Databricks Machine Learning Professional Question 137
Select 3A data team is building a recommendation system for a retail website. To ensure users experience low latency when receiving recommendations, the team decides to query precomputed batch predictions stored in a Delta table instead of performing online inference. What are the benefits of this approach?
- A
Reduced latency for serving predictions to users
- B
Lower compute costs compared to real-time model inference
- C
Improved ability to handle sudden spikes in user traffic
- D
Increased accuracy of predictions compared to online inference
- E
Simplified model retraining process
Show answer and explanation
Correct answers: A, B, C
Explanation
Querying precomputed batch predictions offers several advantages, including reduced latency, lower compute costs, and better scalability during traffic spikes. These benefits make it a practical choice for live-serving scenarios where real-time inference is not necessary. However, the accuracy of predictions and the retraining process are independent of whether predictions are served in real-time or precomputed.
- A. Correct.
Querying precomputed batch predictions eliminates the need for real-time inference, significantly reducing latency when serving predictions.
- B. Correct.
Precomputing predictions in a batch process reduces the need for high-performance compute resources at inference time, lowering overall costs.
- C. Correct.
Since predictions are precomputed and stored, the system can handle traffic spikes more effectively without requiring additional real-time compute resources.
- D. Incorrect.
The accuracy of predictions is determined by the model and data used during training, not by whether predictions are computed in real time or in a batch. Therefore, this option is incorrect.
- E. Incorrect.
Batch prediction storage does not inherently simplify the model retraining process; retraining still depends on the pipeline setup and data preparation. This option is not relevant in this context.