Agentic Solution for Data Warehouse Access by Can Lin & Uday Ramesh Savagaonkar

Can Lin and Uday Ramesh Savagaonkar from Meta discuss Meta's approach to data warehouse access management in the age of AI agents, focusing on how they're building agentic solutions to streamline access and minimize security risks. They explain the complexities of managing data access at Meta's scale, where numerous teams and engineers work with sensitive user data. They introduce a multi-agent system with data user agents and data owner agents to negotiate data access, supported by a hierarchical data warehouse structure, context management, and intention management. They detail an end-to-end use case involving data discovery, exploration, and use, highlighting the importance of context-driven decisions, fine-grained query-level provisioning, data access budgets, and rule-based safe learning engines. They share performance metrics from a partial data simulation, showing a high recall rate and a conservative approach tuned towards safety, and discuss future work involving agentic collaboration, infrastructure evolution, and continuous improvement through feedback loops.