Question 10
A data engineering group wants stable and secure storage access from Azure Machine Learning jobs. The immediate design goal is to let a pipeline write generated feature files back to Blob storage as well as read source blobs. Which approach addresses that goal?
- Datastore URI reference
- Storage Blob Data Contributor role
- Registered datastore
- Storage Blob Data Reader role
Explanation
Correct Answer
B:The data contributor role allows blob content modification at its assignment scope. It is broader than a read-only role and should be limited to locations where the pipeline writes.
Incorrect Answers
A:An Azure ML datastore URI identifies a path through registered storage connection metadata; the consumer still needs correct identity and network access to resolve the referenced files; the outstanding need is to let a pipeline write generated feature files back to Blob storage as well as read source blobs.
C:A datastore describes how Azure ML reaches an underlying storage service; its registration is connection metadata and does not duplicate the entire contents into the workspace; the needed capability is to let a pipeline write generated feature files back to Blob storage as well as read source blobs.
D:This data-plane role authorizes blob reads at the assigned scope; a generic Azure resource management role does not necessarily provide equivalent access to blob payloads; the requirement, by contrast, is to let a pipeline write generated feature files back to Blob storage as well as read source blobs.
Question 11
A data engineering group wants stable and secure storage access from Azure Machine Learning jobs. The team is required to reference a registered datastore path as job input so paths remain portable across job definitions. Which option is the best fit?
- Datastore URI reference
- Storage Blob Data Contributor role
- Registered datastore
- Credential rotation procedure
Explanation
Correct Answer
A:An Azure ML datastore URI identifies a path through registered storage connection metadata. The consumer still needs correct identity and network access to resolve the referenced files.
Incorrect Answers
B:The data contributor role allows blob content modification at its assignment scope; it is broader than a read-only role and should be limited to locations where the pipeline writes; the team would still need to reference a registered datastore path as job input so paths remain portable across job definitions.
C:A datastore describes how Azure ML reaches an underlying storage service; its registration is connection metadata and does not duplicate the entire contents into the workspace; the specified outcome instead requires the team to reference a registered datastore path as job input so paths remain portable across job definitions.
D:Rotating the underlying credential and updating any dependent connection removes reliance on the exposed value; registering a new dataset does not invalidate an existing leaked storage key; the implementation under review needs to reference a registered datastore path as job input so paths remain portable across job definitions.
Question 12
An Azure ML training pipeline still authenticates to Blob Storage with a shared key that was accidentally committed to a repository. Storage data-plane RBAC and network connectivity are already correctly scoped for the job identity. Security requires that the exposed key become unusable and that subsequent training runs stop depending on stored storage keys. Which TWO actions jointly meet those requirements? Choose TWO.
- Identity-based datastore authentication
- Datastore URI reference
- Storage Blob Data Reader role
- Credential rotation procedure
- Storage Blob Data Contributor role
Explanation
Correct Answers
A:Identity-based access avoids long-lived shared storage credentials and relies on correctly scoped data-plane RBAC; network access to the storage endpoint is still required. Here, it removes shared-key dependence from future jobs without expanding their assigned permissions.
D:Rotating the underlying credential and updating any dependent connection removes reliance on the exposed value; registering a new dataset does not invalidate an existing leaked storage key. Here, it invalidates the exposed value during the incident response rather than merely hiding it.
Incorrect Answers
B:An Azure ML datastore URI identifies a path through registered storage connection metadata; the consumer still needs correct identity and network access to resolve the referenced files; the system would still need a remedy for the disclosed storage key and the job's continued key-based authentication.
C:This data-plane role authorizes blob reads at the assigned scope; a generic Azure resource management role does not necessarily provide equivalent access to blob payloads; this mechanism leaves unresolved the disclosed storage key and the job's continued key-based authentication.
E:The data contributor role allows blob content modification at its assignment scope; it is broader than a read-only role and should be limited to locations where the pipeline writes; the observed fault still involves the disclosed storage key and the job's continued key-based authentication.
Question 13
An ML engineering organization runs both experimentation and automated workloads. The immediate design goal is to give one developer a persistent, managed development workstation for interactive notebooks and debugging. Which approach addresses that goal?
- Serverless compute for jobs
- CPU VM selection
- GPU VM selection
- Compute instance
Explanation
Correct Answer
D:An Azure ML compute instance offers a user-focused development environment. It can be stopped when idle, but it is not the preferred shared execution pool for scalable production training.
