SnowPro Specialty: Gen AI Question 77
Single answerCross-region inferenceA global retail company runs a customer-support application on Snowflake in AWS us-west-2. During peak seasonal traffic, calls to Cortex AISQL and COMPLETE functions intermittently fail because the preferred hosted model is temporarily unavailable in the local region. The company wants to improve inference resiliency without redesigning the application, while keeping model invocation inside Snowflake-managed capabilities. Which approach best addresses this requirement?
- A
Enable cross-region inference so Snowflake can route supported model inference requests to another available region when needed
- B
Create a cross-cloud replication group and fail over the entire database to another region before each model call
- C
Increase the size of the virtual warehouse that runs the SQL query so the hosted model has more compute available in the local region
- D
Export prompts to an external LLM endpoint in another region and call it through a custom application service outside Snowflake
Show answer and explanation
Correct answer: A
Explanation
The best answer is to enable cross-region inference for supported Snowflake-hosted model inference workloads. In Snowflake Cortex, cross-region inference is intended to improve resiliency/availability by allowing inference requests to be processed outside the local region when the model is not available there. This is a practical solution when teams want minimal application changes and want to continue using Snowflake-managed AI functions such as AISQL/COMPLETE against supported hosted models. By contrast, warehouse scaling does not solve hosted model regional availability, and database replication/failover is the wrong tool for per-request inference resiliency. External endpoints may be valid in other architectures, but they do not satisfy the requirement to remain within Snowflake-managed capabilities. Candidates should distinguish data-platform DR features from AI inference routing features and understand that cross-region inference applies only where Snowflake documents support for the relevant functions/models.
- A. Correct.
Correct. Cross-region inference is designed to improve availability/resiliency for supported Snowflake-hosted model inference by allowing requests to be served from another region when necessary. This fits the scenario because the company wants to reduce failures during regional model unavailability without changing the application architecture or leaving Snowflake-managed inference paths.
- B. Incorrect.
Incorrect. Database replication and failover address data availability and disaster recovery, not model-serving availability for a single inference request. Failing over an entire database stack before model calls would be operationally heavy and does not reflect how cross-region inference is intended to solve transient hosted-model availability issues.
- C. Incorrect.
Incorrect. Virtual warehouses execute SQL and can affect query processing capacity, but they do not add GPU/model-serving capacity to Snowflake-hosted LLM endpoints in a region. A common misconception is that larger warehouses improve all aspects of AI inference, when in reality hosted model availability is managed separately from warehouse sizing.
- D. Incorrect.
Incorrect. This might provide an alternate inference path, but it violates the stated requirement to keep invocation inside Snowflake-managed capabilities and avoid redesigning the application. It also introduces additional security, networking, and operational complexity that cross-region inference is specifically meant to avoid for supported Snowflake-hosted inference.