Databricks-ML-Pro Model Development — 100 Questions

Question 1

A data scientist is training a Random Forest model on Databricks. They want to ensure that the hyperparameter tuning process is efficient while preventing overfitting during the training phase. Which approach best achieves this goal using MLflow and Hyperopt?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Configure Hyperopt with the Tree of Parzen Estimators (TPE) algorithm and include cross-validation in the objective function.. Using Hyperopt with MLflow allows for parallelized hyperparameter optimization, which is crucial for efficient search spaces. By incorporating cross-validation within the objective function, the model is evaluated on multiple data folds, effectively preventing overfitting. This integrated approach ensures that models are both performant and robust, leveraging Databricks' distributed compute clusters to minimize training time while maintaining high-quality hyperparameter selection.

Answer

Perform a grid search on a single node without cross-validation.

Answer

Utilize MLflow autologging to automatically log all hyperparameter trials during a random search.

Answer

Manually tune hyperparameters by iterating through individual values in a notebook loop.

Question 2

A machine learning engineer needs to handle imbalanced datasets for a fraud detection model using PySpark and MLlib. Which TWO techniques are valid for mitigating this issue within the Databricks environment?

Accepted Answer

Increase the weight of the minority class samples using the weight column in the classifier.. Addressing class imbalance is critical for fraud detection where the minority class is the primary focus. Oversampling the minority class or undersampling the majority class are standard statistical practices. Implementing these directly within the Spark pipeline ensures they scale across large distributed datasets, which is vital for maintaining model integrity in big data environments where manual sampling techniques would fail or be inefficient.

Answer

Apply SMOTE (Synthetic Minority Over-sampling Technique) using standard scikit-learn libraries.

Answer

Filter out majority class rows to create a balanced dataset before training.

Answer

Use the standard cross-validation object with no modification.

Question 3

Refer to the exhibit. A data scientist is deploying a model to the Model Registry. What is the impact of the registered_model_name parameter in the log_model function?

Accepted Answer

It creates a new version of the registered model if the name already exists.. The registered_model_name parameter instructs MLflow to create or update a registered model entry in the Model Registry. This is a critical step for transitioning from training to deployment, as it creates a centralized location for managing model versions. By providing this parameter, the model is automatically cataloged, allowing downstream systems to reference the latest version in production automatically.

Answer

It forces the model to immediately transition to the 'Production' stage.

Answer

It saves the model locally to the driver node's disk.

Answer

It deletes all existing versions of the model with that name.

Question 4

A data scientist is training a deep learning model using TensorFlow on Databricks. They observe that training speed is slow despite using a GPU cluster. Which action should they prioritize to optimize performance?

Accepted Answer

Convert the dataset into TFRecords format and use tf.data API for efficient loading.. Deep learning models in Databricks often suffer from I/O bottlenecks when data is not optimally staged or formatted. Using the Petastorm library or converting data into TFRecords allows for efficient, parallelized loading of large datasets directly into the GPU memory. This minimizes idle time for the GPUs, ensuring that the computational power of the cluster is fully utilized during the training phase of the neural network.

Answer

Increase the number of threads on the driver node.

Answer

Switch from GPU to CPU instances to reduce costs.

Answer

Reduce the batch size to the smallest possible value.

Question 5

A machine learning engineer needs to deploy a custom model that uses a non-standard pre-processing library. How should they package the model to ensure the environment is correctly replicated during inference?

Accepted Answer

Pass a dictionary of dependencies to the 'conda_env' argument in the log_model function.. When using non-standard libraries, the MLflow 'conda_env' or 'pip_requirements' parameter in 'log_model' is critical. It explicitly defines the environment dependencies required to run the model. This ensures that the inference environment is an exact mirror of the training environment, which is the most reliable way to prevent 'it works on my machine' issues when deploying models into production clusters or serving endpoints.

Answer

Rely on the cluster's pre-installed libraries and assume they are present.

Answer

Manually install the required library on every node of the inference cluster.

Answer

Only log the model object and omit the environment specification.

Question 6

When sharing a Databricks notebook that contains sensitive model training logic, which feature should be used to provide users access without exposing the underlying source code?

Accepted Answer

Use the 'Run as owner' credential passthrough for the notebook.. Providing access to model results and visualizations without exposing the underlying source code is handled via Workspace permissions and 'Run as owner' notebook workflows. This allows authorized users to interact with the model's outputs (like plots, logs, or metrics) without compromising the intellectual property stored in the code. This is essential for secure collaboration in large organizations where different roles have different access needs.

Answer

Convert the notebook to a PDF and email it to the users.

Answer

Share the notebook with 'Can View' permissions for all users.

Answer

Copy the code into a public GitHub repository.

Question 7

A data scientist is training a PyTorch model on Databricks using MLflow. They need to ensure that the model architecture and training parameters are automatically logged without explicitly calling mlflow.pytorch.autolog() in every notebook. Which approach is the most efficient?

