Databricks Certified Machine Learning Professional practice questions

Question 1

A company is migrating its model development to MLflow 3 on Databricks and must decide which set of capabilities MLflow 3 provides specifically for ML model development. Per the MLflow 3 on Databricks documentation, which option matches?

Accepted Answer

Experiment tracking, model evaluation, a production model registry, and model deployment tools for model development.

Answer

Source-control hosting, CI runners, and container builds, replacing the team's existing Git and image-registry tooling.

Answer

Only prompt management and LLM-as-a-judge scoring, with no support for classical experiment tracking or registries.

Answer

Feature computation and online serving of feature tables, taking over the role of the Databricks Feature Store entirely.

Question 2

A payments fraud ML team is designing data-versioning for its training pipeline and must decide which Databricks technology supplies ACID transactions and time travel for the training data itself. Per Databricks CI/CD best practices, which technology fits?

Accepted Answer

Delta Lake, which provides ACID transactions and time travel for data versioning across the pipeline.

Answer

MLflow Model Registry, which provides ACID transactions and time travel for the training data itself.

Answer

Unity Catalog access controls, which version the raw training data by snapshotting permissions over time.

Answer

Declarative Automation Bundles, which version the training data by deploying it atomically with YAML.

Question 3

An MLOps lead at a real-time SaaS personalization platform is rolling out the OpenAI-compatible SDK against a Databricks Model Serving endpoint and is being asked which client patterns the documentation lists for querying the endpoint. Per the Model Serving documentation, which TWO options match?

Accepted Answer

Querying Databricks Model Serving endpoints supports multiple client interfaces including the OpenAI-compatible SDK, REST API, MLflow Deployments SDK, and the Databricks Python SDK.

Accepted Answer

The REST API endpoint follows the pattern POST /serving-endpoints/{name}/invocations with JSON-formatted inference requests.

Answer

Querying a Databricks Model Serving endpoint is only supported through a per-endpoint custom gRPC stub generated at deployment time.

Answer

The REST API endpoint at Databricks Model Serving follows a per-tenant DNS pattern with the model serving endpoint name encoded only in the host portion of the URL and not in any documented path-style invocations URL.

Question 4

An ML platform engineer at a healthcare diagnostic AI team is comparing distributed hyperparameter-tuning libraries for a single-machine PyTorch model on a Spark cluster and must decide which TWO documented patterns are supported. Per the distributed training documentation, which TWO options match?

Accepted Answer

Optuna with MlflowSparkStudy distributes optimization across PySpark executors using MlflowStorage as a shared backend.

Accepted Answer

Hyperopt SparkTrials distributes single-machine model tuning to Spark workers.

Answer

Optuna can only run on the Spark driver node as a single-process loop.

Answer

Hyperopt SparkTrials runs entirely on the Spark driver and never dispatches trials to Spark workers.

Question 5

An engineer is defining a Databricks job and a DLT pipeline for a recommender as version-controlled YAML in a Declarative Automation Bundle. Per the ML validation testing and Databricks Asset Bundles documentation, what do bundles make it possible to describe as source files?

Accepted Answer

Databricks resources such as jobs and pipelines, described alongside the project's metadata as source files.

Answer

Only Python wheel artifacts, while jobs and pipelines must still be clicked together by hand in the workspace UI.

Answer

Cluster billing reports, so the finance team can reconcile compute spend directly from the project's source repo.

Answer

User identities and group memberships, so the bundle provisions Unity Catalog principals at each deploy time.

Question 6

A team is evaluating how to access foundation models on Databricks for a new assistant feature and is weighing the available access modes and tooling. Per the Databricks Foundation Model APIs overview, which TWO statements are correct?

Accepted Answer

Base models such as Meta-Llama-3.3-70B-Instruct are available for immediate use with pay-per-token pricing.

Accepted Answer

Workloads needing performance guarantees or fine-tuned variants can be deployed with provisioned throughput.

Answer

Externally hosted models such as GPT-4 must be self-hosted on the team's own clusters before they can be queried.

Answer

The AI Playground is the required gateway that all production scoring requests to these models must pass through.

Question 7

A data science team is evaluating open-source libraries that Databricks documents for automating and scaling hyperparameter tuning across multiple computes. Per the Databricks hyperparameter tuning documentation, which option matches?

Accepted Answer

Optuna, Ray Tune, and Hyperopt, which automate and scale hyperparameter tuning across multiple computes.

Answer

Petastorm, Delta Live Tables, and Auto Loader, the libraries Databricks documents for tuning models.

Answer

TorchDistributor, DeepSpeed, and Ray, which Databricks documents for scaling hyperparameter search batches.

Answer

MLflow Tracking, Feature Store, and Unity Catalog, which together perform hyperparameter optimization.

Question 8

An MLOps lead at a fintech must decide how access to a registered credit-risk model is governed for analysts, data scientists, and platform engineers across catalogs. Per the Unity Catalog Model Registry documentation, which option matches?

Accepted Answer

Access control through Unity Catalog GRANTS manages who can read, write, or manage models.

