I85PREP
⌘ K
MS MLO study materials
MS MLO study resources

MS MLO practice questions

One question at a time. Build your MS MLO understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Design and implement ML solutions

In Azure Machine Learning, which workspace artifact stores a reusable, versioned snapshot of a dataset definition so training pipelines can reference it reproducibly?

  1. An environment
  2. A registered data asset Correct answer.
  3. A compute cluster
  4. An endpoint

Explanation

A registered data asset captures a versioned reference to data (a URI or a definition) inside the workspace, letting pipelines pin a specific version for reproducibility. Compute clusters provide execution capacity, environments capture runtime dependencies, and endpoints serve models.

Question 2Design and implement ML solutions

A team wants each training run to record its metrics, parameters, and output artifacts automatically in Azure Machine Learning. Which approach achieves this with minimal custom code?

  1. Use MLflow tracking integrated with the workspace during the job Correct answer.
  2. Enable diagnostic settings on the workspace
  3. Attach an Application Insights resource to the workspace
  4. Manually copy logs into a blob container after each run

Explanation

Azure Machine Learning integrates with MLflow, so calling mlflow.log_metric, log_param, and log_artifact inside a job automatically records run data in the workspace. Manual copying is error-prone, and diagnostic settings and Application Insights cover telemetry rather than experiment tracking.

Question 3Design and implement ML solutions

Which Azure Machine Learning compute type is designed for interactive development in notebooks with automatic scale-down after idle time?

  1. Kubernetes-attached compute
  2. Compute cluster
  3. Compute instance Correct answer.
  4. Serverless compute

Explanation

A compute instance is a single managed VM intended for interactive authoring such as notebooks, and it can be configured to shut down automatically when idle. Clusters, serverless, and Kubernetes targets are oriented toward running jobs or deployments rather than interactive sessions.

Question 4Design and implement ML solutions

You need to deploy a model so it can be scored synchronously with low latency for a web application. Which Azure Machine Learning deployment type fits?

  1. Batch endpoint
  2. Online endpoint Correct answer.
  3. Pipeline job
  4. Data drift monitor

Explanation

Online endpoints expose a model behind a REST API for real-time, low-latency synchronous scoring. Batch endpoints process large files asynchronously, pipeline jobs orchestrate training steps, and drift monitors observe data rather than serve predictions.

Question 5Design and implement ML solutions

Which component of an Azure Machine Learning environment primarily controls the installed Python packages and their versions for a training job?

  1. The compute cluster SKU
  2. The conda or pip dependency specification Correct answer.
  3. The registered model version
  4. The datastore

Explanation

An environment's dependency specification (conda or pip) determines which Python packages and versions are installed in the job container. Compute SKU affects hardware, datastores locate data, and model versions are outputs rather than runtime dependencies.

Question 6Design and implement ML solutions

A pipeline must run a data-preprocessing step only after the data-validation step succeeds. Which Azure Machine Learning pipeline concept expresses this ordering?

  1. A datastore
  2. An environment version
  3. A pipeline component dependency Correct answer.
  4. A model registration

Explanation

Components in a pipeline declare inputs that reference other components' outputs, which creates explicit dependencies and enforces execution order. Datastores, environments, and model registrations are supporting resources, not ordering constructs.

Question 7Design and implement ML solutions

Which Azure Machine Learning feature lets a data scientist compare metrics across multiple training runs and select the best-performing run?

  1. The run history and experiment view Correct answer.
  2. The compute quota page
  3. The datastore explorer
  4. The access control blade

Explanation

The experiment view aggregates runs and their logged metrics so they can be compared and one selected. The datastore explorer browses storage, the quota page manages capacity, and access control handles authorization.

Question 8Design and implement ML solutions

What is the primary purpose of registering a trained model in an Azure Machine Learning workspace registry?

  1. To retrain the model automatically on a schedule
  2. To convert the model to ONNX format
  3. To allocate GPU capacity for inference
  4. To provide a versioned, named artifact that deployments can reference Correct answer.

Explanation

Model registration assigns a name and version to a model artifact so it can be tracked, audited, and referenced by deployments. It does not schedule retraining, convert formats, or provision compute.

