Feature Store

Overview

A Feature Store is a centralized platform for storing, managing, and serving ML features. It ensures consistency between training and serving, enables feature reuse across teams, and provides point-in-time correct feature retrieval.

Why Feature Store?

graph LR
    subgraph "Without Feature Store"
        A1[Team A: Feature X] --> D1[Different Implementation]
        B1[Team B: Feature X] --> D2[Different Values]
    end
    
    subgraph "With Feature Store"
        A2[Team A] --> FS[Feature Store]
        B2[Team B] --> FS
        FS --> C[Consistent Features]
    end

Architecture

graph TB
    subgraph "Ingestion"
        B[Batch Sources] --> I[Feature Pipeline]
        S[Streaming Sources] --> I
    end
    
    subgraph "Feature Store"
        I --> O[Offline Store<br/>Training]
        I --> N[Online Store<br/>Serving]
    end
    
    O --> T[Training Pipeline]
    N --> P[Prediction Service]
    
    R[Feature Registry] --> O
    R --> N

Key Concepts

1. Offline Store

• Stores historical feature values
• Used for model training
• Supports point-in-time joins
• Technologies: Hive, BigQuery, Delta Lake

2. Online Store

• Stores latest feature values
• Used for real-time serving
• Low-latency access (ms)
• Technologies: Redis, DynamoDB, Bigtable

3. Point-in-Time Correctness

# Avoid data leakage - only use features available at prediction time
# BAD: Using future data
features_jan = get_features("2024-01-31")  # Includes Feb data!

# GOOD: Point-in-time join
features = feature_store.get_historical_features(
    entity_df=events,
    features=["user_avg_spend_30d"],
    timestamp_column="event_time"
)

Feature Store Comparison

Interview Questions

1. What problem does a feature store solve?
2. Explain the difference between offline and online stores.
3. How do you ensure point-in-time correctness?
4. How would you design a feature store for a recommendation system?
5. What are the trade-offs of using a feature store?

Common Mistakes

• Training-serving skew: Computing features differently in training vs serving
• Data leakage: Using future data in training features
• Over-engineering: Not every project needs a full feature store
• Ignoring freshness: Online store not updated frequently enough

Summary

A Feature Store solves critical ML infrastructure problems: feature reuse, consistency between training and serving, and point-in-time correctness. It bridges the gap between data engineering and ML engineering, enabling teams to share and serve features reliably.

Cross-References

• ML Feature Engineering
• Data Pipeline
• Storage Distributed
• MLOps Pipelines
• System Design Feature Store