Ivx Cf Feature Store

Feature Store

When to Use

• Sharing features across multiple models
• Ensuring training/serving consistency
• Real-time and batch feature serving
• Feature versioning and lineage

Architecture

Raw Data ──► Feature Engineering ──► Feature Store ──► Training / Serving
                      │
                      ▼
                Feature Registry

Tools

Feast Pattern

Define Features

from feast import Entity, Feature, FeatureView, ValueType
from feast.types import Float, Int64

# Entity
video = Entity(
    name="video_id",
    value_type=ValueType.STRING,
    description="Video ID",
)

# Feature View
video_features = FeatureView(
    name="video_metadata",
    entities=["video_id"],
    ttl=timedelta(days=1),
    features=[
        Feature(name="duration", dtype=Int64),
        Feature(name="resolution", dtype=Int64),
        Feature(name="frame_count", dtype=Int64),
    ],
    batch_source=video_source,
)

Materialize

from feast import FeatureStore

store = FeatureStore(repo_path=".")

# Materialize latest features (batch)
store.materialize(
    start_date=datetime(2024, 7, 1),
    end_date=datetime.now(),
)

# Get features for training
training_df = store.get_historical_features(
    entity_df=pd.DataFrame({"video_id": video_ids, "event_timestamp": timestamps}),
    features=["video_metadata:duration", "video_metadata:resolution"],
).to_df()

Online Serving

# Get current features for prediction (low latency)
features = store.get_online_features(
    features=["video_metadata:duration", "video_metadata:resolution"],
    entity_rows=[{"video_id": "abc123"}],
).to_dict()
# 5ms latency, used in production

Feature Engineering

Online vs Offline

Consistency

• Same transformation code for training and serving
• Unit tests for feature transforms
• Shadow serving to validate

CF-Specific

• Store video metadata features (duration, resolution, tags)
• User preference features for personalization
• Content embedding features (text, image, audio)
• Pipeline performance features (latency, cost)
• Version features with model versions

Anti-Patterns

• ❌ Training and serving code paths different
• ❌ No feature versioning
• ❌ Storing PII in feature store
• ❌ No TTL on features
• ❌ No monitoring for drift