Feature Groups

Feature groups organize features for machine learning with point-in-time correctness.

---

Overview

Feature groups are collections of features for a specific entity (e.g., customer, product). They enable feature reuse and prevent data leakage through point-in-time joins.

---

Feature Group Concepts

Entities

Entities define the join key for features:

from seeknal.entity import Entity

customer_entity = Entity(
    name="customer",
    join_keys=["customer_id"]
)

Materialization

Control when and how features are computed:

from seeknal.featurestore.duckdbengine.feature_group import Materialization

materialization = Materialization(
    event_time_col="transaction_time",
    lookback_days=30
)

---

Creating Feature Groups

DuckDB Feature Groups

from seeknal.featurestore.duckdbengine.feature_group import FeatureGroupDuckDB

fg = FeatureGroupDuckDB(
    name="customer_features",
    entity=customer_entity,
    materialization=materialization,
    project="my_project"
)

# Auto-detect features from DataFrame
fg.set_dataframe(df).set_features()

# Write to offline store
fg.write(feature_start_time=datetime(2024, 1, 1))

Spark Feature Groups

from seeknal.featurestore.feature_group import FeatureGroup

fg = FeatureGroup(
    name="customer_features",
    entity=customer_entity,
    project="my_project"
)

# Set features explicitly
fg.set_features({
    "total_purchases": "sum(purchase_amount)",
    "transaction_count": "count(*)"
})

---

Point-in-Time Joins

Prevent data leakage by joining features as they appeared at prediction time. Use FeatureFrame.pit_join() in transforms:

from seeknal.pipeline import transform

@transform(name="training_data")
def training_data(ctx):
    # Spine with entity keys and prediction dates
    labels = ctx.ref("source.churn_labels")

    # PIT join: get features as of each application_date
    features_df = ctx.ref("feature_group.customer_features").pit_join(
        spine=labels,
        date_col="application_date",
        keep_cols=["label"],
    )
    return features_df

---

Online Serving

Retrieve features for low-latency serving using ctx.features():

from seeknal.pipeline import transform

@transform(name="predictions")
def predictions(ctx):
    # Get latest features from entity store
    features = ctx.features("customer", [
        "customer_features.total_purchases",
        "customer_features.avg_order_value",
    ])
    return features

---

Feature Group Configuration

Common Options

Option	Type	Description	Default
name	string	Unique identifier	Required
entity	Entity	Entity definition	Required
materialization	Materialization	Materialization config	Optional
project	string	Project name	Required

---

Best Practices

1. Always use point-in-time joins for ML features
2. Define clear entities with join keys
3. Version feature groups for production
4. Document feature logic clearly
5. Test feature serving before deployment

---

Related Topics

• Point-in-Time Joins - Preventing data leakage
• Training to Serving - End-to-end ML workflow
• ML Engineer Path - Complete ML tutorial

---

Next: Learn about Semantic Models or return to Building Blocks