Databricks-GenAI-Assoc Design Applications Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Configure the Model Serving endpoint with autoscaling enabled and set a minimum and maximum number of concurrent requests per replica.

  Why this is correct

  Autoscaling allows the endpoint to add or remove replicas based on load, which helps maintain consistent latency during peaks. Setting concurrency bounds per replica controls how many requests each replica handles, preventing overload. Together they balance throughput and response time. This is a supported configuration for Databricks Model Serving and directly addresses the reliability and scaling requirements.

• ✗

  Deploy the model to a single large GPU replica with maximum memory to handle all traffic without autoscaling.

  Why it's wrong here

  A single large replica creates a single point of failure and cannot scale horizontally. If traffic exceeds its capacity, latency spikes and requests fail. Autoscaling is preferred for handling variable loads. While a large GPU may handle more concurrent requests, it does not provide the elasticity or fault tolerance needed for peak loads. This design is not recommended for production RAG applications.

• ✓

  Implement client-side retries with exponential backoff and a timeout for each request to the serving endpoint.

  Why this is correct

  Retries with exponential backoff help the application recover from transient errors or throttling, improving reliability. A request timeout prevents hanging calls and allows the client to fail fast or retry. This pattern is essential when calling a model serving endpoint under load. It complements server-side autoscaling by handling temporary capacity issues gracefully, ensuring consistent user experience.

• ✗

  Precompute all possible user questions and their answers and store them in a lookup table for instant retrieval.

  Why it's wrong here

  Precomputing all possible questions is infeasible for open-ended natural language and would not cover new or paraphrased queries. It also does not scale with a growing knowledge base. This approach does not address peak load handling or consistent response times for real user queries. It is not a practical design pattern for generative AI applications.

• ✗

  Cache the LLM responses for identical user queries in a Delta table and serve them without invoking the model.

  Why it's wrong here

  Caching identical queries can reduce load, but it does not address scaling for diverse queries or peak loads. It also risks serving stale answers and is not a substitute for proper endpoint scaling. While caching is useful, it does not ensure consistent response times under varying load. It is not one of the two primary design choices for handling peak loads gracefully.