The Open Source Enterprise Feature Platform for Machine Learning

Declarative Feature Engineering

Data scientists can define feature transformations in familiar Python. The platform automatically compiles these definitions into efficient data processing jobs and orchestrates their execution, eliminating the need to manually write and maintain ETL scripts.

@batch_feature_view(
  sources=[transactions],
  entities=[user],
  mode='spark_sql',
  ttl='30d',
  schema=[Field('user_id', Int32), Field('total_spend', Float32)]
)
def user_total_spend(transactions_df):
    return f"""
        SELECT
            user_id,
            SUM(amount) as total_spend,
            CURRENT_TIMESTAMP() as timestamp
        FROM {transactions_df}
        GROUP BY user_id
    """
                                    

Simple Python definitions replace hundreds of lines of complex ETL code.

High-Performance Online Serving

Delivers a high-concurrency feature serving API with millisecond latency to meet the strict performance demands of real-time applications like fraud detection. The architecture is production-hardened, highly available, and auto-scaling.

One-Click Training Data Generation

Accurately generates training datasets from any point in history, guaranteeing no data leakage. It leverages materialized features in the offline store to rapidly build terabyte-scale datasets, dramatically reducing model iteration cycles.