McAfee-Secured Website

Exam Code: AI-300

Exam Name: Operationalizing Machine Learning and Generative AI Solutions

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Machine Learning Operations Engineer Associate

100% Microsoft AI-300 Guaranteed Success With Testking Training Materials

Get Real AI-300 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

213 Questions & Answers with Testing Engine

"Operationalizing Machine Learning and Generative AI Solutions Exam", also known as AI-300 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-300 Exam Engine. Your AI-300 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-300 Practice Test 3

Question 20

A job can read a datastore from a user's notebook but fails on a managed compute cluster. Why should the platform team inspect identity context?

  1. Configure the datastore to reference the intended ADLS Gen2 filesystem and directory path.
  2. Review datastore resource references and cross-subscription identity permissions explicitly.
  3. Grant storage data access to the job's runtime identity and verify datastore authentication mode.
  4. Renew the expired delegated credential or migrate the connection to managed identity access.

Explanation

Correct Answer

C:

Interactive user authorization may not transfer to managed job identities; compute execution must possess its own permitted access.

Incorrect Answers

A:

Configure the datastore to reference the intended ADLS Gen2 filesystem and directory path. It addresses another case: An Azure ML job expects nested directories in an ADLS Gen2 container, but its configured datastore points to a different blob root. For that case, the distinguishing justification or acceptance check is: The datastore must identify the actual backing storage location; changing experiment names cannot fix a misaddressed filesystem. The present case instead states: A job can read a datastore from a user's notebook but fails on a managed compute cluster. It calls for Grant storage data access to the job's runtime identity and verify datastore authentication mode. Showing success for the other case would not establish: Interactive user authorization may not transfer to managed job identities; compute execution must possess its own permitted access.

B:

The alternative, Review datastore resource references and cross-subscription identity permissions explicitly, belongs to a different technical condition: A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. There the source-specific mechanism or proof is: Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization. The condition asked about here is: A job can read a datastore from a user's notebook but fails on a managed compute cluster. A suitable correction is Grant storage data access to the job's runtime identity and verify datastore authentication mode; the targeted justification or result is: Interactive user authorization may not transfer to managed job identities; compute execution must possess its own permitted access.

D:

Renew the expired delegated credential or migrate the connection to managed identity access. That operation would be relevant if: A datastore configured with a temporary SAS token worked yesterday but now returns 403 responses on every job. Its corresponding reason or verification is: Time-limited SAS credentials expire independently of the model code; credential renewal or identity-based access restores authorization. However the problem in this question is: A job can read a datastore from a user's notebook but fails on a managed compute cluster. The requirement instead calls for Grant storage data access to the job's runtime identity and verify datastore authentication mode; the question-specific check is: Interactive user authorization may not transfer to managed job identities; compute execution must possess its own permitted access.

Question 21

A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. Which configuration should be reassessed?

  1. Create a named datastore targeting the governed production account and use it in job inputs.
  2. Review datastore resource references and cross-subscription identity permissions explicitly.
  3. Grant storage data access to the job's runtime identity and verify datastore authentication mode.
  4. Check workspace or compute network path and storage firewall/private endpoint rules for the datastore.

Explanation

Correct Answer

B:

Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization.

Incorrect Answers

A:

Create a named datastore targeting the governed production account and use it in job inputs. This is useful for the distinct issue: A pipeline reads from its workspace default datastore, but production data belongs in a separate governed storage account. The relevant technical rationale or evidence for that issue is: Explicit data locations avoid silently relying on the default backing store and make production access auditable. Here the reported issue is: A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. The relevant response is: Review datastore resource references and cross-subscription identity permissions explicitly. The first result cannot verify this different correction: Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization.

C:

Grant storage data access to the job's runtime identity and verify datastore authentication mode. It addresses another case: A job can read a datastore from a user's notebook but fails on a managed compute cluster. For that case, the distinguishing justification or acceptance check is: Interactive user authorization may not transfer to managed job identities; compute execution must possess its own permitted access. The present case instead states: A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. It calls for Review datastore resource references and cross-subscription identity permissions explicitly. Showing success for the other case would not establish: Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization.