Question 9Design and implement ML solutions

A training script must read a registered data asset regardless of whether the underlying storage is Azure Blob or Data Lake Storage. Which argument type should the script accept to keep it storage-agnostic?

  1. A mounted or downloaded input path passed as an argument Correct answer.
  2. A storage account connection string
  3. A hard-coded HTTPS URL with a SAS token
  4. An mlclient datastore object

Explanation

Passing an input as a path argument lets Azure Machine Learning resolve the registered asset to a mounted or downloaded local path, so the script stays independent of the storage backend and credentials. Hard-coded URLs, SDK objects, and connection strings couple the code to specific storage details.

Question 10Design and implement ML solutions

You want to reuse the same preprocessing logic across several pipelines while keeping its code, environment, and interface consistent. What should you create?

  1. A notebook cell copied into each pipeline
  2. A registered component with defined inputs and outputs Correct answer.
  3. A batch endpoint
  4. A separate workspace per pipeline

Explanation

A registered component packages code, environment, and a typed input/output interface so it can be reused consistently across pipelines. Copying cells duplicates logic, separate workspaces fragment assets, and batch endpoints serve inference.

Question 11Design and implement ML solutions

A pipeline job fails because a dependency library version differs from local testing. Which change most directly makes the job reproducible?

  1. Increase the compute cluster node count
  2. Switch the job to a compute instance
  3. Register the model before rerunning
  4. Pin the library version in the environment definition and reference that environment in the job Correct answer.

Explanation

Pinning the dependency in the environment definition and referencing that environment ensures the job container installs the same version every run. Scaling nodes, changing to a compute instance, or registering a model does not control installed package versions.

Question 12Design and implement ML solutions

A scheduled pipeline should retrain weekly but must skip execution when the input data asset version has not changed. Which design supports this?

  1. Set the schedule to run every minute
  2. Use a batch endpoint for training
  3. Add a condition step that compares the current data version and gates downstream training Correct answer.
  4. Disable run history logging

Explanation

A gating condition that inspects the current data asset version lets the pipeline short-circuit training when nothing changed. Very frequent schedules waste resources, disabling logging removes auditability, and batch endpoints are for inference.

Question 13Design and implement ML solutions

You must deploy a model that requires a custom scoring script to transform raw JSON into the model's expected feature vector. Which deployment configuration element holds that logic?

  1. The scoring script referenced by the deployment Correct answer.
  2. The environment's base Docker image
  3. The endpoint's authentication mode
  4. The model's registration description

Explanation

The scoring script defines init and run functions that preprocess incoming requests and invoke the model, so custom transformation belongs there. The base image supplies the runtime, authentication controls access, and the description is metadata.

Question 14Design and implement ML solutions

A team needs to detect when production input feature distributions diverge from the training data and trigger alerts. Which Azure Machine Learning capability addresses this?

  1. Workspace private link configuration
  2. Compute cluster autoscaling
  3. Pipeline component caching
  4. Model data drift monitoring on a deployed endpoint Correct answer.

Explanation

Data drift monitoring compares production input distributions against a baseline and can raise alerts when divergence exceeds thresholds. Autoscaling, caching, and private links address capacity, performance, and networking rather than distribution shift.

Question 15Design and implement ML solutions

You need to run a hyperparameter sweep that stops poorly performing trials early to save cost. Which configuration should you use?

  1. A single job with a fixed learning rate
  2. A batch endpoint with a large instance count
  3. A registered component without parameters
  4. A sweep job with an early-termination policy Correct answer.

Explanation

A sweep job explores a search space and an early-termination policy such as bandit or median stopping ends underperforming trials before completion. A fixed single job does no search, batch endpoints serve inference, and a parameterless component cannot vary hyperparameters.

Question 16Design and implement ML solutions

A model must be retrained whenever a new version of a registered data asset appears, without a fixed calendar schedule. Which trigger is most appropriate?

  1. A data-asset version change trigger Correct answer.
  2. A model registration event
  3. A manual notebook run
  4. A cron schedule every hour

Explanation

A trigger keyed to data asset version changes fires when new data is registered, matching event-driven retraining. Hourly cron ignores whether data changed, manual runs are not automated, and model registration is an output rather than an input event.