Accepted Answer

Call mlflow.autolog() at the start of the notebook session.. Databricks Runtime for Machine Learning includes pre-configured MLflow autologging. By setting the environment-wide configuration via Spark properties or using the mlflow.autolog() function at the beginning of the session, the environment automatically captures model artifacts, parameters, and metrics for supported frameworks. This reduces boilerplate code and prevents oversight, ensuring experiment reproducibility and auditability across the entire team's experimentation lifecycle within the workspace.

Answer

Manually wrap every training loop in a with mlflow.start_run() block to capture parameters.

Answer

Call mlflow.pytorch.autolog() at the start of the notebook cell.

Answer

Configure the MLflow tracking URI to a remote server using a custom tracking client.

Question 8

Which TWO of the following steps are required to properly register a custom Python model in the Unity Catalog Model Registry using the MLflow Fluent API?

Accepted Answer

Use the 'uc:' prefix in the model_uri parameter during the registration process.. Registering models in Unity Catalog requires a specific workflow to ensure governance and lineage. First, you must log the model using a supported flavor or custom pyfunc. Second, you must use the 'uc:' URI prefix in the log_model or register_model call to direct the artifact to the Unity Catalog rather than the workspace-local registry. This ensures the model is discoverable and governable.

Answer

Call mlflow.register_model() using a standard 'models:/' URI.

Answer

Manually upload model pickle files to an external S3 bucket before registration.

Answer

Convert the model to an ONNX format before registering it to the catalog.

Question 9

When training a model using Scikit-Learn in a Databricks notebook, which MLflow component should a data scientist use to group multiple related training runs under a single experiment?

Accepted Answer

mlflow.set_experiment(). MLflow Experiments serve as the primary organizational unit for model development. By associating runs with a specific experiment, teams can compare performance across different iterations, track progress, and manage metadata. Proper use of experiments is essential for maintainable machine learning workflows, as it prevents the cluttering of run history and enables effective result visualization in the MLflow UI.

Answer

mlflow.start_run()

Answer

mlflow.log_param()

Answer

mlflow.register_model()

Question 10

Which THREE of the following are valid ways to improve the performance of a model during the development phase in Databricks?

Accepted Answer

Using Hyperopt for automated hyperparameter tuning.. Model performance is improved through a combination of feature engineering, hyperparameter optimization, and proper data sampling. Hyperopt provides automated search for parameters, feature stores ensure consistency and reuse, and cross-validation provides a robust estimate of model performance. Using these tools within the Databricks ecosystem ensures that improvements are statistically sound and easily reproducible for production deployment.

Answer

Deploying the model to a Multi-Model Serving endpoint.

Answer

Increasing the number of hidden layers in a neural network blindly.

Question 11

Which of the following describes the correct relationship between an MLflow Run and a Model Version?

Accepted Answer

A model version references a specific artifact logged during a run.. A Run is an execution of training code that logs metrics and parameters. A Model Version is an entity in the Model Registry that wraps the resulting model artifact from one or more runs. You can create multiple versions of a model over time, each originating from different successful runs, ensuring that every deployment is tied back to a specific, auditable training history.

Answer

A run is a child object of a model version.

Answer

A run automatically converts into a model version upon completion.

Answer

A model version is created by merging multiple runs together.

Question 12

A data scientist is training a Random Forest model on Databricks. They want to ensure that the hyperparameter tuning process is efficient while preventing overfitting during the training phase. Which approach best achieves this goal using MLflow and Hyperopt?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use MLflow's log_model with a custom flavor that supports distributed serialization.. Memory pressure on the driver during aggregation typically occurs when large model artifacts or prediction results are collected into memory. By leveraging MLflow's native integration with Spark, the model can be distributed or persisted to storage without pulling the entire object into the driver's memory space. This ensures scalability when handling large-scale datasets, which is essential for production-grade machine learning pipelines in distributed environments.

Answer

Increase the driver node instance type to a memory-optimized class.

Answer

Downsample the dataset by 90% to fit into the current driver memory.

Answer

Collect the entire dataset into a local Pandas DataFrame before training.

Question 13

Which TWO of the following configurations are required to effectively track and reproduce a deep learning experiment using MLflow in Databricks?

Accepted Answer

Include the current Git commit hash in the MLflow run tags.. Tracking and reproducibility require capturing both the code state and the environment dependencies. By logging the git commit hash and the conda.yaml or requirements.txt environment file, you ensure that any collaborator can recreate the exact training environment. This is critical for model governance, auditing, and debugging performance regressions in complex deep learning pipelines where environment variations often lead to non-deterministic training outcomes.

Answer

Log the training dataset as an artifact using mlflow.log_artifact().

Answer

Store the model weights in a local directory on the driver node.

Answer

Increase the number of threads for data preprocessing.

Question 14

Which THREE actions are required when migrating a custom Scikit-learn model to the Databricks Model Registry to ensure it can be served via MLflow Model Serving?

