Databricks-ML-Assoc Model Development Practice Test 2 — 15 Questions

Question 1

A data scientist is building a custom model that uses a pre-trained external library. Which THREE requirements are essential for correctly packaging this model using the MLflow 'pyfunc' flavor?

Accepted Answer

The custom model must be defined as a class inheriting from mlflow.pyfunc.PythonModel.. Packaging a custom pyfunc model requires a class that inherits from 'PythonModel'. This class must implement the 'predict' method, which is the entry point for inference. Furthermore, the external library must be specified in the environment configuration, and all necessary artifacts (like model weights or configuration files) must be passed to the log_model function to ensure the code has access to these dependencies during execution.

Answer

The predict method must be defined as a static method to avoid serialization issues.

Answer

The custom model must be serialized using pickle.dump before logging to the model registry.

Question 2

A machine learning engineer needs to log a model that was trained outside of Databricks but needs to be managed within the Databricks Model Registry. What is the recommended workflow to achieve this?

Accepted Answer

Log the model using mlflow.pyfunc.log_model, providing the path to the external artifacts.. The recommended approach is to use the MLflow log_model function locally or within a notebook, pointing to the external model files. By explicitly defining the model flavor and providing the required environment metadata, the model is registered as a first-class artifact. This centralizes the model lifecycle management, enabling versioning, staging, and deployment tracking within Databricks, even if the actual training compute was off-platform.

Answer

Manually copy the files into the Databricks DBFS and link them to a Model Registry entry via the UI.

Answer

Upload the model to Unity Catalog as a generic file and point to the location in the notebook.

Answer

Convert the model to a Spark DataFrame and store it as a Delta table in the Hive Metastore.

Question 3

A data scientist is using Databricks to train a model and wants to ensure that the model is automatically registered to the Model Registry if it meets specific accuracy thresholds. How should this be configured?

Accepted Answer

Use the 'registered_model_name' parameter within the log_model function.. The Model Registry provides a programmatic interface for managing model versions. By using the MLflow 'register_model_name' parameter in log_model, you can automate the registration process. This allows teams to create CI/CD pipelines where models are automatically evaluated against metrics, and only those that perform well are promoted to the registry, reducing manual effort and ensuring consistent quality in model deployment pipelines.

Answer

Update the Model Registry manually via the Databricks UI after the run is completed.

Answer

Create a separate job in Databricks Workflows that scans the experiment and registers models.

Answer

Enable the 'Auto-Register' setting in the Databricks cluster configuration.

Question 4

A team is developing a machine learning pipeline where multiple data scientists contribute to the same experiment. What is the most effective way to organize these runs?

Accepted Answer

Create separate MLflow experiments for each developer or logical model task.. Using unique experiment names is crucial for preventing data collisions. When multiple developers work on a project, they should either use distinct experiments or leverage the 'run_name' parameter within a shared experiment to identify their specific contributions. This ensures that metrics and artifacts are properly attributed, allowing for clear comparisons and preventing one user from accidentally overwriting or cluttering the results of another user.

Answer

Share a single experiment name and use tags to identify the developer.

Answer

Use the default experiment, but manually delete unsuccessful runs to keep it clean.

Answer

Require all team members to use the same run name for consistency.

Question 5

A data scientist is working in a notebook and notices that their MLflow tracking logs are not appearing in the expected experiment. What is the most likely cause?

Accepted Answer

The mlflow.set_experiment() function was not called with the correct path.. In Databricks, if the experiment context is not explicitly set, MLflow defaults to a specific path or the notebook-linked experiment. If the scientist is working in a workspace where multiple experiments exist, or if they accidentally used a different experiment name in a previous cell, the tracking will go to the wrong location. Checking the current experiment ID is essential for ensuring that logs are captured in the correct, organized location.

Answer

The cluster is running in 'Single User' mode, which disables MLflow logging.

Answer

The notebook has not been attached to a Unity Catalog-enabled compute resource.

Answer

The MLflow tracking server is down for the entire Databricks workspace.

Question 6

Which TWO of the following are benefits of using the Databricks Model Registry for managing model versions?

Accepted Answer

It provides a clear audit trail of model version history and transitions.. The Model Registry serves as the source of truth for model deployments. It allows teams to track the lineage of a model from the specific experiment that created it to the version history in the registry. It also provides a centralized location for managing staging, testing, and production transitions, which is vital for enabling collaborative development and ensuring that the correct model versions are being served in production environments.