Question 17Design and implement ML solutions

You want to compare two model versions fairly by evaluating both on the same held-out registered data asset with identical preprocessing. What should you do?

  1. Evaluate each with a different random split and average results
  2. Compare training loss from each run's logs
  3. Evaluate only on the training data
  4. Reference the same registered test asset and the same component version in both evaluation jobs Correct answer.

Explanation

Using one registered test asset and one evaluation component version ensures both models see identical data and preprocessing, isolating model differences. Different splits introduce variance, training data leaks the answer, and training loss is not a held-out comparison.

Question 18Design and implement ML solutions

An online endpoint must handle variable traffic without manual intervention while limiting cost during quiet periods. Which setting supports this?

  1. Autoscaling on the deployment with a minimum and maximum instance count Correct answer.
  2. A batch endpoint with a long timeout
  3. Disabling authentication
  4. A fixed instance count equal to peak traffic

Explanation

Autoscaling adjusts instance count between configured minimum and maximum bounds based on load, handling peaks while scaling down when idle. A fixed peak count wastes money, disabling authentication weakens security, and batch endpoints do not serve synchronous traffic.

Question 19Design and implement ML solutions

Which Azure Machine Learning construct best isolates secrets such as storage keys used by a training job?

  1. Storing the key in a workspace key vault and referencing it via a secret in the job Correct answer.
  2. Embedding the key in the environment YAML
  3. Committing the key to the Git repository
  4. Hard-coding the key in the training script

Explanation

The workspace's linked key vault stores secrets, and jobs can retrieve them at runtime through secret references without exposing values in code or configuration. Hard-coding, committing, or embedding keys in environment files leaks credentials into version control and logs.

Question 20Design and implement ML solutions

A pipeline component produces a large intermediate dataset consumed by the next component. Which design avoids unnecessary copying between steps?

  1. Email the dataset to the next team
  2. Write the output to a mounted output path so the next step reads it in place Correct answer.
  3. Store the dataset in the component's Docker image
  4. Serialize the dataset into the component's log output

Explanation

Using a mounted output path lets the downstream component read the intermediate data directly from storage rather than duplicating it. Logging data, emailing it, or baking it into an image is inefficient and impractical for large datasets.

Question 21Design and implement ML solutions

A team wants to reuse a trained model for batch scoring of millions of records nightly. Which deployment choice is most appropriate?

  1. A compute instance notebook
  2. An online endpoint with a single instance
  3. A batch endpoint with a scoring script Correct answer.
  4. A hyperparameter sweep job

Explanation

A batch endpoint is built for asynchronous, high-volume scoring of large datasets on a schedule. Online endpoints target low-latency single requests, compute instances are for authoring, and sweep jobs tune hyperparameters.

Question 22Design and implement ML solutions

A model registered in the workspace must be promoted to production only after it beats the current production model on a registered evaluation dataset. Which mechanism enforces this gate?

  1. Enabling data drift monitoring on the endpoint
  2. Setting the endpoint to autoscale
  3. A pipeline step that compares metrics and conditionally registers or deploys the new version Correct answer.
  4. Manually deploying every trained model

Explanation

A conditional pipeline step evaluates the candidate against the production baseline and only registers or deploys it when the metric threshold is met, creating an automated promotion gate. Manual deployment, autoscaling, and drift monitoring do not enforce quality gates.

Question 23Design and implement ML solutions

A pipeline run intermittently fails only on multi-node clusters when two nodes write to the same output location. What is the most likely design flaw?

  1. The output path is shared across parallel nodes instead of being partitioned per node Correct answer.
  2. The schedule uses a cron expression
  3. The environment uses a pinned library version
  4. The endpoint uses key-based authentication

Explanation

Parallel nodes writing to a single shared output location cause race conditions and intermittent corruption; partitioning output per node or using a merge step resolves it. Pinned versions, authentication mode, and cron syntax do not cause concurrent write conflicts.

Question 24Design and implement ML solutions

You need to serve a model that must access a database using credentials that rotate every 30 days without redeploying. Which design is best?

