17 min
Lesson 3 of 4
Production Deployment & Governance
From the course Snowflake Cortex AI
Deploying Cortex AI in production requires attention to performance, cost management, and governance.
Performance Optimization
Compute Sizing
- LLM functions use serverless compute (no warehouse needed for basic functions)
- Complex pipelines benefit from dedicated AI warehouses
- Batch processing is more efficient than row-by-row
- Use appropriate model sizes: smaller models for simple tasks
Model Selection Guide
| Task | Recommended Model | Speed | Quality |
|---|---|---|---|
| Sentiment analysis | Built-in function | Fastest | High |
| Text classification | mistral-7b | Fast | Good |
| Summarization | mistral-large | Medium | High |
| Complex reasoning | llama3-70b | Slower | Highest |
Caching
- Cache AI results in tables for repeated queries
- Use materialized views for frequently accessed enrichments
- Incremental processing: only run AI on new/changed records
Cost Management
- LLM functions are priced per token processed
- Monitor usage with Snowflake's cost management tools
- Set resource monitors and alerts
- Use smaller models when quality is sufficient
- Batch process during off-peak hours for lower costs
Governance & Security
Access Control
- Cortex functions respect Snowflake's RBAC
- Control who can call LLM functions with role grants
- Audit AI usage through query history
- Mask sensitive data before AI processing
Data Governance
- AI outputs inherit the governance of input data
- Tag AI-generated columns for lineage tracking
- Document which models were used for compliance
- Implement data quality checks on AI outputs
Streamlit Integration
Build interactive AI apps with Streamlit in Snowflake:
- Create chat interfaces powered by Cortex
- Build dashboards with AI-generated insights
- Deploy internal tools with semantic search
- No external hosting required — runs inside Snowflake
Key Takeaways
- •Select model sizes based on task complexity — smaller models for simple tasks
- •Cache AI results and process incrementally to manage costs effectively
- •Cortex functions respect Snowflake RBAC and audit controls
- •Streamlit in Snowflake enables interactive AI applications without external hosting
0 of 4 lessons marked complete
Progress is kept in this browser only.