Answer

It automatically retrains models when data drift is detected in production.

Answer

It replaces the need for MLflow experiment tracking.

Answer

It automatically encrypts the training data used to build the model.

Question 7

A machine learning engineer needs to track a custom scikit-learn model's hyperparameters and evaluation metrics during a distributed hyperparameter tuning run in Databricks. Which approach adheres to MLflow best practices for nested runs?

Accepted Answer

Instantiate a parent run using mlflow.start_run() before the optimization loop, and within each trial loop, invoke mlflow.start_run(nested=True) to log specific parameters and metrics.. MLflow nested runs organize hyperparameter tuning trials hierarchically under a parent run, isolating metrics per trial while aggregating overall optimization progress. This structure is crucial for large-scale Databricks workflows where managing hundreds of individual model artifacts efficiently prevents clutter and enhances experiment traceability.

Answer

Manually invoke mlflow.start_run() without the nested parameter inside each iteration, allowing MLflow to automatically infer parent-child relationships based on timestamp ordering.

Answer

Set the MLFLOW_NESTED_RUN environment variable inside the driver node before executing the training loop to force automatic nesting of child runs.

Answer

Log all trial metrics to a single global run using unique metric key suffixes, avoiding child runs entirely to minimize database overhead on the tracking server.

Question 8

A data scientist is preparing to train a deep learning model using MLflow autologging on a Databricks cluster. Which TWO statements accurately describe the behavior of MLflow autologging for frameworks like PyTorch or TensorFlow? (Select TWO)

Accepted Answer

Framework-specific metrics such as loss and accuracy, along with optimizer hyperparameters like learning rate and momentum, are automatically logged to the active MLflow run.. MLflow autologging simplifies experiment tracking by automatically capturing framework-specific artifacts, parameters, and metrics without requiring manual logging statements. Understanding these default behaviors ensures data scientists know precisely what metadata is preserved and where manual intervention remains necessary for complex custom pipelines.

Answer

Autologging automatically captures model checkpoints at the end of every epoch and registers them directly to the Databricks Model Registry as production-ready versions.

Answer

Autologging explicitly disables the creation of parent runs, requiring users to manually manage run nesting if they wish to execute hyperparameter sweeps alongside deep learning training.

Answer

Autologging overrides any custom logging statements defined later in the training script, preventing developers from logging custom business-specific metrics or tags.

Question 9

A data science team is developing a custom machine learning model on Databricks that does not have a native MLflow flavor. To ensure the model can be logged, tracked, and deployed successfully using MLflow, which TWO actions must the ML engineer perform? (Choose two.)

Accepted Answer

Create a subclass of mlflow.pyfunc.PythonModel and implement the predict method to handle input transformations and inference execution.. Custom models lacking native flavors require defining a custom Python model class inheriting from mlflow.pyfunc.PythonModel and implementing the predict context method. These steps are vital on Databricks because they wrap arbitrary preprocessing logic and non-standard libraries into a standardized scoring interface required for model registry deployment.

Answer

Register the custom model directly into the Delta Lake catalog using standard SQL ALTER TABLE statements combined with MLflow decorators.

Answer

Configure cluster-level Apache Spark configuration properties to serialize Python bytecode directly into distributed worker memory nodes.

Answer

Export the model weights into a plain text JSON file and upload them to DBFS using standard Databricks CLI file copy commands.

Question 10

A data scientist is optimizing a model's hyperparameters using Hyperopt. Which TWO steps are strictly necessary to effectively manage the search space and evaluate the objective function in a Databricks environment?

Accepted Answer

Define the search space using hyperopt.hp distributions.. Hyperopt requires a clearly defined search space and an objective function that returns the loss to be minimized. In Databricks, managing trials effectively often involves parallelizing the search using SparkTrials. Selecting the correct search space distributions and returning the correct loss metric are foundational, ensuring the optimization process converges on the best hyperparameter configuration efficiently while respecting the distributed nature of the cluster environment.

Answer

Manually restart the Spark cluster for every single trial performed.

Answer

Only use a sequential search algorithm to ensure repeatability.

Answer

Hard-code the hyperparameter values into the model training loop.

Question 11

A data scientist is training a Random Forest model on a 500GB dataset using Databricks. They notice that the training process is extremely slow and eventually crashes with an OutOfMemory error. Which approach should be prioritized to optimize the model training process?