  1. Have the scoring script retrieve the credential from the workspace key vault at runtime Correct answer.
  2. Store the credential in the endpoint's environment variables permanently
  3. Hard-code the credential in the scoring script
  4. Bake the current credential into the deployment image

Explanation

Retrieving the credential from the key vault at runtime lets rotation happen centrally without rebuilding or redeploying the image. Baking credentials into images, permanent environment variables, or scripts requires redeployment and risks exposure.

Question 25Design and implement ML solutions

A team observes that an online endpoint's latency rises sharply under load while CPU stays low. Which analysis is most plausible before changing the model?

  1. The model is overfitting the training data
  2. The environment pins an older library
  3. The deployment is under-provisioned on instances so requests queue Correct answer.
  4. The registered data asset has a new version

Explanation

Low CPU with rising latency suggests requests are queuing because too few instances handle the concurrency, pointing to capacity rather than model quality. Overfitting affects accuracy, a new data version affects inputs, and pinned libraries affect reproducibility, none of which explain idle CPU with high latency.

Question 26Design and implement ML solutions

A data scientist needs to create an Azure Machine Learning workspace that supports managed network isolation with private endpoints for the storage account. Inbound workspace access must also use a private endpoint, with public access disabled. Which configuration should they choose when provisioning the workspace?

  1. A workspace with a public endpoint and a managed virtual network.
  2. A workspace with a private endpoint and a managed virtual network. Correct answer.
  3. A workspace with a private endpoint and no managed virtual network.
  4. A workspace with a public endpoint and no managed virtual network.

Explanation

The workspace private endpoint provides private inbound connectivity. The managed virtual network separately supports managed outbound isolation and storage private endpoints. Meeting both requirements also requires disabling public workspace access and configuring the appropriate isolation mode and storage private-endpoint rules.

Question 27Design and implement ML solutions

A machine learning team wants to track the lineage of data used in experiments. They need to ensure that every time a dataset is modified, a new version is created and the previous versions remain immutable. Which Azure Machine Learning data asset type should they use?

  1. Data asset (v2) with versioning enabled. Correct answer.
  2. File dataset (v1) with versioning enabled.
  3. Datastore with versioning enabled.
  4. Data asset (v1) with versioning enabled.

Explanation

Data assets (v2) support versioning, where each modification creates a new immutable version, preserving lineage. Datastores do not support versioning of data; they are connection information to storage. Data assets (v1) and file datasets (v1) are legacy types that do not provide the same versioning capabilities as v2 data assets.

Question 28Design and implement ML solutions

A developer is writing an Azure Machine Learning pipeline and needs to pass a small configuration value (such as a random seed) to a component at runtime. Which approach is most appropriate?

  1. Define the value as a pipeline parameter and reference it in the component. Correct answer.
  2. Create a new data asset for the value and register it.
  3. Store the value in a file in the default datastore and read it from the component.
  4. Hardcode the value in the component's source code.

Explanation

Pipeline parameters are designed to pass small values like hyperparameters or configuration settings to components at runtime. They are easy to change per run and are tracked in the job. Storing in a file or data asset is overkill for a single value, and hardcoding reduces flexibility and reproducibility.

Question 29Design and implement ML solutions

A team needs to ensure that their Azure Machine Learning environment is reproducible across different compute targets. They want to specify Python packages and their versions explicitly. Which file should they use to define the environment?

  1. A conda.yaml file. Correct answer.
  2. A requirements.txt file.
  3. A Dockerfile.
  4. An environment.yml file.

Explanation

Azure Machine Learning uses conda.yaml files to define environments, specifying Python packages and versions. While requirements.txt and environment.yml can list packages, conda.yaml is the native format for Azure ML environments and supports both conda and pip dependencies. A Dockerfile is used for custom images but is not the primary way to define package dependencies for reproducibility in Azure ML.

Question 30Design and implement ML solutions

A data engineer is creating a pipeline that includes a component to train a model. They want to ensure that the component can access a registered data asset. What should they do when defining the component?

  1. Pass the data asset as an input to the component. Correct answer.
  2. Hardcode the data asset's path in the component script.
  3. Download the data asset to the default datastore before running the pipeline.
  4. Mount the data asset to the compute instance manually.