Incorrect Answers
A:Serverless job compute delegates infrastructure allocation to Azure ML for compatible jobs; it does not mean the workload executes without a configured runtime environment or compute billing; the actual design criterion is to give one developer a persistent, managed development workstation for interactive notebooks and debugging.
B:A CPU-based node avoids paying for hardware the script cannot use; compute choice should reflect workload profiling rather than treating a GPU as universally faster; the outstanding need is to give one developer a persistent, managed development workstation for interactive notebooks and debugging.
C:A GPU-capable compute SKU supplies accelerator devices to supported frameworks; selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment; this approach leaves the team unable to give one developer a persistent, managed development workstation for interactive notebooks and debugging.
Question 14
An ML engineering organization runs both experimentation and automated workloads. The team is required to run submitted training jobs on a scalable pool that can grow for parallel demand and shrink afterward. Which option is the best fit?
- GPU VM selection
- Compute instance
- Compute attachment
- AML compute cluster
- CPU VM selection
Explanation
Correct Answer
D:An Azure ML compute cluster supports managed job execution with configurable node sizes and scaling bounds. Its lifecycle differs from a persistent interactive compute instance.
Incorrect Answers
A:A GPU-capable compute SKU supplies accelerator devices to supported frameworks; selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment; the needed capability is to run submitted training jobs on a scalable pool that can grow for parallel demand and shrink afterward.
B:An Azure ML compute instance offers a user-focused development environment; it can be stopped when idle, but it is not the preferred shared execution pool for scalable production training; the requirement, by contrast, is to run submitted training jobs on a scalable pool that can grow for parallel demand and shrink afterward.
C:Attaching a supported compute target makes it selectable for compatible Azure ML jobs; it differs from registering a model or dataset, which describes artifacts rather than execution capacity; the team would still need to run submitted training jobs on a scalable pool that can grow for parallel demand and shrink afterward.
E:A CPU-based node avoids paying for hardware the script cannot use; compute choice should reflect workload profiling rather than treating a GPU as universally faster; the implementation under review needs to run submitted training jobs on a scalable pool that can grow for parallel demand and shrink afterward.
Question 15
An ML engineering organization runs both experimentation and automated workloads. An operator is tasked to submit supported training workloads without manually creating and maintaining a named compute cluster. Which mechanism provides that capability?
- Compute attachment
- Serverless compute for jobs
- CPU VM selection
- GPU VM selection
Explanation
Correct Answer
B:Serverless job compute delegates infrastructure allocation to Azure ML for compatible jobs. It does not mean the workload executes without a configured runtime environment or compute billing.
Incorrect Answers
A:Attaching a supported compute target makes it selectable for compatible Azure ML jobs; it differs from registering a model or dataset, which describes artifacts rather than execution capacity; the specified outcome instead requires the team to submit supported training workloads without manually creating and maintaining a named compute cluster.
C:A CPU-based node avoids paying for hardware the script cannot use; compute choice should reflect workload profiling rather than treating a GPU as universally faster; the requested task remains to submit supported training workloads without manually creating and maintaining a named compute cluster.
D:A GPU-capable compute SKU supplies accelerator devices to supported frameworks; selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment; this alternative does not achieve the requirement to submit supported training workloads without manually creating and maintaining a named compute cluster.
Question 16
An ML engineering organization runs both experimentation and automated workloads. The immediate design goal is to provide acceleration to a compatible deep-learning training script that requires CUDA-capable hardware. Which approach addresses that goal?
- GPU VM selection
- Serverless compute for jobs
- AML compute cluster
- Compute attachment
Explanation
Correct Answer
A:A GPU-capable compute SKU supplies accelerator devices to supported frameworks. Selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment.
Incorrect Answers
B:Serverless job compute delegates infrastructure allocation to Azure ML for compatible jobs; it does not mean the workload executes without a configured runtime environment or compute billing; the actual design criterion is to provide acceleration to a compatible deep-learning training script that requires CUDA-capable hardware.
C:An Azure ML compute cluster supports managed job execution with configurable node sizes and scaling bounds; its lifecycle differs from a persistent interactive compute instance; the outstanding need is to provide acceleration to a compatible deep-learning training script that requires CUDA-capable hardware.
D:Attaching a supported compute target makes it selectable for compatible Azure ML jobs; it differs from registering a model or dataset, which describes artifacts rather than execution capacity; this approach leaves the team unable to provide acceleration to a compatible deep-learning training script that requires CUDA-capable hardware.
Question 17
An ML engineering organization runs both experimentation and automated workloads. The team is required to minimize unnecessary accelerator cost for tabular preprocessing that has no GPU implementation. Which option is the best fit?
