The DBOS project, co-founded by Stonebraker, argues operating systems primarily manage data at scale. Replacing core OS components (like the file system and scheduler) with a database engine can lead to faster performance, built-in high availability, and transactional guarantees for system operations, with "really no downside."
Stonebraker clarifies that GPUs excel at parallel processing (SIMD), but database indexing (e.g., traversing a B-tree) is a serial process. Each step involves following a pointer to a new memory location, a sequence of operations that cannot be parallelized effectively, making GPUs unsuitable for accelerating this core database function.
Stonebraker predicts that the next evolution of AI agents will involve performing actions that modify state, such as transferring money. This transforms the problem from simple prediction to a complex distributed systems challenge where atomicity, consistency, and isolation (ACID properties) are critical, making it a classic distributed database problem.
Stonebraker's research reveals that on real production data warehouse benchmarks, LLMs achieve 0% accuracy. This is due to messy, non-mnemonic schemas, complex 100+ line queries, and domain-specific data not found in training sets—factors absent from simplified academic benchmarks like Spider and Bird.
Stonebraker asserts that specialized database architectures (e.g., column stores, stream processors) are an order of magnitude faster for their specific use cases than general-purpose row stores like Postgres. While Postgres is a great "lowest common denominator," at the high end, a tailored solution is necessary for optimal performance.
Ingress, Stonebraker's first database, couldn't handle non-standard data types like polygons for GIS or custom calendars for financial bonds. Postgres was engineered with an extendable type system to solve this fundamental limitation, making it vastly more flexible for diverse applications beyond standard business data processing.
According to Stonebraker, Oracle's early success against the technically superior Ingress was due to Larry Ellison's aggressive sales tactics. Ellison would claim features existed when they didn't, shipping buggy software and essentially using initial customers as a QA team. This highlights how market success isn't always tied to technical merit.
Stonebraker claims the tech world blindly followed Google's lead on MapReduce, which was "ridiculously inefficient" compared to distributed databases. He also slams eventual consistency for failing to guarantee data integrity (e.g., preventing stock from going below zero), a tradeoff most enterprises cannot make. Google later abandoned both concepts.