13. Running it for real | Semantic Layer course

Section 1

Start small, and write down what you skipped

Next section

Section 2

Showcase questions become your test suite

Next section

Section 3

The rest of the practices, with what Meridian did

Next section

Section 4

Seven ways this goes wrong

Next section

Section 5

A phased plan that survives contact

Next section

Section 6

Local stack to AWS, component by component

Next section

Section 7

Tuesday, one more time

Hands on

Ship it on your tailnet

30 min

Run the full course stack with docker compose, verify each service is healthy, confirm the course app persists your progress across a container restart, and expose it over Tailscale.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab13-deploy

Full instructions, starter files and the solution are in labs/modules/lab13-deploy/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

7 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 A team has spent nine months building an enterprise-wide ontology covering all thirteen business domains and has not yet answered a single business question. Which anti-pattern is this, and what is the fix?

Materializing everything; virtualize with R2RML instead Boiling the ocean; scope to a question somebody is already arguing about, ship that, then extend Treating the ontology as a one-time project; assign a standing owner No versioning; put the ontology in Git

2 A senior engineer argues that SHACL is redundant because the OWL ontology already declares that every loan application requires at least one KYC document. What is wrong with that?

Nothing, provided the OWL cardinality axiom is set to minCardinality 1 OWL is open-world, so a missing KYC document makes it infer that one must exist somewhere rather than reporting a violation; SHACL is closed-world and reports the violation with a focus node and a path SHACL is faster, so it should be preferred for performance reasons OWL cannot express cardinality constraints at all, so the axiom does not exist

3 Which two things do NOT change when you move this course's stack from the local docker compose setup to AWS? (pick two)

Select all that apply.

The SPARQL queries, including the federated ones The document embeddings and the k-NN index built from them The R2RML mappings and the SHACL shapes Authentication and the endpoint URLs

4 Meridian's data team built a database view called loan_semantic_v1 that resolves dpd, stat_cd, npa_flag and wo_dt into one arrears_class column. What did that buy them?

It removed the need for an ontology, since the view now carries the definition The R2RML mapping stopped containing business logic and became reviewable, and the default count query went from 4.2 seconds to 0.9 because Postgres evaluates the CASE on indexed columns It let them delete the Metric Store, because the view computes the metric It made materialization unnecessary for the inferred graph

5 A team wants to skip phase 2 (reasoning, SHACL, governance) and go straight from a virtualized ontology to building agents, because agents are what the business asked for. What is the consequence?

None, provided the ontology is complete enough to answer the showcase questions An agent over an ungoverned ontology produces unexplainable answers faster, which is the Module 1 problem with better latency The agent cannot be built at all without SHACL shapes MCP tool discovery will fail because the ontology has no versions

6 The showcase-question suite asserts on rules_fired, not just on the numeric answer. Why bother?

Because SPARQL results are non-deterministic without an ORDER BY clause Because a right answer produced by the wrong rule is right today and wrong the day somebody edits that rule, so you assert on the derivation as well as the number Because the reasoner is slower than the query engine and the timing needs isolating Because reasoning_steps is required by the ask response schema

7 After thirteen modules, Meridian's three teams still report 1,847, 4,102 and 1,203. What did the semantic layer actually change?

Nothing meaningful, since the disagreement persists Each definition now has an IRI, a version, a rule that computes it, a citation to the policy page that authorizes it and a committee that approved it, so any answer states which one it used and anybody can check The three numbers now converge to 1,203, which is the approved definition The teams now share one query, so only one number is reachable