Bitzu

Machine Learning Associate ยท Feature tables in Unity Catalog

Online vs offline feature tables

Differences between offline feature tables (Delta, batch training and scoring) and online tables/stores (low-latency lookup for real-time serving) and when each is used.

Key points

  • Offline feature tables: Unity Catalog Delta tables with full history, source of truth for training and batch scoring.
  • Online store (Databricks Online Feature Store, backed by Lakebase): published copy for low-latency lookup by real-time apps and Model Serving endpoints.
  • Without a time series key: only the latest value per entity.
  • Publish with fe.publish_table; unused online stores bill continuously, so delete them.

Source: https://docs.databricks.com/aws/en/machine-learning/feature-store/online-tables

Practice questions

Single choice

What is an offline feature table in Unity Catalog?

  1. A Delta table used for training and batch scoring
  2. A key-value cache that answers in milliseconds
  3. A registered model
  4. A streaming checkpoint
Show answer

Answer: A. A Delta table used for training and batch scoring

Offline tables are Delta tables used for batch training and scoring.

Select all that apply

Which statements about publishing to an online store are correct?

  1. fe.publish_table copies an offline table to the online store
  2. The source table needs a primary key without NULLs
  3. Unused online stores should be deleted because they are billed continuously
  4. The online store replaces the Delta table as the source for training
  5. DROP TABLE fully removes the online table's backing storage
Show answer

Answer: A. fe.publish_table copies an offline table to the online store, B. The source table needs a primary key without NULLs, C. Unused online stores should be deleted because they are billed continuously

Publishing uses publish_table, needs a non-null primary key, and the store bills continuously. DROP TABLE does not remove the backing storage, and the Delta table remains the training source.

True or false

An online feature table without a time series key keeps the full history of every entity's feature values.

Show answer

Answer: False

False. Without a time series key only the latest value per entity is available; full history lives in the offline table.

Practice 8 questions on this topic in Bitzu

Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.

Download on the App StoreGet it on Google Play

More Machine Learning Associate topics in this module