Answer

Unity Catalog Model Registry has no concept of per-model access control and grants every workspace user.

Answer

Access to models registered in Unity Catalog Model Registry is governed exclusively by workspace-level.

Answer

Access control over registered models requires writing a custom OPA policy bundle and attaching it to AI Gateway.

Question 9

A developer is packaging a custom MLflow Python model whose embedding weights take several seconds to load, and wants inference latency to stay low on every request. Per the Databricks custom Python code deployment guidance, where should the developer load those weights?

Accepted Answer

Load the weights inside load_context so they load only once, keeping the predict function fast on each call.

Answer

Load the weights inside predict so they refresh on every request and never go stale between successive calls.

Answer

Load the weights in the model __init__ constructor, which MLflow re-runs for each invocation of predict.

Answer

Pass the weights as a field in each scoring request payload so the model never has to load them itself.

Question 10

An MLOps lead at a real-time SaaS personalization platform is enabling MLflow autologging for a SparkML CrossValidator job and must decide what autolog() actually captures. Per the MLflow PySpark ML autologging documentation, which option matches?

Accepted Answer

MLflow PySpark ML autologging via mlflow.pyspark.ml.autolog() automatically logs hyperparameters, metrics, and models from CrossValidator and TrainValidationSplit runs.

Answer

mlflow.pyspark.ml.autolog() only captures the final aggregated CrossValidator metric in a flat run and discards every per-combination hyperparameter trial.

Answer

mlflow.pyspark.ml.autolog() runs only inside Databricks Connect from a local machine and is not supported on Databricks Runtime clusters.

Answer

mlflow.pyspark.ml.autolog() logs only the trained model artifact and never records hyperparameters or evaluation metrics.

Question 11

An MLOps engineer at an e-commerce recommendation team is tasked with serving a model in real time and must decide how the latest feature values reach the model at inference. Per the Databricks Feature Store documentation, which option matches?

Accepted Answer

The model automatically looks up the latest feature values at inference from the store

Answer

The caller must pass every current feature value inside each scoring request payload sent

Answer

Lakehouse Monitoring streams the latest feature values to the endpoint ahead of each call

Answer

The model reads its feature values from the MLflow run artifacts that were saved earlier

Question 12

A company needs to automatically capture the request inputs and prediction outputs of an existing production model serving endpoint so analysts can later debug drift using Databricks SQL. Per the Databricks Model Serving guidance, which capability should the company enable?

Accepted Answer

Enable inference tables so request inputs and responses are continuously logged to a Delta table in Unity Catalog.

Answer

Enable the Serving UI query panel so each request is displayed and then exported as a CSV file for later review.

Answer

Attach a Lakehouse Monitoring metric table that recomputes endpoint latency without retaining the raw request payloads.

Answer

Turn on cluster event logs so the serving compute writes inputs and predictions into the driver log files.

Question 13

An ML platform engineer must choose between the FeatureEngineeringClient and the FeatureStoreClient when calling create_training_set, depending on whether the feature tables live in Unity Catalog or the Workspace Feature Store. Per the Train models with feature tables documentation, which option matches?

Accepted Answer

Use FeatureEngineeringClient.create_training_set for Feature Engineering in Unity Catalog and FeatureStoreClient.create_training_set for Workspace Feature Store.

Answer

Use FeatureStoreClient.create_training_set for Unity Catalog feature tables and FeatureEngineeringClient.create_training_set for the Workspace Feature Store tables.

Answer

Use FeatureEngineeringClient.create_training_set for both stores, because the Workspace Feature Store shares the same Unity Catalog client entry point now.

Answer

Use FeatureStoreClient.create_training_set for both stores, since Unity Catalog feature tables are reached through the legacy workspace client by default.

Question 14

A department is housekeeping several stale model versions in the legacy Workspace Model Registry and must decide what the Archived stage implies and who is allowed to move a version between stages. Per the Workspace Model Registry documentation, which option matches?

Accepted Answer

Archived means the version is inactive and a deletion candidate, and a user with appropriate permission can transition versions between stages.

Answer

Archived means the version is locked permanently, so no user, regardless of their permission level, can ever transition it again at all.

Answer

Archived means the version is actively serving fallback traffic, so only the original model author is ever allowed to transition it to another stage.

Answer

Archived means the version remains in active testing, so any workspace viewer can freely transition it straight to Production at will.

Question 15

An engineer is designing the provisioned concurrency for a custom model serving endpoint that must handle about 50 queries per second, where each request takes roughly 0.2 seconds of model execution time. Per the Databricks custom model serving guidance, how should the engineer estimate the required concurrency?

Accepted Answer

Multiply QPS by model execution time, giving about 10 provisioned concurrency units for the endpoint.

Answer

Divide QPS by model execution time, giving roughly 250 provisioned concurrency units for the endpoint.

Answer

Add QPS to model execution time, giving approximately 50 provisioned concurrency units for the endpoint.

Answer

Set concurrency equal to QPS alone, giving about 50 provisioned concurrency units regardless of execution time.

Databricks Certified Machine Learning Professional

Sample questions

Sources

Similar exams