Accepted Answer

Specify an input signature (ModelSignature) when logging the model.. To serve a model, MLflow needs to understand how to load the model (the flavor), the environment requirements, and the input schema. By defining a custom signature and environment, you ensure the model behaves consistently during inference. Serving infrastructure relies on these components to construct a containerized environment that matches the training setup, preventing runtime errors and ensuring the model can handle incoming production requests.

Answer

Rename the local variable to 'model_final' before saving to registry.

Answer

Convert the model to a PMML file for faster inference.

Question 15

Refer to the exhibit. A data engineer attempts to transition a model version to 'Staging' but receives this error. What is the most likely cause?

Accepted Answer

The version number '5' was deleted or never successfully registered.. The error indicates the registry cannot locate the specified version, often because it was deleted or the name is slightly misspelled. In Databricks, model versioning is strictly managed by the Model Registry. This error highlights the necessity of verifying artifact existence through the API before performing state transitions, which prevents pipeline failures in automated CI/CD workflows that rely on registry state changes.

Answer

The model version is archived and must be unarchived before transitioning.

Answer

The model name 'DemandForecast' does not exist in the Model Registry.

Answer

The user lacks the 'CAN_MANAGE' permission on the model.

Question 16

A data scientist is using Databricks to train a model with a very large vocabulary, leading to high-dimensional sparse inputs. Which storage format for the feature table is most efficient for both storage and retrieval?

Accepted Answer

Delta Lake tables with Z-ordering on the primary key.. Delta tables are the standard for Databricks Feature Store, as they support efficient reads and writes, schema enforcement, and time travel. When dealing with high-dimensional sparse data, Delta's columnar storage format allows the engine to only read the necessary partitions and columns, significantly reducing I/O. This is critical for performance-sensitive feature engineering pipelines, as it allows rapid processing of large-scale datasets.

Answer

Parquet files on DBFS without using Delta Lake.

Answer

Storing the feature data as a series of JSON blobs in a single file.

Answer

Using a local SQLite database on the Spark driver node.

Question 17

Which TWO of the following are valid ways to register a model in the Databricks Model Registry?

Accepted Answer

Pass `registered_model_name='my_model'` to `mlflow.sklearn.log_model()`.. Registering models is a core step in the MLOps lifecycle. You can either register directly during the training flow (using log_model) or after the fact using the UI or the Python API. Understanding these methods is vital for automating the transition from development to staging and production, ensuring that all models undergo the necessary governance and versioning processes required by an enterprise Databricks environment.

Answer

Upload the model pickle file directly to the Model Registry's SQL database.

Answer

Manually copy the model code into the Databricks Model Registry UI.

Answer

Use the Spark `saveAsTable` command on the model object.

Question 18

When designing a custom MLflow model flavor for a proprietary framework, what is the most critical method to implement to ensure compatibility with `mlflow.pyfunc`?

Accepted Answer

The `load_model` method to enable loading the model artifact.. The `load_model` method is the entry point for the `pyfunc` flavor. It defines how the model is reconstructed from serialized files, which is necessary for serving. A well-implemented `load_model` method ensures that the model can be loaded by any deployment environment that supports the generic `pyfunc` interface, providing flexibility in model deployment across different platforms and runtimes.

Answer

The `train_model` method to support online learning.

Answer

The `register_model` method to support registry integration.

Answer

The `log_model` method to automatically record metrics.

Question 19

Refer to the exhibit. You are attempting to log a model to the MLflow Model Registry, but receive the serialization error shown. What is the cause of this error?

Accepted Answer

MLflow cannot resolve the absolute DBFS path during the serialization process.. The error occurs because the model object contains hardcoded references to absolute file paths that are not accessible on all worker nodes or serving endpoints. MLflow expects model artifacts to be self-contained or stored in a managed artifact repository. Referencing absolute paths creates a dependency on a specific storage mount that may not exist in the production environment.

Answer

The model file size exceeds the limit for the DBFS workspace.

Answer

The DBFS mount is currently in read-only mode during the training job.

Answer

The MLflow provider is not configured to use DBFS as the artifact store.

Question 20

Refer to the exhibit. You have multiple experimental runs in your MLflow tracking server. You want to retrieve the best-performing model based on accuracy. Which command should you use?

Accepted Answer

mlflow.search_runs(experiment_ids=['1'], order_by=['metrics.accuracy DESC']).iloc[0]. Using the MLflow client's search_runs method with an order_by parameter is the standard way to retrieve experiment results programmatically. This approach is superior to manual inspection, as it allows for automated pipeline selection of the top-performing model, ensuring that the model version promoted to the registry is objectively the most accurate according to the defined metrics.

Answer

mlflow.get_all_runs(experiment_id='1').sort_by('accuracy')

Answer

mlflow.list_experiments().get_best_run('accuracy')

Answer

mlflow.get_run(run_id='training_run_123').accuracy