Databricks-ML-Assoc Practice Exam 3 — 20 Questions

Question 1

You are using Databricks Feature Store to create a feature table that will be used for both batch training and online inference. The feature table must be refreshed daily with new data, and the online store must serve the latest feature values within minutes of the refresh. Which configuration should you use?

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

Accepted Answer

Create a feature table with online=True and use a Databricks Job to run a daily write that updates both the offline Delta table and the online store.. A feature table created with online=True is backed by both an offline Delta table and an online store. Writing to it via FeatureStoreClient.write_table publishes the new feature values to both stores. Running this write in a daily Databricks Job refreshes the online store with the latest data, enabling low-latency serving within minutes. Other options either omit the online store, use unsupported manual exports, or fail to meet freshness requirements.

Answer

Create a feature table with online=False, and configure Model Serving to read from the offline table at inference time.

Answer

Create a standard Delta table and manually export it to a key-value store each day using a notebook.

Answer

Create a feature table with online=True, but only write to it weekly to reduce costs.

Question 2

You are deploying a model using MLflow, and you want to log the custom pre-processing logic alongside the model so it is automatically applied during inference. How should you achieve this?

Accepted Answer

Use the mlflow.pyfunc.log_model() function with a custom code wrapper.. MLflow allows you to log custom Python functions or entire pipelines as part of the model using the 'pyfunc' flavor. By wrapping your pre-processing steps and model into a custom pyfunc class and logging it, you encapsulate both the transformation and the prediction logic. This is a critical best practice for production deployments, as it ensures consistent pre-processing between training and inference, eliminating training-serving skew and simplifying the deployment process for the serving endpoint.

Answer

Create a separate inference job that runs the pre-processing code before the model.

Answer

Store the pre-processing logic in a shared Delta table and call it at inference.

Answer

Log the pre-processing parameters as tags in the MLflow run and re-apply them.

Question 3

When using MLflow to track experiments, what happens if you invoke mlflow.end_run() inside a nested loop when the parent run is already active?

Accepted Answer

The active run is terminated, and subsequent logging calls outside the loop will fail if they assume the run is active.. Calling end_run() within a nested loop terminates the current active run. If it is part of a parent-child structure, ending the child prematurely could lead to incomplete data or orphaned metrics. Proper management requires careful scoping, often using context managers (with statements), to ensure that the run lifecycle is handled cleanly and that all sub-runs correctly report back to the primary experiment tracking context.

Answer

The entire experiment is deleted from the workspace.

Answer

The parent run automatically restarts to compensate for the closed child run.

Answer

The nested loop continues, but logging is redirected to the default root run.

Question 4

A machine learning engineer is using MLflow Tracking on Databricks to compare multiple runs. They want to programmatically retrieve the best run based on a metric called 'rmse' from an experiment. Which MLflow API call should they use?

Accepted Answer

`mlflow.search_runs()` with `order_by=['metrics.rmse ASC']`. `mlflow.search_runs()` is designed to query runs within an experiment and can order results by metrics. By specifying `order_by=['metrics.rmse ASC']`, the run with the lowest RMSE appears first. This is efficient and programmatic, unlike manually iterating runs or using functions that retrieve single runs or experiment metadata.

Answer

`mlflow.get_experiment_by_name()` and then filter runs by metric

Answer

`mlflow.get_run()` with the run ID of the best run

Answer

`mlflow.list_run_infos()` and manually iterate to find the minimum 'rmse'

Question 5

A data scientist is using Hyperopt with SparkTrials on a Databricks cluster to tune a scikit-learn model. They set max_evals=100 and parallelism=4. After the tuning completes, they notice that some trials failed due to memory errors on the workers. What is the most likely cause of these failures?

Accepted Answer

Each trial is training on the full dataset, and the dataset is too large to fit in the memory of a single worker.. The correct answer is that each trial is training on the full dataset, and the dataset is too large to fit in the memory of a single worker. SparkTrials distributes trials, but each trial runs on one worker with the full dataset. If the data is large, it can cause out-of-memory errors. To mitigate, you can reduce the dataset size, increase worker memory, or use distributed algorithms. The other options misattribute the cause to model incompatibility, driver memory, or autoscaling.

