Databricks-ML-Pro ML Ops Practice Test 2 — 15 Questions

Question 1

You are designing an ML pipeline on Databricks and want to ensure model reproducibility and auditability. Which TWO of the following actions are considered best practices for achieving this?

Accepted Answer

Log the Conda environment or requirements.txt file along with the model artifacts in MLflow.. Reproducibility relies on capturing both the code and the exact environment configuration. By logging model signatures and environment dependencies, you ensure that the model behaves identically when served in different environments. These practices are cornerstones of MLOps, allowing teams to reconstruct the exact state of a production model for debugging, compliance, or disaster recovery purposes during the model lifecycle.

Answer

Only store the final serialized model object in the MLflow Model Registry to save storage.

Answer

Rely on the default Databricks Runtime version for all training and inference tasks.

Answer

Delete old model versions immediately to keep the registry clean and organized.

Question 2

You have a batch inference pipeline that processes millions of rows daily. The model performance has recently degraded. You need to investigate if the feature distributions have shifted. Which Databricks tool is best suited for monitoring this data drift?

Accepted Answer

Databricks Lakehouse Monitoring. Detecting data drift is essential for model maintenance. Databricks provides native integration with monitoring solutions that analyze input data distributions against historical baselines. By identifying shifts in feature distributions, teams can proactively retrain models before performance impacts the business. This approach is standard in high-maturity MLOps environments where model reliability is closely tied to operational metrics and data quality.

Answer

MLflow Model Registry

Answer

Databricks Feature Store

Answer

MLflow Tracking

Question 3

You are implementing a CI/CD pipeline for your ML projects on Databricks. Which THREE of the following strategies are critical for ensuring reliable and secure model deployment?

Accepted Answer

Use environment variables to manage configuration differences between staging and production.. Reliable MLOps requires decoupling code from configuration, automated testing of artifacts, and strict access controls. By using environment-specific variables, performing integration tests on transient clusters, and utilizing service principals for deployment, you minimize the risk of production outages and security breaches. These strategies are essential for scaling machine learning operations across large enterprise teams, ensuring that deployment processes are repeatable, secure, and fully automated.

Answer

Always run CI/CD jobs using a shared, long-running interactive cluster to save costs.

Answer

Perform automated integration tests on the production model before moving it to staging.

Question 4

Refer to the exhibit. You encounter this error when hitting your model endpoint for the first time after a deployment. What is the most likely cause, and how can you mitigate this?

Accepted Answer

The endpoint is in a cold-start state and needs warm-up time.. The 503 error during initial access often occurs because the model serving endpoint is cold-starting, especially if the model image is large or dependencies take time to load. In a production context, setting up 'provisioned concurrency' or pre-warming the endpoint is essential to prevent latency spikes and service unavailability. Proper MLOps planning involves accounting for startup times in mission-critical applications to ensure reliable end-user experiences.

Answer

The model version is incompatible with the server's runtime version.

Answer

The user does not have permission to access the model endpoint.

Answer

The model artifact has been deleted from the MLflow Model Registry.

Question 5

You are designing a feature store strategy on Databricks. Which TWO of the following statements correctly describe the benefits of using Databricks Feature Store for MLOps?

Accepted Answer

It provides point-in-time joins to prevent data leakage during model training.. The Databricks Feature Store promotes consistency by ensuring the same features used during training are available for inference, preventing training-serving skew. It also enables feature discovery through a centralized registry, facilitating collaboration and code reuse across different data science teams. These capabilities are essential for scaling machine learning operations, reducing redundant work, and maintaining high performance across complex feature pipelines in production environments.

Answer

It automatically generates the raw data ingestion pipelines from external RDBMS sources.

Answer

It guarantees that all models deployed will automatically achieve 99.9% prediction accuracy.

Answer

It replaces the need for MLflow by handling all experiment tracking.

Question 6

Your company uses a multi-workspace architecture for Databricks. You need to deploy a model trained in the 'Development' workspace to the 'Production' workspace. What is the most efficient and secure way to share this model across workspaces?

Accepted Answer

Use the MLflow Model Registry cross-workspace functionality to register and deploy.. Using the Databricks Model Registry with a shared central store or utilizing MLflow model export/import is essential for cross-workspace deployment. By exporting the model artifact and associated metadata, you ensure that the production workspace receives a consistent, versioned model. This setup supports a secure separation of duties, where experimental development is isolated from the production environment, yet models can be promoted via a controlled, reproducible CI/CD process.

Answer

Copy the model notebook code to the Production workspace and retrain the model.

Answer

Share the notebook via a URL link and let production users run it.

Answer

Manually download the model from the UI and upload it to the production workspace.

Question 7

Refer to the exhibit. A junior data scientist runs the provided code snippet to track a training experiment. Why will this code likely result in a runtime error or an illogical MLflow experiment record?

Accepted Answer

It logs hyperparameters as metrics and metrics as hyperparameters.. The code attempts to pass the same dictionary to both 'log_metrics' and 'log_params'. While dictionaries are flexible, 'log_params' expects parameters (immutable settings), and 'log_metrics' expects numbers (values that change over time). The 'accuracy' field is a metric, whereas 'learning_rate' and 'batch_size' are hyperparameters. Mixing them into one structure and logging incorrectly violates MLflow's logical separation of model configuration and performance results.

Answer

MLflow does not support logging dictionaries directly.

Answer

The variable name 'log_data' is a reserved keyword in Python.

Answer

The code requires the MLflow Model Registry to be enabled first.

Question 8

An ML engineer needs to deploy a custom scikit-learn model to Databricks Model Serving while ensuring that all input data payloads are automatically validated against a predefined schema before inference execution. Which deployment approach best achieves this requirement?

