Why PostgreSQL’s Creator Criticizes Oracle and Google – AI‑Generated SQL Still Falls Short

1. Entering the Database Field

Stonebraker describes joining Berkeley after graduation, noting that continuing PhD research offered little future prospects, and how mentor Gene Wong guided him toward database research in 1971, the year after Ted Cod’s seminal CACM paper.

2. Competition with Oracle

He contrasts the early Codasyl network model and IBM’s hierarchical IMS with Ted Cod’s relational approach, arguing that Ingres’s implementation of referential integrity was superior to Oracle’s, which merely documented the feature without implementing it.

3. What Makes Postgres Unique

Postgres was built to support GIS data types (points, lines, polygons) that Ingres could not handle, leading to a flexible, extensible type system that allows custom data types with high efficiency.

He recounts a bond‑trading customer who needed a “bond calendar” that differed from standard calendar arithmetic; Ingres’s hard‑coded date logic could not accommodate this, whereas Postgres’s design enables such domain‑specific extensions.

4. “One Architecture Does Not Fit All”

Stonebraker explains his 2004 paper on specialized database architectures, citing Streambase, column‑store designs like Vertica, and arguing that generic databases lose an order of magnitude in performance compared to purpose‑built systems such as ClickHouse or Pinecone.

5. Disagreeing with Google and Amazon

He criticizes Google’s early Hadoop/MapReduce efforts as inefficient and argues that Google’s pursuit of eventual consistency is impractical for most transactional workloads, contrasting it with Spanner’s use of traditional transactions.

Regarding Amazon, he notes the company’s maintenance of fifteen different database systems, suggesting most should be retired in favor of a smaller, better‑performing set.

6. Choosing Academia Over Industry

Stonebraker prefers academia because corporate environments restrict publishing and open research, and he dislikes corporate bureaucracy.

7. Replacing OS State Management with Databases (DBOS)

He describes the DBOS project (started ~2019) that stores all OS‑level state in a Postgres database, enabling high‑performance scheduling, fault‑tolerance, and atomic workflow execution across languages like TypeScript, Java, Go, and Python.

The project secured funding in 2023, emphasizing that as AI agents gain read‑write capabilities, they will heavily depend on robust database features.

8. Future Challenges and AI‑Generated SQL

Stonebraker outlines why current LLM‑to‑SQL systems fall short: training data lacks warehouse schemas, benchmark queries are far simpler than real workloads, and complex, ambiguous schemas hinder accuracy.

He notes that in their own tests, LLMs achieved 0 % accuracy on real warehouse queries, improving only to ~35 % when the FROM clause is supplied.

9. Advice to His Younger Self

He urges young engineers to break out of conventional thinking, pursue what they love, and be cautious about betting on a forever‑growing computer‑science industry.