Answer

The scikit-learn model is being trained on the driver, causing memory pressure on the driver instead of workers.

Answer

SparkTrials does not support scikit-learn models and should be used only with MLlib.

Answer

The cluster's autoscaling is reducing the number of workers, causing trials to be queued and eventually fail.

Question 6

Refer to the exhibit. Why is including the `signature` and `input_example` in the `log_model` call considered a professional best practice?

Accepted Answer

It enables schema validation and improves model serving usability.. Providing a signature and input example allows MLflow to define the expected schema for the model. This metadata facilitates automatic type validation and allows the model serving infrastructure to generate accurate API documentation. This is critical for downstream consumers who need to integrate the model, as it prevents runtime integration errors and clarifies the interface expectations before the model is even deployed into production.

Answer

It forces the model to run on a GPU during inference.

Answer

It speeds up the model training process.

Answer

It reduces the storage size of the model artifact.

Question 7

Refer to the exhibit. A machine learning engineer wants to promote this model to Production. Which MLflow action should be performed to achieve this while ensuring the model is ready for deployment?

Accepted Answer

Call the transition_model_version_stage API to move the model to Production.. The MLflow Model Registry provides the `transition_model_version_stage` function to promote models between stages like Staging and Production. Promoting a model involves updating its state in the registry, which then allows downstream deployment pipelines to identify the correct model version. This structured approach ensures that only verified models are promoted to production environments, maintaining governance and stability in the model deployment process.

Answer

Use the mlflow.register_model() function with the production flag.

Answer

Delete the model from Staging and re-upload it to the Production path.

Answer

Update the tag 'status' to 'Production' in the model metadata.

Question 8

Which TWO actions should be taken to ensure reproducibility of a Databricks ML model experiment?

Accepted Answer

Log the Git commit hash along with experiment metadata in MLflow.. Reproducibility in Databricks is achieved by capturing the state of the code, the environment, and the data. By utilizing Git integration and MLflow tracking, data scientists can point to the exact version of the code and environment used for an experiment. Using versioned Delta tables ensures that the data state can also be reconstructed, preventing 'data drift' from invalidating the results of historical experiments as underlying data tables change.

Answer

Hardcode the data file paths to local desktop directories.

Answer

Only log the model artifact and ignore the environment dependencies.

Answer

Manually delete all previous MLflow runs to keep the tracking server clean.

Question 9

An ML engineer deployed a model to a Databricks Model Serving endpoint and enabled inference tables. After a week, the team notices that some requests returned HTTP 200 but the corresponding rows in the inference table show null prediction values. They need to diagnose why predictions are missing for those requests. Which explanation is most consistent with this symptom?

Accepted Answer

The model returned a response shape that the serving layer could not map to the prediction column.. Inference tables capture request and response payloads for each served request, and the serving layer derives the prediction column from the model's response structure. When a model returns output in a format the logging layer does not recognize as a prediction, the request still succeeds with HTTP 200 but the logged prediction is null. Sampling, autoscaling, and read permissions produce different symptoms such as missing rows or errors, not present rows with null predictions.

Answer

The endpoint's autoscaling reduced replicas, causing dropped responses.

Answer

The endpoint's service principal lacked SELECT permission on the inference table.

Answer

The inference table was configured with a sampling fraction that excluded those rows.

Question 10

A data scientist has registered a scikit-learn model in Unity Catalog as `ml_team.prod.churn_model` and wants it served by a Databricks Model Serving endpoint that performs online inference for a web app. The endpoint must automatically pick up new model versions as they are promoted, without the data scientist editing the endpoint each time. Which approach should the data scientist use to configure the served entity?

Accepted Answer

Serve the model by name with the version set to the alias `champion`, so the endpoint resolves the alias to whatever version it currently points to.. Serving by model name with an alias makes the endpoint follow promotions automatically: when a new version is assigned the `champion` alias, subsequent requests resolve to that version. Pinning a numeric version, copying artifacts to DBFS, or polling with a job all require manual or scheduled intervention and cannot deliver the immediate, configuration-free update the scenario demands.