Accepted Answer

Log the model artifacts alongside an explicit MLflow Model Signature and input example derived from training data.. Integrating MLflow signature validation directly into the model logging phase enforces strict input contract checks at the serving endpoint. Databricks Model Serving automatically leverages the logged signature to reject malformed requests before passing them to the scoring function, preventing runtime failures and maintaining robust ML pipeline integrity.

Answer

Log the model with a custom inference payload handler that parses incoming JSON payloads manually within the scoring function.

Answer

Attach a notebook validation widget to the serving endpoint configuration using the Databricks REST API.

Answer

Configure Unity Catalog governance rules to drop unapproved columns from incoming inference tables automatically.

Question 9

Refer to the exhibit. A data scientist is attempting to load a model saved in Unity Catalog, but the job fails. What is the most likely cause?

Accepted Answer

Insufficient permissions on the Unity Catalog volume.. When using Unity Catalog for model management, the permissions must be explicitly granted for the service principal or user running the job to access the specific volume or schema. This error indicates that the environment context lacks the required connectivity or authorization to resolve the artifact path, highlighting the importance of managing cross-workspace permissions in Unity Catalog.

Answer

The model was not registered in the MLflow Model Registry.

Answer

The MLflow version is incompatible with the model type.

Answer

The cluster does not have the 'ML' runtime installed.

Question 10

Which THREE practices are considered best for maintaining high-quality ML models in production within Databricks?

Accepted Answer

Automated unit testing for data processing pipelines.. High-quality MLOps requires automated testing, monitoring for data drift, and a formalized promotion process. Automated tests prevent regression, while monitoring ensures that models do not decay due to changing data distributions. Together, these practices ensure that production environments remain stable and that models continue to provide business value as the underlying environment evolves.

Answer

Deploying models immediately after training completes.

Answer

Sharing a single MLflow experiment for all users.

Question 11

Refer to the exhibit. What is the effect of this policy on your Databricks Model Registry?

Accepted Answer

Only Admins can modify models in the Production stage.. This policy implements a 'write-once' protection for production models, preventing unauthorized users from overwriting or modifying production-ready models. In an MLOps environment, this is critical to ensure that once a model is approved for production, it cannot be accidentally or maliciously altered. This enforces the separation of concerns between model promotion and model development, ensuring integrity in the production lifecycle.

Answer

Admins are denied access to the Production stage.

Answer

All users can promote models to the Production stage.

Answer

The policy prevents reading models in Production.

Question 12

Which TWO of the following are essential for implementing an effective CI/CD process for ML on Databricks?

Accepted Answer

Integration with Databricks Repos for Git-based versioning.. Effective CI/CD requires both environment consistency and automated verification. Using Databricks Repos ensures that code versions are tracked and easily deployed, while automated testing ensures that model artifacts meet performance criteria. These together form the backbone of a reliable, automated deployment pipeline, minimizing manual errors and ensuring that every production change is tested.

Answer

Manually downloading logs to a local machine.

Answer

Running all tests manually in the production workspace.

Answer

Sharing all secrets in plain-text notebooks.

Question 13

You are preparing a scikit-learn model for deployment to a Databricks Model Serving endpoint. When you log the model with MLflow, which flavor should you use to ensure the endpoint can load and serve the model with a minimal, optimized Python environment?

Accepted Answer

mlflow.sklearn. Logging a scikit-learn model with the mlflow.sklearn flavor ensures that the model is saved in a format that Databricks Model Serving can directly load. This flavor automatically includes the necessary scikit-learn dependencies and a pyfunc-compatible signature, enabling efficient, low-latency serving without manual environment specification.

Answer

mlflow.pyfunc

Answer

mlflow.onnx

Answer

mlflow.tensorflow

Question 14

You are building a CI/CD pipeline for a machine learning model using Databricks Repos and MLflow. The pipeline must automatically test, register, and deploy the model to a staging endpoint. Which TWO steps are essential to include in the pipeline to ensure the model can be served with the correct dependencies? (Choose two.)

Accepted Answer

Log the model using MLflow with a conda environment that specifies all required libraries and their versions.. Logging the model with a conda environment ensures that all libraries and versions are captured, allowing Model Serving to recreate the runtime. Logging the model signature defines the input and output schema, which the endpoint uses to validate and format requests. Together, these steps guarantee that the model can be served correctly with its dependencies.

Answer

Set the model version's description to include the Git commit hash of the training code.

Answer

Create a custom Docker image that includes the model and its dependencies, and deploy it to a serving endpoint.

Answer

Register the model in the MLflow Model Registry with a stage transition to 'Staging'.

Question 15

You are deploying a model to a Databricks Model Serving endpoint that requires a custom Python library not available in the default environment. You have logged the model with MLflow and specified the library in the conda environment. However, when you attempt to serve the model, the endpoint fails to start. What is the most likely cause?

Accepted Answer

The model was logged with a conda environment that includes pip dependencies, but the library requires system packages that are not available in the serving environment.. Databricks Model Serving installs Python libraries from the conda environment but does not install system-level packages. If the custom library requires such packages, the endpoint will fail to start. Other options are less likely because Model Serving supports conda-based libraries and Python version selection, and a missing signature would not prevent startup.

Answer

The MLflow model was logged with a conda environment that specifies Python 3.8, but the serving endpoint only supports Python 3.9.

Answer

The custom library is not pre-installed on the serving cluster, and Databricks Model Serving does not support installing libraries from conda at runtime.

Answer

The model signature is missing, so the endpoint cannot determine the input schema and fails to initialize.