D:

The alternative, Check workspace or compute network path and storage firewall/private endpoint rules for the datastore, belongs to a different technical condition: A datastore shows healthy metadata but read operations fail only after Storage firewall restrictions were enabled. There the source-specific mechanism or proof is: A valid datastore reference does not override network denial; the data plane must be reachable from executing compute. The condition asked about here is: A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. A suitable correction is Review datastore resource references and cross-subscription identity permissions explicitly; the targeted justification or result is: Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization.

Question 22

An engineer proposes making training storage public to resolve datastore read failures. What safer diagnostic sequence should be preferred?

  1. Confirm data-plane permissions and private network connectivity without opening public access.
  2. Review datastore resource references and cross-subscription identity permissions explicitly.
  3. Register authorized datastores that reference the same permitted storage location.
  4. Create a named datastore targeting the governed production account and use it in job inputs.

Explanation

Correct Answer

A:

Public exposure is unnecessary when the real fault may be identity or private endpoint routing configuration.

Incorrect Answers

B:

Review datastore resource references and cross-subscription identity permissions explicitly. A separate source case supports that approach when: A workspace is moved to a new subscription and its datastore still points to the previous subscription's storage account. The supporting rationale or observable result is: Subscription changes do not transparently update storage references or their RBAC; the new workspace needs valid data-plane authorization. This question describes a different issue: An engineer proposes making training storage public to resolve datastore read failures. Address it through Confirm data-plane permissions and private network connectivity without opening public access, with the required justification or evidence: Public exposure is unnecessary when the real fault may be identity or private endpoint routing configuration.

C:

Register authorized datastores that reference the same permitted storage location. This is useful for the distinct issue: Two Azure ML training teams need read-only access to the same large files without copying them into each workspace. The relevant technical rationale or evidence for that issue is: Datastores provide reusable pointers and access configuration without forcing another physical copy of every dataset. Here the reported issue is: An engineer proposes making training storage public to resolve datastore read failures. The relevant response is: Confirm data-plane permissions and private network connectivity without opening public access. The first result cannot verify this different correction: Public exposure is unnecessary when the real fault may be identity or private endpoint routing configuration.

D:

Create a named datastore targeting the governed production account and use it in job inputs. It addresses another case: A pipeline reads from its workspace default datastore, but production data belongs in a separate governed storage account. For that case, the distinguishing justification or acceptance check is: Explicit data locations avoid silently relying on the default backing store and make production access auditable. The present case instead states: An engineer proposes making training storage public to resolve datastore read failures. It calls for Confirm data-plane permissions and private network connectivity without opening public access. Showing success for the other case would not establish: Public exposure is unnecessary when the real fault may be identity or private endpoint routing configuration.

Question 23

A data scientist wants a persistent managed notebook development machine for interactive experiments. Which Azure ML compute resource fits best?

  1. Select a compute SKU with sufficient memory per node for the job's actual footprint.
  2. Provision a user-dedicated Azure Machine Learning compute instance for interactive development.
  3. Increase the cluster's maximum node count within quota and workload capacity requirements.
  4. Request the correct VM-family core quota or select another permitted SKU.

Explanation

Correct Answer

B:

Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads.

Incorrect Answers

A:

Select a compute SKU with sufficient memory per node for the job's actual footprint. That operation would be relevant if: A large training run repeatedly exhausts memory but CPU utilization remains low on current nodes. Its corresponding reason or verification is: Memory pressure depends on VM characteristics and job data; increasing CPU utilization settings does not add physical RAM. However the problem in this question is: A data scientist wants a persistent managed notebook development machine for interactive experiments. The requirement instead calls for Provision a user-dedicated Azure Machine Learning compute instance for interactive development; the question-specific check is: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads.