Explanation

In Azure Machine Learning pipelines, data assets are passed as inputs to components, allowing the component to access the data at runtime. This approach ensures proper tracking and versioning. Hardcoding paths or manually mounting data is not scalable or tracked, and downloading to the default datastore is unnecessary and inefficient.

Question 31Design and implement ML solutions

A machine learning engineer is building a pipeline that must run a data validation step and then a training step. They want to ensure that the training step only runs if the validation step succeeds. How should they configure the pipeline?

  1. Use a condition in the pipeline to check the validation step's output.
  2. Set the validation step's continue_on_step_failure property to False.
  3. Set the training step's continue_on_step_failure property to False.
  4. Set the training step's depends_on property to the validation step. Correct answer.

Explanation

In Azure Machine Learning pipelines, the depends_on property establishes a dependency, ensuring that the training step runs only after the validation step completes successfully. By default, if a step fails, subsequent steps do not run. Setting continue_on_step_failure to False on the validation step would cause the pipeline to stop on failure, but depends_on is the direct way to enforce order and success.

Question 32Design and implement ML solutions

A data scientist wants to use the Azure Machine Learning designer to build a regression model. They need to split the data into training and testing sets. Which module should they use?

  1. Split Data Correct answer.
  2. Partition and Sample
  3. Score Model
  4. Tune Model Hyperparameters

Explanation

The Split Data module in Azure Machine Learning designer is specifically designed to divide a dataset into two parts, typically for training and testing. Partition and Sample is used for sampling and partitioning but is not the standard module for train-test split in the designer. Tune Model Hyperparameters and Score Model serve different purposes.

Question 33Design and implement ML solutions

A team is using Azure Machine Learning compute clusters for training. They want to minimize cost by automatically scaling down to zero nodes when no jobs are running. Which setting should they configure?

  1. Set the idle seconds before scale down to 0.
  2. Enable low priority nodes.
  3. Set the minimum nodes to 0. Correct answer.
  4. Set the maximum nodes to 0.

Explanation

Setting the minimum nodes to 0 allows the compute cluster to scale down to zero nodes when idle, minimizing cost. The maximum nodes setting controls the upper limit, not the scale-down behavior. Idle seconds before scale down determines how long to wait before scaling down, but setting it to 0 might cause premature scale down. Low priority nodes reduce cost but do not automatically scale to zero.

Question 34Design and implement ML solutions

A machine learning engineer needs to deploy a model to an Azure Kubernetes Service (AKS) cluster for online inference. They want to enable authentication for the scoring endpoint. Which authentication method should they use?

  1. Key-based authentication
  2. No authentication
  3. Azure Active Directory authentication
  4. Token-based authentication Correct answer.

Explanation

For online endpoints on AKS, Azure Machine Learning supports key-based and token-based authentication. Token-based authentication uses Azure Active Directory to obtain a token, providing more secure and granular access control. Key-based authentication uses static keys, which are less secure. Azure AD authentication is not a direct option for the endpoint itself; token-based leverages Azure AD.

Question 35Design and implement ML solutions

A data scientist is using automated machine learning (AutoML) in Azure Machine Learning. They want to ensure that the best model is selected based on a specific metric. Which configuration should they set?

  1. The primary metric in the AutoML configuration. Correct answer.
  2. The training budget in the AutoML configuration.
  3. The validation dataset in the AutoML configuration.
  4. The featurization mode in the AutoML configuration.

Explanation

In AutoML, the primary metric determines which model is considered best. By setting the primary metric, the data scientist controls the optimization objective. The training budget limits time, validation dataset is used for evaluation, and featurization mode controls preprocessing, but none of these select the best model based on a metric.

Question 36Design and implement ML solutions

A team is using Azure Machine Learning to train a model on a large dataset stored in Azure Data Lake Storage Gen2. They want to minimize data movement and improve training performance. Which storage option should they use for the compute cluster?

  1. Copy the data to the default blob storage.
  2. Mount the data lake storage to the compute cluster. Correct answer.
  3. Use a datastore to reference the data lake storage.
  4. Download the data to the compute cluster's local disk.