Accepted Answer

Increase the number of partitions to enable better parallelization of the workload.. Optimizing memory usage in Spark-based machine learning requires balancing data partitioning and hardware resources. By increasing the number of partitions, the data scientist can ensure that the workload is distributed more evenly across the executor nodes, preventing individual executors from being overwhelmed by large data chunks. Proper configuration of memory settings and caching strategies are essential for scaling model development in a distributed environment like Databricks.

Answer

Increase the number of executors while keeping the partition count low.

Answer

Switch from Random Forest to a Gradient Boosted Tree model immediately.

Answer

Reduce the number of features in the dataset without performing any data repartitioning.

Question 12

A data scientist is using MLflow to track experiments in Databricks. They want to log both the model parameters and the resulting performance metrics for a complex pipeline. Which TWO actions should they take to ensure reproducible and organized tracking?

Accepted Answer

Call mlflow.set_experiment() at the start of the notebook to organize results.. Effective MLflow tracking requires logging both input parameters and output metrics to reconstruct experiment history. Using 'mlflow.log_param' allows for tracking the configuration, while 'mlflow.log_metric' captures the outcome. Integrating these within an 'mlflow.start_run' block ensures that all artifacts are associated with a specific, identifiable run, which is critical for model governance, comparison across different trials, and maintaining a clear audit trail of the development lifecycle.

Answer

Manually store parameters in a CSV file and upload them to DBFS.

Answer

Use the log_artifact method to record the model architecture as a raw string.

Answer

Disable the autologging feature to ensure the MLflow tracking server is not overloaded.

Question 13

Refer to the exhibit. The data scientist receives a TypeError when trying to log parameters in bulk. What is the cause of this error?

Accepted Answer

The mlflow module was incorrectly imported or overwritten in the current session.. The error occurs because the MLflow tracking API expects a dictionary for bulk parameter logging, but the syntax used in the exhibit implies a misunderstanding of how the API handles dictionaries. The method 'log_params' is correct, but users often confuse parameter names or nested structures, or they may be using an outdated version of the MLflow library where bulk operations were handled differently. Validating input types is crucial.

Answer

The log_params method does not accept dictionary objects as inputs.

Answer

The parameter keys in the dictionary are not strings.

Answer

The max_depth value must be a string instead of an integer.

Question 14

A data scientist needs to share their trained model with a colleague. The model is saved as an MLflow artifact in the workspace. What is the most effective way to share this model while maintaining full reproducibility?

Accepted Answer

Register the model in the MLflow Model Registry.. Using the MLflow Model Registry is the industry standard for sharing models. By registering a model, the scientist provides their colleague with a versioned, immutable artifact that includes the model's code, environment dependencies, and signature. This ensures that the colleague can load the exact same model in their own environment without encountering dependency conflicts or version mismatch errors, which is critical for collaborative data science.

Answer

Export the model as a pickle file and send it via email.

Answer

Save the model code to a Git repository and have the colleague retrain it.

Answer

Upload the model weights to a cloud storage bucket without metadata.

Question 15

A machine learning engineer is tuning a scikit-learn GradientBoostingClassifier on a 40 GB Delta table using Spark ML and Hyperopt inside a Databricks notebook. Each trial currently re-reads the Delta table from cloud storage and re-fits the model from scratch, and a 60-trial search is taking over nine hours. The engineer wants to reduce wall-clock time without changing the search space or the model family. Which change should the engineer make to the Hyperopt objective function to achieve this?

Accepted Answer

Load the Delta table once outside the objective function, convert it to a Pandas DataFrame with toPandas(), and reference that object inside the objective function.. The repeated cost in this search is the per-trial Delta scan and Spark job launch, not the model fitting itself. Reading the table a single time and holding it as an in-memory Pandas DataFrame for the objective function removes that repeated I/O across all trials while leaving the search space and algorithm untouched. Caching, broadcasting, or autologging do not address the per-trial data reload and refit pattern described.

Answer

Broadcast the Delta table into a Spark variable inside the objective function so every trial reads from executor memory.

Answer

Persist the training DataFrame with .cache() on the driver right before calling fmin, so trials reuse the materialized partitions.

Answer

Call mlflow.spark.autolog() before the search and let autolog cache the feature vectors between trials.