Training a 4B model to produce 81% faster query plans than Postgres

TEXT START: Leis et al. asked this exact question in 2015.

The Dissection

This is a proof-of-capability for agentic cognitive automation. A difficult search problem is wrapped in inspection, execution, and measurement tools; SFT and RL then train a model to search the plan space against an objective reward.

The 4B model is not independently replacing Postgres. The operational achievement belongs to the entire pipeline: pg_hint_plan, planner statistics, repeated executions, calibration, container control, benchmark curation, and the measurement harness. The model is the search component inside a machine built to make search tractable.

The headline is also stronger than the supplied evidence. The headline claims 81% faster plans, while the body reports a 44.7% latency reduction across 113 queries. The excerpt does not reconcile those figures. It also says the untrained model initially failed to produce plans for 99 queries. This is a trained, tool-mediated result on a fixed IMDb workload—not raw general-purpose superiority.

The Core Fallacy

The central error is treating execution time as the entire optimization problem and a benchmark win over Postgres’s default plan as durable superiority. Production optimization also includes planning overhead, data drift, changing indexes, hardware, concurrency, memory pressure, tail latency, correctness, maintainability, and operational risk.

The deeper Discontinuity Thesis error is confusing capability with durable economic position. Once this agentic loop is standardized, the 4B model becomes a commodity. The value migrates to whoever controls the workload traces, database runtime, evaluation harness, compute, and deployment surface. Human optimization judgment is not protected; it is being converted into training data and executable feedback.

Hidden Assumptions

• IMDb and 113 join-heavy queries represent enough of production workload diversity.
• Disjoint JOB and CEB topologies eliminate meaningful overfitting.
• The cost of exploratory executions, training, and inference can be amortized over enough repeated queries.
• A discovered plan remains superior as data distributions, indexes, hardware, PostgreSQL versions, and concurrency change.
• Hint-driven plans remain safe and maintainable rather than becoming brittle overrides.
• The reported speedup uses a comparable baseline and includes all relevant overhead; the excerpt does not establish this.
• The result reflects transferable optimization skill rather than learned database regularities plus a powerful harness.
• Humans remain necessary at the control point. The architecture suggests the opposite: humans build the loop, then the loop performs the search.

Social Function

Classification: partial truth, transition management, and prestige signaling.

The technical result is real within its stated boundary. But the presentation packages a labor-substitution mechanism as an attractive engineering experiment involving RL, H100s, containers, and benchmark design. That framing hides the important event: expert optimization work has been decomposed into tool calls, measurable outcomes, and machine iteration.

This is not pure copium. It is an early, narrow P1 demonstration. Its anesthetic function is the suggestion that narrowness makes the threat harmless. Narrowness is how automation enters: first where verification is cheap, then wherever a comparable feedback loop can be built.

The Verdict

Technically credible, economically under-read, and headline-inflated.

The excerpt does not prove generalized replacement of query optimizers or substantiate the 81% figure. It does demonstrate the mechanism that matters under the Discontinuity Thesis: expert cognitive search can be converted into a closed feedback loop, improved by a small model, and deployed against an incumbent system.

Postgres is not the corpse. Human query-planning judgment is the tissue being stripped from it. The durable moat will belong to whoever owns the data, execution environment, evaluation harness, and distribution. Everyone else becomes a servitor maintaining the machinery—or a redundant commentator explaining why the old planner was once considered intelligent.