Explanation

Mounting Azure Data Lake Storage Gen2 to the compute cluster allows direct access to the data without copying, minimizing data movement and improving performance. Downloading to local disk incurs transfer time and may not fit. Using a datastore is for referencing, but mounting provides the performance benefit. Copying to blob storage adds unnecessary data movement.

Question 37Design and implement ML solutions

A machine learning engineer is designing a pipeline that includes a step to train a model and a step to evaluate it. They want to ensure that the evaluation step runs even if the training step fails, to capture failure metrics. Which property should they set on the evaluation step?

  1. depends_on = training step
  2. continue_on_step_failure = False
  3. enabled = False
  4. continue_on_step_failure = True Correct answer.

Explanation

Setting continue_on_step_failure to True on the evaluation step allows it to run even if the training step fails, enabling capture of failure metrics. depends_on ensures order but does not control failure behavior. Setting continue_on_step_failure to False would prevent the evaluation step from running on failure. enabled=False would disable the step entirely.

Question 38Design and implement ML solutions

A data scientist is using Azure Machine Learning to run a hyperparameter tuning experiment. They want to use Bayesian sampling. Which early termination policy is compatible with Bayesian sampling?

  1. Truncation selection policy
  2. Median stopping policy
  3. Bandit policy
  4. No early termination policy Correct answer.

Explanation

Bayesian sampling in Azure Machine Learning does not support early termination policies because it requires completing all runs to build a probabilistic model. Therefore, no early termination policy can be used. Median stopping, truncation selection, and bandit policies are incompatible with Bayesian sampling.

Question 39Design and implement ML solutions

A team is deploying a model to an Azure Machine Learning online endpoint. They want to enable automatic scaling based on CPU usage. Which setting should they configure?

  1. Set the scale_type to 'Default' and let Azure manage scaling.
  2. Set the scale_type to 'Manual' and specify the number of instances.
  3. Set the scale_type to 'TargetUtilization' and provide a target utilization percentage. Correct answer.
  4. Set the scale_type to 'Auto' and provide a target utilization percentage.

Explanation

For automatic scaling based on CPU usage, the scale_type should be set to 'TargetUtilization' with a target utilization percentage. This configures the endpoint to scale instances to maintain the target CPU utilization. Manual scaling requires specifying a fixed number of instances. 'Default' and 'Auto' are not valid scale types for Azure ML online endpoints.

Question 40Design and implement ML solutions

A machine learning engineer needs to ensure that a model deployed to an online endpoint can be updated without downtime. Which deployment strategy should they use?

  1. Shadow deployment
  2. Canary deployment Correct answer.
  3. Blue-green deployment
  4. Rolling deployment

Explanation

Canary deployment allows routing a small portion of traffic to the new model version while the majority goes to the existing version, enabling testing and gradual rollout without downtime. Blue-green requires switching all traffic at once, which can cause downtime if not managed carefully. Rolling deployment updates instances in batches but may still cause brief downtime. Shadow deployment mirrors traffic but does not serve the new model to users.

Question 41Data engineering and feature management

In Azure Machine Learning, which asset type is used to store a pointer to a specific file or folder in a datastore so that pipelines can consume it by name without hardcoding paths?

  1. A compute cluster definition
  2. A model registration entry
  3. A registered Data asset (URI file or URI folder) Correct answer.
  4. An environment specification

Explanation

Data assets register a URI that points at a file or folder in a datastore, giving pipelines a stable named reference instead of a hardcoded path. Compute, environments, and models are separate asset types that do not resolve to storage locations.

Question 42Data engineering and feature management

A data scientist wants a versioned, reusable snapshot of a table stored in the workspace's default blob store so that training jobs read identical rows each run. Which approach fits best?

  1. Copy the table into each job's output folder at runtime
  2. Register the table's folder as a versioned URI folder data asset Correct answer.
  3. Point the job at a live SQL query that re-runs each time
  4. Reference the storage account key directly inside the script

Explanation

Registering the folder as a data asset with explicit versions pins the exact data location so every run reads the same snapshot. Re-querying live sources or copying at runtime introduces drift and duplicated logic.