Answer

Create a Databricks job that runs every five minutes, downloads the latest model version, and calls the Model Serving REST API to overwrite the endpoint configuration with the new numeric version.

Answer

Serve the model by name with the version pinned to the numeric version that is currently in production, then re-create the endpoint whenever a new version is registered.

Answer

Export the MLflow model to a local directory, upload the artifacts to DBFS, and configure the endpoint to load the model from that DBFS path.

Question 11

Which approach is most efficient for tracking hyperparameter tuning metrics across thousands of runs on Databricks?

Accepted Answer

Using MLflow Experiments UI.. Using the MLflow Search API or the Experiments UI is the standard approach to manage and analyze large volumes of experiment results. These tools provide filtering, sorting, and visualization capabilities that allow data scientists to quickly identify the best performing models from thousands of candidates. This efficiency is necessary for iterative machine learning, where the ability to derive insights from vast amounts of experiment data drives faster and more accurate model development.

Answer

Logging results to a text file.

Answer

Printing metrics to the notebook logs.

Answer

Manually updating a spreadsheet.

Question 12

A model has been trained on a dataset containing categorical features with high cardinality. Which technique is most effective for preparing these features for a linear model in a Databricks environment?

Accepted Answer

Use Target Encoding to map categories to their mean target value.. Linear models require numeric inputs and are sensitive to the scale and representation of categorical variables. One-hot encoding high-cardinality features results in sparse, high-dimensional matrices that degrade performance. Target encoding or using specialized embeddings is more efficient. Choosing the right encoding strategy is a critical model development decision that directly impacts training speed, convergence, and the overall predictive accuracy of the final model deployed into production.

Answer

One-hot encode all categorical variables.

Answer

Drop the categorical features entirely.

Answer

Convert all features into a single string representation.

Question 13

A data scientist is working in a Databricks notebook and wants to use MLflow to log a trained scikit-learn model. They want to ensure that the model can be loaded later for inference. What is the correct MLflow function to log the model?

Accepted Answer

mlflow.sklearn.log_model. To log a scikit-learn model, use mlflow.sklearn.log_model. This function captures the model, its dependencies, and optionally a signature, making it easy to load later with mlflow.sklearn.load_model. It is the standard way to persist scikit-learn models in MLflow, ensuring reproducibility and compatibility with model serving.

Answer

mlflow.log_artifact

Answer

mlflow.log_model

Answer

mlflow.register_model

Question 14

A data scientist is using Databricks Feature Store to create a feature table for a recommendation model. They want to ensure that the feature table can be used for both training and batch scoring, and that it supports point-in-time correctness. Which two actions must they take when creating the feature table? (Choose two.)

Accepted Answer

Define a primary key for the feature table.. To create a feature table in Databricks Feature Store that supports training, batch scoring, and point-in-time correctness, the data scientist must define a primary key and a timestamp key. The primary key identifies entities for joins, and the timestamp key enables time-aware joins to prevent leakage. Other actions like enabling time travel manually, partitioning, or using streaming tables are not required and may be unnecessary or counterproductive. Thus, the correct choices are defining a primary key and specifying a timestamp key.

Answer

Set the table as a streaming table.

Answer

Partition the feature table by the primary key.

Answer

Enable time travel on the feature table.

Question 15

A data scientist is preparing a dataset for training a model on Databricks. They want to split the data into training and testing sets, and they need to ensure that the split is reproducible across different runs. Which PySpark method should they use to split the DataFrame with a fixed random seed?

Accepted Answer

df.randomSplit([0.8, 0.2], seed=42). randomSplit is the PySpark method designed to split a DataFrame into multiple parts based on weights, and it accepts a seed for reproducibility. This allows the data scientist to obtain consistent training and testing sets across runs. Other methods like sample, repartition, or cache do not provide a reproducible train-test split.

