Postgres Summit US2026

Liberatii Advisor is not an agent that invokes a collection of tuning scripts. It is an optimizer in the same architectural tradition as a database query optimizer: it constructs a search space, derives properties, prices alternatives, and selects a winner. Given the same workload facts, transformation rules, constraints, and prices, it generates the same alternatives, assigns the same costs, and makes the same decision—including the decision to keep the original design.

This session presents Liberatii Advisor as a deterministic optimizer for Postgres workloads. Advisor captures the workload and database facts, applies a fixed set of transformations to construct alternative physical designs, and evaluates every eligible alternative with a target-aware cost model. The model compares expected read benefit with write-maintenance cost, workload frequency, deployment constraints, and the cost of reaching data across the network. Candidates that do not beat the original plan remain rejected; selected and rejected alternatives retain their cost evidence and provenance.

In the first live demonstration, Advisor isolates one application workload and finds a complex five-join bottleneck. We will inspect the alternatives and their costs, approve only safe online changes, apply the selected indexes while traffic continues, and verify the result at the application boundary. Operator input supplies policy—scope, safety, and priority—not optimization logic.

The second demonstration makes two important cost components visible. We run the same analytical question over raw tables on a remote three-node YugabyteDB cluster, a maintained result on YugabyteDB, and a fresh materialization beside the application. The comparison separates distributed compute cost from network delivery cost and shows why a remote materialized result can still leave a WAN round trip on every request.

Both demonstrations use ordinary SQL and standard Postgres clients. Attendees will see the alternatives, the cost reasoning, the executed SQL, and the measured outcome—not just a generated recommendation or a polished dashboard.

Audience takeaways

• How to distinguish a database optimizer from an agent that invokes tuning tools.
• Why deterministic optimization produces repeatable, auditable decisions.
• How a workload cost model compares the original design with physical alternatives.
• How read benefit, write cost, frequency, safety, and network delivery affect selection.
• Why a cheaper query plan is only a prediction until application throughput validates it.
• Why remote materialization removes compute but not network latency.