Hybrid Features: Python + SQL

Fabra v1.1.0 introduces Hybrid Features, allowing you to choose the best tool for the job: Python for complex logic, and SQL for heavy lifting.

The Two Modes

1. Python Features (On-Demand)

Perfect for real-time transformations, complex math, or logic that is hard to express in SQL.

@feature(entity=User)
def name_length(user_id: str) -> int:
    # Complex logic in pure Python
    user = fetch_user_profile(user_id)
    return len(user.name)

How it works:

Online: The function is executed in real-time when you request the feature.
Offline: The function is applied row-by-row to your entity dataframe using apply().

2. SQL Features (Batch / Materialized)

Perfect for aggregations, joins, and large-scale data processing that should happen in the warehouse.

@feature(
    entity=User,
    sql="SELECT user_id, count(*) as txn_count FROM transactions GROUP BY user_id",
    materialize=True,
    refresh=timedelta(minutes=10)
)
def user_txn_count(user_id: str) -> int:
    return 0 # Fallback value or type hint

How it works:

Offline Retrieval: Fabra executes the SQL query against your Offline Store (DuckDB or Postgres) and joins the result to your entity dataframe.
Materialization: The Scheduler executes the SQL query periodically and bulk-loads the results into Redis.
Online Serving: The API serves the pre-computed value from Redis.

[!IMPORTANT] Point-in-Time Correctness: For SQL features, your entity dataframe must have a timestamp column (default: event_timestamp or timestamp) to perform the ASOF join correctly.

Mixing Both

You can mix both types in the same request!

training_df = await store.get_training_data(
    entity_df,
    features=["name_length", "user_txn_count"]
)

Fabra automatically handles the orchestration:

Computes name_length in Python.
Delegates user_txn_count to the database.
Merges everything into a single DataFrame.