Advanced

Enhancements

Streaming features, governance, team onboarding, and frequently asked questions.

Enhancement 1: Streaming Features

# Stream features from Kafka using Feast
from feast import StreamFeatureView, KafkaSource

kafka_source = KafkaSource(
    name="driver_trips_stream",
    kafka_bootstrap_servers="localhost:9092",
    topic="driver_trips",
    timestamp_field="event_timestamp",
    message_format={"json": {"schema": "..."}},
)

driver_trips_stream = StreamFeatureView(
    name="driver_trips_stream",
    entities=[driver],
    schema=[
        Field(name="trips_today", dtype=Int64),
        Field(name="total_distance", dtype=Float32),
    ],
    source=kafka_source,
    online=True,
)

Enhancement 2: Feature Governance

# Feature catalog with metadata and ownership
feature_catalog = {
    "driver_stats": {
        "owner": "ml-platform-team",
        "description": "Core driver performance metrics",
        "sla": "99.9% availability, <5ms p99 latency",
        "pii": False,
        "tags": ["driver", "performance"],
        "consumers": ["trip-prediction", "driver-ranking"],
    }
}

# Access control
def check_access(team, feature_view):
    allowed = {
        "ml-team": ["driver_stats", "trip_features"],
        "analytics": ["driver_stats"],
    }
    return feature_view in allowed.get(team, [])

Enhancement 3: Team Onboarding

# Feature discovery CLI
# feast feature-views list
# feast entities list
# feast feature-services list

# SDK usage for new teams
from feast import FeatureStore

store = FeatureStore(repo_path="feature_repo")

# Discover available features
for fv in store.list_feature_views():
    print(f"{fv.name}: {[f.name for f in fv.features]}")
    print(f"  Entity: {fv.entities}")
    print(f"  Tags: {fv.tags}")

Frequently Asked Questions

❓

Q: Feast vs Tecton vs Hopsworks?
Feast is open-source and self-hosted. Tecton is a managed platform with streaming support. Hopsworks focuses on feature pipelines. Feast is best for teams wanting full control.

❓

Q: How often should I materialize?
Depends on feature freshness requirements. Real-time: use streaming. Hourly: cron job. Daily: batch pipeline.

❓

Q: How do I handle feature versioning?
Use feature view versions (e.g., driver_stats_v2) and stage transitions. Keep old versions alive until all consumers migrate.

❓

Q: Can I use this with non-Python models?
Yes. The FastAPI serves features as JSON, so any language can consume them. You can also use gRPC for lower latency.

Congratulations

You have built a complete ML feature platform with feature definitions, offline/online stores, a serving API, and monitoring. These are the same patterns used by leading ML teams at scale.

← PreviousFeature Monitoring