SnowPro Specialty: Gen AI Question 247
Single answer3.3 Monitor and optimize Snowflake Cortex costs.A company has launched a customer-support assistant built on Snowflake Cortex. Over the last two weeks, finance has reported a sharp increase in AI-related spend. The engineering team suspects that some prompts are much larger than necessary and that one application is using a more expensive model than intended. The team wants to identify which users, applications, and models are driving the cost increase so they can take corrective action. Which approach would BEST help them monitor and optimize Snowflake Cortex costs?
- A
Query Snowflake account usage views for Cortex usage to analyze consumption by user, query, and model, then use that data to find oversized prompts and unexpected model selection patterns
- B
Increase the size of the virtual warehouse that serves the application so Cortex requests finish faster and therefore cost less overall
- C
Replicate the Cortex-related schemas to another region and compare storage costs there, because Cortex charges are primarily determined by replicated metadata size
- D
Disable query history retention for the application users so that prompt text is not stored, which directly reduces Cortex model inference charges
Show answer and explanation
Correct answer: A
Explanation
To monitor and optimize Snowflake Cortex costs, teams should start with usage observability and attribution. The scenario specifically calls for identifying cost drivers by user, application behavior, and model choice. Snowflake best practices for cost management emphasize reviewing account usage and query/AI-related monitoring data to understand consumption patterns before making changes. After identifying the source of spend, common optimization actions include reducing unnecessary prompt/token volume, selecting the lowest-cost model that still meets quality requirements, and preventing accidental routing to more expensive models. Warehouse resizing, replication changes, or query history settings do not address the underlying Cortex inference cost drivers described here.
- A. Correct.
Correct. The most effective first step is to use Snowflake's monitoring and usage data to identify where Cortex spend is coming from. In practice, teams should analyze Snowflake account usage/monitoring data for AI and query activity to attribute usage by user, workload, and model. This helps surface patterns such as unnecessarily large prompts, repeated calls, or use of a higher-cost model than expected. Once those drivers are identified, the team can optimize prompts, routing logic, caching strategy, or model choice.
- B. Incorrect.
Incorrect. Cortex model inference cost is not reduced simply by increasing warehouse size. Warehouses affect compute for SQL and data processing workloads, but Cortex model usage pricing is tied to the AI service usage itself, not to making a warehouse larger. Someone might choose this option because faster completion sometimes reduces infrastructure time in traditional architectures, but that is not the primary lever for Cortex inference cost optimization.
- C. Incorrect.
Incorrect. Cross-region replication and storage metadata are unrelated to the main drivers of Snowflake Cortex inference spend in this scenario. The issue described is prompt size and model choice, which are application usage characteristics. This option confuses storage/replication costs with AI model consumption costs.
- D. Incorrect.
Incorrect. Query history retention settings do not directly reduce Cortex inference charges. Even if a team has governance concerns about sensitive prompt text, disabling or limiting history is not a cost optimization method for model execution itself. This option reflects a common misconception that reducing observability data lowers service consumption charges.