- GPU VM selection
- AML compute cluster
- Compute instance
- CPU VM selection
Explanation
Correct Answer
D:A CPU-based node avoids paying for hardware the script cannot use. Compute choice should reflect workload profiling rather than treating a GPU as universally faster.
Incorrect Answers
A:A GPU-capable compute SKU supplies accelerator devices to supported frameworks; selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment; this approach leaves the team unable to minimize unnecessary accelerator cost for tabular preprocessing that has no GPU implementation.
B:An Azure ML compute cluster supports managed job execution with configurable node sizes and scaling bounds; its lifecycle differs from a persistent interactive compute instance; the needed capability is to minimize unnecessary accelerator cost for tabular preprocessing that has no GPU implementation.
C:An Azure ML compute instance offers a user-focused development environment; it can be stopped when idle, but it is not the preferred shared execution pool for scalable production training; the requirement, by contrast, is to minimize unnecessary accelerator cost for tabular preprocessing that has no GPU implementation.
Question 18
An ML engineering organization runs both experimentation and automated workloads. An operator is tasked to use an existing supported compute resource as a training target without confusing it with ML asset registration. Which mechanism provides that capability?
- AML compute cluster
- CPU VM selection
- Compute attachment
- GPU VM selection
Explanation
Correct Answer
C:Attaching a supported compute target makes it selectable for compatible Azure ML jobs. It differs from registering a model or dataset, which describes artifacts rather than execution capacity.
Incorrect Answers
A:An Azure ML compute cluster supports managed job execution with configurable node sizes and scaling bounds; its lifecycle differs from a persistent interactive compute instance; the team would still need to use an existing supported compute resource as a training target without confusing it with ML asset registration.
B:A CPU-based node avoids paying for hardware the script cannot use; compute choice should reflect workload profiling rather than treating a GPU as universally faster; the specified outcome instead requires the team to use an existing supported compute resource as a training target without confusing it with ML asset registration.
D:A GPU-capable compute SKU supplies accelerator devices to supported frameworks; selecting it does not automatically parallelize code or resolve missing CUDA dependencies in the environment; the requested task remains to use an existing supported compute resource as a training target without confusing it with ML asset registration.
Question 19
The platform operations team must manage bursty training demand without uncontrolled costs. The immediate design goal is to allow a shared training cluster to scale all the way down between scheduled runs. Which approach addresses that goal?
- Minimum node count zero
- Idle scale-down time
- Regional quota request
- Low-priority compute selection
Explanation
Correct Answer
A:Setting the minimum node count to zero lets a compatible cluster deallocate idle nodes. Running jobs can still trigger provisioning, which introduces startup latency.
Incorrect Answers
B:The idle time policy governs scale-down delay for managed cluster nodes; it influences cost and startup churn without changing how a model's metrics are evaluated; this alternative does not achieve the requirement to allow a shared training cluster to scale all the way down between scheduled runs.
C:Azure VM family quotas constrain compute provisioning even when ML job YAML is otherwise valid; capacity or quota must be addressed before the requested nodes can be allocated; the actual design criterion is to allow a shared training cluster to scale all the way down between scheduled runs.
D:Low-priority nodes trade stability for lower compute cost and may be preempted; they are inappropriate for a job that must never be interrupted without recovery logic; the outstanding need is to allow a shared training cluster to scale all the way down between scheduled runs.
Question 20
The platform operations team must manage bursty training demand without uncontrolled costs. The team is required to cap simultaneous training capacity to a financially approved ceiling. Which option is the best fit?
- Minimum node count zero
- Maximum node count
- Regional quota request
- Job concurrency limit
Explanation
Correct Answer
B:The maximum node setting bounds cluster growth regardless of queued work. It is a capacity guardrail rather than a guarantee that every queued job starts immediately.
Incorrect Answers
A:Setting the minimum node count to zero lets a compatible cluster deallocate idle nodes; running jobs can still trigger provisioning, which introduces startup latency; the outstanding need is to cap simultaneous training capacity to a financially approved ceiling.
C:Azure VM family quotas constrain compute provisioning even when ML job YAML is otherwise valid; capacity or quota must be addressed before the requested nodes can be allocated; the needed capability is to cap simultaneous training capacity to a financially approved ceiling.
D:A concurrency setting controls parallel task execution and pressure on downstream resources; it is distinct from the total number of compute nodes available to the workspace; the requirement, by contrast, is to cap simultaneous training capacity to a financially approved ceiling.