Databricks Generative AI Engineer Associate Question 180
Select 3You are deploying a generative AI application on Databricks that serves a fine-tuned large language model (LLM) for real-time text generation. The application requires high availability for end-users and should scale automatically based on workload. Which steps should you take to ensure the application meets these requirements?
- A
Configure a Databricks Serving endpoint with autoscaling enabled.
- B
Deploy the model using a notebook and manually restart the cluster during high traffic.
- C
Enable request batching for the endpoint to optimize resource utilization.
- D
Use the Databricks REST API to monitor endpoint performance and adjust configurations dynamically.
- E
Deploy the application on a single-node cluster to reduce latency.
Show answer and explanation
Correct answers: A, C, D
Explanation
To ensure high availability and scalability for a generative AI application in Databricks, you should leverage Databricks Serving endpoints with autoscaling, enable features like request batching to optimize resource usage, and utilize APIs for dynamic monitoring and adjustments. These practices ensure the application meets performance requirements under variable workloads, while manual interventions or single-node setups are inadequate for production-grade deployments.
- A. Correct.
Configuring a Databricks Serving endpoint with autoscaling enabled ensures that the application can handle fluctuating traffic while maintaining availability and performance.
- B. Incorrect.
Manually restarting the cluster during high traffic is not a scalable or reliable solution for ensuring high availability.
- C. Correct.
Enabling request batching improves resource utilization and reduces overhead by combining multiple requests into a single operation.
- D. Correct.
Using the Databricks REST API allows for dynamic monitoring and configuration adjustments, ensuring optimal performance and availability.
- E. Incorrect.
Deploying on a single-node cluster may reduce latency in some scenarios, but it does not provide high availability or scalability, as the single node becomes a single point of failure.