Question 43Data engineering and feature management

Which datastore type in Azure Machine Learning connects a workspace to an Azure Data Lake Storage Gen2 container using either account key or credential-based authentication?

  1. Azure Blob datastore
  2. Azure Data Lake Gen2 datastore Correct answer.
  3. Azure File share datastore
  4. Azure SQL database datastore

Explanation

The Azure Data Lake Gen2 datastore type is specifically designed to reference ADLS Gen2 storage with hierarchical namespace and service-principal or key auth. Blob, file share, and SQL datastores target different storage services.

Question 44Data engineering and feature management

You need a feature that computes the 7-day rolling mean of transaction amount per customer and must be identical during training and online scoring. What is the most reliable design?

  1. Define the transformation once in a shared feature pipeline and reuse it for scoring Correct answer.
  2. Let the model learn the rolling mean from raw amounts
  3. Recompute the mean in the scoring service with separate SQL
  4. Compute the rolling mean in the training notebook only

Explanation

Training-serving skew is avoided when the same transformation logic is authored once and applied in both the batch training path and the online scoring path. Duplicating the logic in separate codebases invites divergence in windowing and null handling.

Question 45Data engineering and feature management

Which Azure Machine Learning feature lets you define a time-windowed transformation over a source table and materialize it into a feature store entity for reuse across teams?

  1. An inference endpoint traffic rule
  2. A feature set specification with a transformation and source Correct answer.
  3. A registered model
  4. A compute instance startup script

Explanation

A feature set specification declares the source data, transformation code, and window so the feature can be materialized and reused. Models, compute startup scripts, and traffic rules do not define feature transformations.

Question 46Data engineering and feature management

A team notices that a categorical feature has 40 percent unseen categories at inference because the encoder was fit on a small training sample. Which change most directly addresses this?

  1. Switch the endpoint to batch scoring
  2. Increase the model's learning rate
  3. Fit the encoder on the full historical category distribution and map unknowns to an explicit bucket Correct answer.
  4. Drop the feature from the model entirely

Explanation

Fitting the encoder on a representative category vocabulary and routing unseen values to a dedicated bucket keeps encoding stable at inference. Hyperparameters, feature removal, and scoring mode do not fix vocabulary coverage.

Question 47Data engineering and feature management

Which statement correctly describes how point-in-time correctness prevents label leakage in a feature pipeline?

  1. Features for a training row must only use data with timestamps at or before the label event time Correct answer.
  2. Features for a training row may use any data available at training time
  3. Features must be recomputed on every scoring request
  4. Labels must be recomputed after features are materialized

Explanation

Point-in-time joins restrict each training row's features to values that existed before the label event, which mirrors what would be known at inference. Using future data leaks the outcome into the inputs and inflates offline metrics.

Question 48Data engineering and feature management

In a medallion-style lakehouse, which layer typically holds cleansed, conformed, and query-optimized tables consumed by feature engineering?

  1. Silver
  2. Bronze
  3. Landing
  4. Gold Correct answer.

Explanation

The gold layer holds business-level, conformed, and optimized tables that downstream consumers such as feature pipelines read. Bronze and landing hold raw ingestion data, while silver holds cleaned but not fully modeled data.

Question 49Data engineering and feature management

A pipeline must train on data that changes daily but must be reproducible for audits. Which combination best satisfies both needs?

  1. A single mutable folder path read by every run
  2. Manual notebook execution by an analyst each day
  3. A live dashboard query feeding the training script
  4. A fixed registered data asset version pinned in the pipeline plus daily new versions Correct answer.

Explanation

Pinning an explicit data asset version in the pipeline records exactly which data produced each model, while new daily versions capture fresh data. Mutable paths and manual runs break auditability.

Question 50Data engineering and feature management

Which practice most reduces the risk of schema drift breaking a scheduled feature materialization job?

  1. Disabling job alerts to reduce noise
  2. Storing raw data without any schema checks
  3. Validating incoming schema against an expected contract and failing fast with clear errors Correct answer.
  4. Increasing the job's retry count to ten

Explanation

An explicit schema contract catches added, removed, or retyped columns before they corrupt materialized features, and failing fast surfaces the problem clearly. Retries and alert suppression hide rather than detect drift.