AI and the Data Platform: Connecting AI to All Organizational Data

Alexander Polomodov and Nikolay Golov compare four ways to connect AI agents to organizational data. A conventional warehouse provides curated, documented marts but takes time to build and delivers data with a delay. A federated engine queries multiple databases through common SQL; large joins depend on data movement, source performance, and failures in individual branches. A universal database combines workloads within a product, while shared storage separates data from specialized compute engines.

Practical implementations combine these approaches. Microsoft and AWS manage change streams into analytical storage, while lag and failure recovery remain concerns. ClickHouse and Google combine access to multiple engines with replication. Designers still choose tables to copy, calculation locations, and acceptable latency. Oracle and SAP HANA illustrate the promise of combined transactional and analytical processing alongside price and supplier dependence.

Databricks and Snowflake illustrate movement toward connecting analytics and PostgreSQL. The Databricks discussion includes Neon, Mooncake, and a scenario in which an agent creates a service and other agents use its data. Golov stresses that components are still being assembled. Snowflake appears through Crunchy Data and pg_lake. These scenarios indicate development directions; the discussion does not establish that a universal solution is ready for every workload.

Polomodov finds shared storage attractive but asks how to preserve efficiency across different queries. Golov discusses frequent Iceberg writes, growing file counts, and object-storage load. An open format preserves tool choices but does not eliminate operations or maturity checks. A detailed open-source review is left for a future conversation. Timestamps are based on the Podster audio version and may differ slightly from YouTube.