Databricks GenAI assembling and deploying apps
Assembling and Deploying Apps is worth 22% of the Databricks Generative AI Engineer Associate exam — around 10 of the 45 scored questions and the second largest section. Together with application development it makes building and shipping 52% of the exam.
This section is about turning something that works in a notebook into something that runs.
Packaging an application
- Assembling the components — retriever, chain, prompts, model configuration — into a deployable artefact.
- Managing dependencies so the deployed version behaves like the one you tested.
- Configuration that differs by environment, kept out of the code.
MLflow
Central to this section and to the platform generally:
- Tracking — recording parameters, metrics and artefacts from experiments.
- Model registry — versioning what exists and controlling what is promoted.
- Packaging and deployment — MLflow is the route from a registered model to a serving endpoint.
- Reproducibility — knowing exactly which version is live and what produced it.
If a question asks how you would know which version is deployed, or how to manage promotion from development to production, MLflow is almost always in the answer.
Model serving
- Serving endpoints that host the application and expose it to callers.
- Real-time serving versus batch processing, and which a requirement calls for.
- Scaling for load, and what that costs.
- Latency — a RAG application has several sequential steps, so latency accumulates across retrieval, prompt assembly and generation.
Unity Catalog
Governance of data and models, and the answer to most access-control questions on this exam:
- Centralised access control over tables, volumes and models.
- Lineage — what a model was built from, and who used it.
- Consistent permissions across the workspace rather than per-object exceptions.
Cost in production
Databricks expects cost awareness:
- Tokens are the dominant cost driver in a generative application. More retrieved chunks means more tokens per query, every query.
- Endpoints bill while provisioned. An internal tool used occasionally does not need capacity held permanently.
- A smaller model that meets the requirement beats a larger one that exceeds it.
Sample questions
Question 1. A team must be able to state exactly which version of a generative AI application is serving production traffic, and roll back quickly if a new version behaves badly. Which capability supports this?
- A. Unity Catalog permissions on the underlying tables
- B. The MLflow model registry with versioned deployments
- C. The vector search index configuration
- D. A shared notebook documenting the current version
Show answer
Answer: B
The MLflow model registry versions artefacts and controls promotion between stages, which is what identifying the live version and rolling back requires. Unity Catalog governs access and lineage, a vector index stores embeddings, and a notebook is a development surface rather than a deployment control.
Question 2. An internal RAG assistant is used roughly 40 times a day at unpredictable intervals. The team wants to minimise cost and tolerates occasional slower first responses. What should they consider?
- A. A permanently provisioned endpoint sized for peak load
- B. A larger model to answer faster
- C. An endpoint configuration that scales down when idle
- D. Retrieving more chunks so fewer follow-up questions are needed
Show answer
Answer: C
An endpoint that scales down when idle avoids paying for capacity that sits unused between sporadic requests, which suits low unpredictable volume where cold starts are acceptable. Permanently provisioned capacity bills continuously, a larger model raises cost per query, and retrieving more chunks increases token spend.
Question 3. Different teams must have different access to the documents a RAG application can retrieve, with a record of who accessed what. Which Databricks capability governs this?
- A. Unity Catalog
- B. The MLflow tracking server
- C. Serving endpoint configuration
- D. Access instructions written into the system prompt
Show answer
Answer: A
Unity Catalog provides centralised access control and lineage across data and models, which is what per-team permissions with an audit record requires. MLflow manages model lifecycle, endpoint configuration controls serving, and prompt instructions do not enforce access.
What to practise
Take the RAG application from your development work and actually ship it: register it in MLflow, deploy it to a serving endpoint, govern its data in Unity Catalog, then deploy a second version and roll back to the first. That sequence is this entire section, and doing it once removes any ambiguity about which tool does what.