehicles.

Feast Configuration

# feature_repo/feature_store.yaml
project: vroom_forecast
provider: local
registry: ${FEAST_REGISTRY}       # feast-data/registry.db
online_store:
  type: redis
  connection_string: ${FEAST_REDIS}  # localhost:6379
offline_store:
  type: file

Key Files

• feature_repo/feature_store.yaml — Feast config (offline: file, online: Redis)
• feature_repo/definitions.py — Entity, FileSource, FeatureView, feature refs
• seed.py — Loads CSVs into SQLite (idempotent, used by the materialization pipeline)
• pipeline.py — Computes features and writes to stores (called by Airflow)

Database Schema

The SQLite database is the mutable source of truth for vehicle data. It is populated by seed.py (CSV vehicles) and the serving API (UI vehicles).

erDiagram
    vehicles {
        int vehicle_id PK
        int technology
        float actual_price
        float recommended_price
        int num_images
        int street_parked
        int description
        text source
    }
    reservations {
        int id PK
        int vehicle_id FK
        text created_at
    }
    vehicles ||--o{ reservations : has