Task 4.4: Configure a persistent datastore to store and retrieve intermediate memory or structured information; Integrate managed, external, and custom MCP servers based on a given application requirements; Develop an appropriate interactive user facing interface for an agent usage scenario (Apps, Slack, Teams, etc.)

• AAn AI Search index of the profiles, with cases written to a Delta table through a SQL warehouse

  Incorrect

  Similarity search is not a keyed profile lookup, and a SQL warehouse over Delta is built for analytics, not per-request transactional writes.

• BProfile lookups and case writes against the Delta gold table through a SQL warehouse

  Incorrect

  The lakehouse is optimized for analytics; fast lookup-style queries with full ACID transactions are what Lakebase is designed for.

• CA Lakebase synced table of the profiles, with each case inserted and updated in that table

  Incorrect

  The sync pipeline owns a synced table's data: it is read-only for applications, and direct writes are overwritten on the next refresh.

• DA Lakebase synced table of the profiles, with cases in the app's own Postgres table beside it

  Correct

  A managed sync pipeline serves the gold table in Postgres for low-latency reads, and the app inserts and updates its own tables in the same database.

Synced tables bring Unity Catalog data into Lakebase with a fully managed pipeline, so no sync scripts are needed. Applications query the synced data alongside their own operational tables, which they read and write with standard Postgres drivers.

• ARegister it as an external MCP in Unity Gateway with an HTTP connection for its OAuth

  Correct

  External MCP registration creates a Unity Catalog connection for the credentials and makes the server a governed service that can be shared.

• BRebuild the vendor's tools as a custom MCP server on Databricks Apps and share that app with each team

  Incorrect

  Custom MCP servers on Apps are for your own tools or APIs; rewriting a vendor's hosted server adds work and maintenance.

• CHard-code the vendor token in each agent's environment variables

  Incorrect

  Per-agent credentials are not shared, governed or audited centrally.

• DAdd the vendor server to the Genie One MCP configuration

  Incorrect

  Genie One is a Databricks-provided MCP for business questions over governed data; it does not proxy other vendors' servers.

Use a Databricks-provided MCP when one exists, register a self-hosted or third-party server as an external MCP, or build a custom server on Databricks Apps for your own tools. Registered MCPs use Unity Catalog permissions.

• AThe app URL with /mcp appended

  Correct

  Copy the app URL from its Overview page and append /mcp to connect an agent.

• BThe workspace URL with /serving-endpoints

  Incorrect

  Serving endpoints are a different service.

• CThe app URL with /api/2.0/genie

  Incorrect

  That is a Genie API path, not the app's MCP endpoint.

• DAny URL, if the app name starts with mcp-

  Incorrect

  The mcp- prefix is a naming convention that helps Databricks tools recognize the server; the endpoint is still /mcp.

Databricks Apps provides the compute and an HTTPS URL for a custom MCP server. For the Starter and OpenAPI templates the MCP path is /mcp; an existing server uses the path it was configured with.

• AThe app can adopt the job's existing service principal at creation

  Incorrect

  You cannot specify an existing service principal at app creation.

• BThe app runs as the workspace admin

  Incorrect

  App authorization uses the app's own service principal.

• CThe app always gets its own new service principal

  Correct

  Databricks provisions a dedicated service principal per app, which cannot be changed or replaced with an existing one.

• DThe app runs without any identity

  Incorrect

  Every app has an identity for app authorization.

App authorization uses the app's own service principal; it stays the same across deployments and is deleted with the app.

• ARead replicas

  Incorrect

  Replicas add compute for read scaling.

• BScale to zero

  Correct

  Scale to zero suspends Lakebase compute after inactivity, which suits development and test databases.

• CHigh availability

  Incorrect

  High availability adds standby compute for failover.

• DSynced tables

  Incorrect

  Synced tables serve lakehouse data; they do not suspend compute.

Lakebase offers autoscaling, scale to zero, branches, read replicas, instant restore and high availability.

• AHard-code the key in the app's source instead of the environment

  Incorrect

  Databricks says never to store sensitive values in app code or directly in environment variables.

• BRaise the secret resource to Can manage so that the value is masked

  Incorrect

  The permission level sets what the app may do with the scope's secrets; it does not hide a value that is injected directly.

• CStore the key in a second scope and add both scopes as resources

  Incorrect

  More scopes do not change how the value reaches the app or where it is shown.

• DDefine the variable with valueFrom: partner_key in app.yaml

  Correct

  A variable defined with valueFrom keeps the secret value managed by Databricks and out of plain text; values injected directly appear on the Environment page.

Databricks Apps injects each secret resource as an environment variable named after its resource key. Define the variable in app.yaml with valueFrom and retrieve the value in app code at runtime, so the secret is never written out in plain text.