Choosing an example
The examples are independent. The EKS walkthroughs use the same five-table retail model so you can compare engines and catalogs directly. The laptop walkthrough uses three SaaS models to show how an agent composes semantic queries with richer DataHub context.
Choose the walkthrough that matches the infrastructure and behavior you want to evaluate.
Each walkthrough contains its own run order. Read Prerequisites once for the shared workstation, cluster, image, and credential requirements.
Pick a stack
Section titled “Pick a stack”| Walkthrough | Model | Catalog | Engine | Good for |
|---|---|---|---|---|
| Retail on Glue and StarRocks | Retail | AWS Glue | StarRocks | The reference EKS path |
| Retail on Glue and Trino | Retail | AWS Glue | Trino | Comparing the same model on another engine |
| DataHub, Polaris and Trino | Retail | Apache Polaris, enriched by DataHub | Trino | Deriving physical structure and importing business meaning |
| Everything on your laptop | SaaS revenue, adoption, and support | Apache Polaris, enriched by DataHub | StarRocks | Multiple models, two MCP servers, OPA, and no cloud account |
What each walkthrough proves
Section titled “What each walkthrough proves”The end-to-end walkthroughs share four checks even though their datasets and infrastructure differ.
A model reconciles. It validates, binds to real tables, passes a drift check against the live schema, and publishes a versioned artifact.
A certified metric returns a verified number. The retail walkthroughs use
store_productivity, a ratio that needs fan-out-safe planning. The laptop walkthrough uses
certified SaaS metrics across three business domains.
The same request is deterministic. Repeating the same authorized request against the same model version and identity produces the same request hash and SQL. A configured cache can reuse the plan or result.
Governance is enforced before SQL. The examples exercise denied fields, row filters, or an external OPA decision depending on the stack.
Consumers share definitions. Agents use MCP, applications use REST, and BI tools can read governed views in the engine.
Beyond the walkthroughs
Section titled “Beyond the walkthroughs”The flights model is a second model with no deployment attached, useful for seeing how a different domain is structured.
See identity propagation by hand is a curl walkthrough on Trino. You mint a real Keycloak token and watch the same query return different results for different users, with the engine enforcing policy under each caller’s identity.
Govern a model with OPA is a curl walkthrough on Trino. You deploy an Open Policy Agent decision engine and watch two authorization layers apply to one query, the model’s built-in governance and an external OPA policy.
Benchmark results has the measured comparison between raw text to SQL and the semantic layer, with the method and the per question verdicts.