SnowPro Specialty: Gen AI Question 265
Single answerUsage quotasA data platform team has enabled Snowflake Cortex AI features for several internal applications. After an unexpected spike in LLM-related charges, the team wants a control that helps prevent runaway consumption for a specific workload while still allowing normal usage until a defined threshold is reached. Which approach best addresses this requirement using Snowflake usage quotas?
- A
Create a usage quota for the relevant AI service or model-backed workload so Snowflake can enforce a defined consumption limit for that usage category.
- B
Create a resource monitor on the virtual warehouse that submits Cortex AI requests, because resource monitors directly cap AI token usage across model invocations.
- C
Set a statement timeout on the application role, because timed-out statements prevent any additional model billing once the timeout is reached.
- D
Restrict the application to a smaller warehouse size, because Cortex AI usage is billed only through warehouse credits and not through service-specific consumption.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use a Snowflake usage quota for the applicable usage category. In real environments, teams often need a guardrail that allows business-as-usual usage but stops further consumption after a defined threshold to avoid surprise spend. Usage quotas are designed for this purpose. By contrast, resource monitors govern warehouse credits, not all Snowflake-managed service usage. Similarly, statement timeouts and warehouse resizing are operational controls, not quota-based spending controls for AI services. This aligns with Snowflake guidance to use the governance mechanism that matches the billing surface being controlled: resource monitors for warehouses and usage quotas for supported service usage categories, including AI-related workloads where applicable.
- A. Correct.
Correct. Usage quotas are intended to limit consumption for supported Snowflake services and are the appropriate control when the goal is to put a defined cap on a particular category of usage, such as AI-related consumption. This directly addresses the requirement to allow normal activity up to a threshold and then prevent further runaway usage in that governed category.
- B. Incorrect.
Incorrect. Resource monitors are designed to track and control warehouse credit consumption, not service-specific AI or model token usage. A common misconception is that all Snowflake costs flow through warehouses. In practice, Cortex and other managed services can have their own billing dimensions, so a warehouse resource monitor does not provide the same protection for AI usage quotas.
- C. Incorrect.
Incorrect. Statement timeouts can stop long-running queries, but they do not function as a consumption quota mechanism for AI services. This option confuses query execution governance with cost-governance controls. It may reduce some waste in specific cases, but it does not enforce a cumulative usage threshold for the workload.
- D. Incorrect.
Incorrect. Reducing warehouse size may affect SQL execution performance and some compute costs, but it does not create a hard limit on supported AI service consumption. This distractor reflects the mistaken assumption that Cortex AI charges are controlled solely by warehouse sizing rather than by service-level billing and governance mechanisms.