Interview, Fireside Chat
Benoit Dageville, Snowflake ｜ RAISE Summit 2025
- Snowflake Strategic Shift to Unstructured Data
- The platform has evolved from integrating structured and semi-structured data to making unstructured data a "first-class citizen."
- Key capability added: AI SQL, elevating SQL to query unstructured data and optimize GPU utilization to prevent idle cycles.
- Current product focus includes RAG (Retrieval-Augmented Generation), search, and vectorization execution pipelines.
- Semantic Views Launch and Capabilities
- Snowflake has officially shipped "Semantic Views" to push business logic and metric definitions directly into the data platform.
- Function: Defines metrics (e.g., "revenue"), aggregation rules, and table relationships so AI can interpret data without human intervention.
- Impact: Reduces the complexity of AI interactions by providing a "level one" semantic layer above the raw data.
- Performance: Benoit reports AI quality for semantic queries has surpassed 95%, with a roadmap goal of 100%.
- Roadmap Priorities:
- Expanding connectors for unstructured data (e.g., SharePoint) while preserving access control lists (ACLs).
- Maintaining a three-layer architecture: Data, Semantic, and Governance.
- Graph Data and Architecture Philosophy
- Graph databases are viewed as the underlying representation of all relationships (including standard relational joins).
- Strategic Choice: Relationships can be materialized for performance or defined within the semantic model as metadata.
- Core Philosophy: The platform prioritizes simplicity by absorbing complexity internally, rather than pushing it to the user.
- Governance and Open Source Strategy
- Snowflake employs a dual approach to data cataloging:
- Horizon: The proprietary, full-featured internal catalog covering all assets, including compute.
- Polaris: The open-source protocol intended to standardize industry-wide cataloging.
- Competitive Distinction: Snowflake differentiates itself from competitors (e.g., Databricks' Unity Catalog) by strictly separating their open-source protocol from their proprietary compute layer to avoid "deception" in naming.
- Security Focus: Governance layer ensures strict access control when ingesting unstructured documents, preserving security policies.
- Snowflake employs a dual approach to data cataloging:
- Market Positioning and Benchmarks
- Performance Claims: Benoit disputes competitor claims of being "faster and less expensive," citing TPCDS benchmarks as the standard for objective comparison.
- Metric Definition: Distinguishes between "efficiency" (infrastructure utilization) and "cost," asserting that high performance is a prerequisite for cost efficiency.
- Gen AI vs. Analytics: Clarifies that Gen AI hype previously confused operational analytics with generative magic; the current focus is on the intersection where AI extracts insights from unstructured data.
- Future Roadmap and Agent Architecture
- Three-Phase AI Integration:
- Ingestion: Using AI to understand and process unstructured data.
- Interaction: Translating natural language to SQL or search queries.
- Agentic Layer: Creating focused AI agents and integrating them to perform complex, multi-step tasks.
- Democratization Goal: Aiming to allow any business user to interact directly with data via natural language, moving beyond "geek-only" technical access.
- Unstructured Data Revolution: The integration of unstructured data is described as a foundational revolution in the data platform sector.
- Three-Phase AI Integration: