Schema · Lineage · Metadata · Access Control · Feature Groups
Connects both paths
ONLINE PATH - Serving
Streaming
Events
→
Feature
Computation
→
Online Store
(Redis )
→
Model
Serving
→
Prediction
Response
Offline Features (batch)
user_lifetime_clicks
user_category_affinity
item_popularity_30d
user_embedding_128d
Online Features (real-time)
user_session_clicks_1h
real_time_trending_score
user_location_context
Point-in-Time Correct Joins
When building a training dataset, you must only use features that were available at the moment of the prediction. Using future data (label leakage) inflates model performance by up to 30%.
Wrong: Join on latest feature values → future data leaks into training → model overfit
Correct: Join on AS-OF timestamp → only use features that existed at event time
Feature Reuse - Models consuming same features
Two-tower retrieval model
LightGBM ranking model
Both models consume the same offline and online features from the registry - no duplication, no skew.
Feature Store
Use Case
Display
Feast Highlight
Open source, cloud-agnostic. Best for teams wanting full control. No managed infra.
Online: Redis / DynamoDB / Bigtable
Offline: BigQuery / Parquet / Redshift
Latency: < 5ms online
Feature Store Architecture - Interactive Visualization
A feature store manages the full lifecycle of ML features: batch computation for offline training and real-time computation for online serving. The offline path uses Spark or Flink to compute features into Parquet or Delta Lake for model training. The online path computes features from streaming events into Redis or DynamoDB for sub-10ms inference. The feature registry acts as a central catalog ensuring both paths use identical feature definitions - eliminating training-serving skew, which is one of the top root causes of production ML degradation.
Offline path: Spark/Flink computes features from raw events into Parquet - used for model training
Point-in-time correct joins: training dataset must use only features available at prediction time, not future data
Feature reuse: multiple models consume the same feature definitions - no duplication, no drift between model versions
Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.
