When using the Databricks Feature Store for real-time model serving, what component is used to retrieve pre-computed features at low latency?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Delta tables are great for batch, but online serving needs millisecond lookups. The Feature Store can publish to an online store (like DynamoDB) — a fast key-value store that returns features in single-digit milliseconds.
Full explanation below image
Full Explanation
For real-time model serving, Delta Lake/Spark queries have too much latency (seconds). The Databricks Feature Store supports publishing feature tables to online stores: fs.publish_table(name='catalog.ml.user_features', online_store=DynamoDBSpec(region='us-east-1', table_name='user_features_online'), mode='merge'). Supported online stores: AWS DynamoDB, Azure Cosmos DB (via Databricks connectors). At serving time, the model endpoint retrieves features for the incoming entity (e.g., user_id) from the online store, achieving P99 latencies of 5-20ms. The offline-to-online sync can be scheduled as a Databricks job to keep online features fresh. Delta Lake queries (seconds), MLflow artifact store (for model files, not features), and broadcast DataFrames (cluster-memory only, not accessible to serving endpoints) are all unsuitable for real-time feature retrieval.