Study Databricks ML-PRO PyFunc Serving Deployment: key concepts, common traps, and exam decision cues.
On this page
Custom serving questions become easier when you separate three layers: how the model is packaged, where it is registered, and how it is queried or deployed.
Deployment-interface map
Requirement
Better first instinct
custom inference logic packaged for serving
custom PyFunc model
keep artifacts and model object governed in Unity Catalog
register and log correctly in the governed lifecycle
query or deploy through Databricks interfaces
REST API or MLflow Deployments SDK depending on the task
What the exam is really testing
If the stem says…
Strong reading
“register a custom PyFunc model”
package custom behavior through the pyfunc interface
“query custom models”
know the serving interface, not just the artifact
“deploy via SDK, REST API, or UI”
interface choice is part of the deployment workflow
Decision order that usually wins
Separate packaging, registration, and querying.
If inference behavior is custom, decide whether PyFunc is the right packaging layer.
Make sure required artifacts travel with the model definition.
Register the governed model artifact before worrying about query interface choice.
Pick REST, SDK, or UI based on whether the task is automation, deployment, or invocation.
ML-PRO wants clean layer boundaries here. A custom model is not the same thing as the endpoint that serves it, and neither is the same thing as the interface used to call it.
Scenario triage
Scenario
Better first move
model needs custom inference behavior
package as custom PyFunc
model depends on supporting artifacts or code
include them in the package intentionally
team needs to automate deployment or querying
choose the appropriate SDK or REST interface
registry and serving concerns are getting blurred
separate artifact governance from invocation surface
Common traps
Trap
Better rule
treating the registered model as identical to the serving interface
registration and serving are different layers
forgetting custom artifacts when the custom model depends on them
packaging matters
assuming one interface is always best
the right interface depends on whether the task is deploy, query, or automate