Answer

df.repartition(2)

Answer

df.sample(fraction=0.8, seed=42)

Answer

df.cache()

Question 16

A data scientist is using MLflow to track experiments on Databricks. They want to compare multiple runs and identify the best performing model based on a custom metric. Which TWO features of MLflow can be used to achieve this? (Choose two.)

Accepted Answer

MLflow tracking API's search_runs method. The MLflow UI's compare runs feature and the tracking API's search_runs method both allow you to compare runs and sort by custom metrics. The UI provides a visual side-by-side comparison, while search_runs enables programmatic querying and sorting. Together, they enable identifying the best performing run based on a custom metric.

Answer

MLflow Projects with a custom entry point

Answer

MLflow Model Registry's stage transitions

Answer

MLflow's automatic logging with autolog()

Question 17

A machine learning engineer has a registered model and wants to expose it as a REST endpoint that their application can call for real-time predictions. They need the endpoint to be created and managed natively within Databricks, with the ability to enable scale-to-zero during idle periods. Which Databricks capability should they use?

Accepted Answer

Databricks Model Serving, creating an endpoint that serves the registered model and configuring the served entity with scale-to-zero enabled.. Databricks Model Serving is the native mechanism for deploying registered models as REST endpoints with managed lifecycle, authentication, and optional scale-to-zero. A Flask process on a job cluster, registry webhooks, or a SQL warehouse cannot provide a stable, secure, autoscaling prediction API, so they fail the scenario's requirements for native management and idle shutdown.

Answer

MLflow Model Registry webhooks that notify an external service whenever a model version is registered, letting that service host the model.

Answer

A Databricks job that runs a Python script loading the model and starting a Flask web server on a driver node, exposed via a public URL.

Answer

A SQL warehouse with a user-defined function that loads the model artifact and returns predictions for incoming queries.

Question 18

A machine learning engineer is orchestrating a multi-step training workflow on Databricks using Databricks Jobs. The workflow includes data preprocessing, model training, and evaluation. The engineer needs to ensure that the evaluation step runs only if the training step completes successfully, and that the preprocessing step runs first. Which feature of Databricks Jobs should be used to define these dependencies?

Accepted Answer

Task dependencies using the depends_on field. Task dependencies via the depends_on field are the core mechanism in Databricks Jobs for orchestrating multi-task workflows. They allow you to specify that a task runs only after its upstream tasks complete successfully, enabling conditional execution. This directly satisfies the need to run preprocessing first, then training, and evaluation only if training succeeds. Other options address compute scaling, failure recovery, or version control, not workflow dependencies.

Answer

Repair run with task-level retries

Answer

Databricks Repos integration with Git

Answer

Job clusters with autoscaling enabled

Question 19

Refer to the exhibit. A user attempts to update a model stage in the Model Registry and receives this error. What is the most appropriate action to resolve this?

Accepted Answer

Request the workspace admin to update the model ACLs. The error clearly indicates a lack of permission to modify the model's state. In Databricks, access control lists (ACLs) are strictly enforced on objects in the registry. To perform administrative actions like changing a model's stage, the user must be assigned the appropriate permission level (CAN_MANAGE) by an owner or admin. This enforces the principle of least privilege, preventing unauthorized users from promoting unvetted models to production.

Answer

Re-register the model under a new name

Answer

Upgrade the model to a higher version

Answer

Switch to a different Databricks cluster

Question 20

Refer to the exhibit. A data scientist is logging a model to MLflow. Why is including an 'input_example' highly recommended in this specific code snippet?

Accepted Answer

It allows MLflow to validate the schema and ensures the model is compatible with deployment endpoints.. The input_example acts as a validation mechanism that MLflow uses to verify that the model can successfully process incoming data. It also enables automatic schema inference, which is crucial for downstream services like Model Serving. By providing this example, developers ensure that the model registry has metadata that allows for schema validation at serving time, reducing the risk of runtime errors when the model is deployed to production endpoints.

Answer

It forces the model to retrain every time it is loaded into production.

Answer

It automatically encrypts the model artifacts for security compliance.

Answer

It creates a dummy model version that cannot be used for inference.