SnowPro Specialty: Gen AI Question 78
Single answerCross-region inferenceA global retail company uses Snowflake Cortex AI functions in an AWS us-east-1 account to summarize customer support tickets. During peak periods, requests to an LLM intermittently fail because capacity for the model in the local region is constrained. The company wants to improve availability without changing application logic to call a different region directly, while still staying within Snowflake-managed capabilities. Which approach should the architect recommend?
- A
Enable cross-region inference so Snowflake can route inference requests to another supported region when local model capacity is unavailable
- B
Create a database replication group and fail over the application database to another region before each inference call
- C
Resize the virtual warehouse that submits the Cortex AI function calls so the LLM has more inference capacity in the local region
- D
Configure Snowpipe Streaming to buffer prompts and automatically replay them in a different region when the model returns a capacity error
Show answer and explanation
Correct answer: A
Explanation
The key requirement is to improve inference availability during regional capacity constraints without rewriting the application to explicitly target another region. Cross-region inference addresses this by allowing Snowflake to route supported model inference requests to another supported region when needed. This is different from Snowflake database replication/failover, which protects data availability, and different from warehouse scaling, which affects SQL compute but not external model-serving capacity. In practice, architects should verify that the required model and region combination supports cross-region inference and review any governance, residency, and compliance implications before enabling it. This aligns with Snowflake guidance that cross-region inference is a managed inference-availability capability rather than a customer-built regional failover pattern.
- A. Correct.
Correct. Cross-region inference is designed for this type of scenario: when model capacity in the local region is constrained, Snowflake can use supported inference capacity in another region without requiring the application to manually target a different region. This improves availability for supported model inference workloads while remaining within Snowflake-managed capabilities.
- B. Incorrect.
Incorrect. Database replication and failover are data availability features, not a mechanism for routing individual Cortex model inference requests to alternate regional model capacity. Failing over a database before each AI call would be operationally impractical and does not address how model inference capacity is allocated.
- C. Incorrect.
Incorrect. Virtual warehouses provide compute for SQL execution and orchestration around the request, but increasing warehouse size does not create additional third-party model inference capacity in a region. A common misconception is to treat LLM availability like warehouse scaling; however, model serving capacity is separate from warehouse sizing.
- D. Incorrect.
Incorrect. Snowpipe Streaming is for ingesting streaming data into Snowflake, not for handling LLM inference retries or rerouting prompts across regions. It does not provide cross-region model execution behavior for Cortex AI functions.