Databricks Machine Learning Associate Question 448
Select 3A machine learning team is building a recommendation system for an e-commerce platform. They plan to use a feature store to manage features used in their models. The team needs to decide whether to use an online or offline feature table for storing user browsing history, which will be used during both training and real-time inference. What are the key considerations when choosing between online and offline feature tables in this scenario?
- A
Online feature tables are optimized for low-latency access during real-time inference.
- B
Offline feature tables are primarily used for batch processing and model training.
- C
Online feature tables are better suited for storing historical data used in training.
- D
Offline feature tables are not suitable for real-time inference due to higher latency.
- E
Online feature tables can store features for both training and inference without performance trade-offs.
Show answer and explanation
Correct answers: A, B, D
Explanation
Online feature tables are optimized for low-latency access, making them suitable for real-time inference. In contrast, offline feature tables are optimized for batch processing and are primarily used for model training. The decision between online and offline feature tables depends on the specific use case: low-latency real-time applications favor online tables, while batch training or historical data analysis favors offline tables. Using offline tables for real-time inference would result in higher latency, and using online tables for training can lead to performance constraints.
- A. Correct.
Correct. Online feature tables are designed for real-time use cases where low-latency access is critical, such as real-time inference.
- B. Correct.
Correct. Offline feature tables are typically used for batch processing tasks like model training, where latency is less of a concern.
- C. Incorrect.
Incorrect. Online feature tables are not ideal for storing large amounts of historical data due to storage and performance constraints.
- D. Correct.
Correct. Offline feature tables are not suitable for real-time inference because they are optimized for high-throughput batch operations, not low-latency access.
- E. Incorrect.
Incorrect. While online feature tables support real-time inference, using the same table for training and inference may lead to performance trade-offs or inconsistent feature values between training and inference.