C:

Increase the cluster's maximum node count within quota and workload capacity requirements. A separate source case supports that approach when: A team wants predictable concurrency for 35 simultaneous jobs but its compute cluster maximum nodes is fixed at one. The supporting rationale or observable result is: Autoscaling capacity limits determine how much work can execute concurrently; a larger job queue alone cannot increase throughput. This question describes a different issue: A data scientist wants a persistent managed notebook development machine for interactive experiments. Address it through Provision a user-dedicated Azure Machine Learning compute instance for interactive development, with the required justification or evidence: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads.

D:

Request the correct VM-family core quota or select another permitted SKU. This is useful for the distinct issue: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. The relevant technical rationale or evidence for that issue is: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically. Here the reported issue is: A data scientist wants a persistent managed notebook development machine for interactive experiments. The relevant response is: Provision a user-dedicated Azure Machine Learning compute instance for interactive development. The first result cannot verify this different correction: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads.

Question 24

A team submits many short training jobs and wants nodes to scale down to zero when no jobs remain. Which Azure ML compute design addresses this?

  1. Use an autoscaling compute cluster with minimum node count set to zero.
  2. Provision a user-dedicated Azure Machine Learning compute instance for interactive development.
  3. Configure idle shutdown or automatic start/stop schedules for the compute instance.
  4. Choose a supported GPU compute target with the necessary GPU quota and runtime.

Explanation

Correct Answer

A:

Managed compute clusters can provision capacity for jobs and scale back when idle, reducing unnecessary compute charges.

Incorrect Answers

B:

The alternative, Provision a user-dedicated Azure Machine Learning compute instance for interactive development, belongs to a different technical condition: A data scientist wants a persistent managed notebook development machine for interactive experiments. There the source-specific mechanism or proof is: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads. The condition asked about here is: A team submits many short training jobs and wants nodes to scale down to zero when no jobs remain. A suitable correction is Use an autoscaling compute cluster with minimum node count set to zero; the targeted justification or result is: Managed compute clusters can provision capacity for jobs and scale back when idle, reducing unnecessary compute charges.

C:

Configure idle shutdown or automatic start/stop schedules for the compute instance. That operation would be relevant if: A compute instance remains billed overnight even when no notebook kernel is active. Its corresponding reason or verification is: Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks. However the problem in this question is: A team submits many short training jobs and wants nodes to scale down to zero when no jobs remain. The requirement instead calls for Use an autoscaling compute cluster with minimum node count set to zero; the question-specific check is: Managed compute clusters can provision capacity for jobs and scale back when idle, reducing unnecessary compute charges.

D:

Choose a supported GPU compute target with the necessary GPU quota and runtime. A separate source case supports that approach when: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. The supporting rationale or observable result is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity. This question describes a different issue: A team submits many short training jobs and wants nodes to scale down to zero when no jobs remain. Address it through Use an autoscaling compute cluster with minimum node count set to zero, with the required justification or evidence: Managed compute clusters can provision capacity for jobs and scale back when idle, reducing unnecessary compute charges.

Question 25

A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. What should change?

  1. Configure idle shutdown or automatic start/stop schedules for the compute instance.
  2. Request the correct VM-family core quota or select another permitted SKU.
  3. Increase the cluster's maximum node count within quota and workload capacity requirements.
  4. Choose a supported GPU compute target with the necessary GPU quota and runtime.

Explanation

Correct Answer

D:

GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity.

Incorrect Answers

A:

Configure idle shutdown or automatic start/stop schedules for the compute instance. It addresses another case: A compute instance remains billed overnight even when no notebook kernel is active. For that case, the distinguishing justification or acceptance check is: Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks. The present case instead states: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. It calls for Choose a supported GPU compute target with the necessary GPU quota and runtime. Showing success for the other case would not establish: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity.

B:

The alternative, Request the correct VM-family core quota or select another permitted SKU, belongs to a different technical condition: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. There the source-specific mechanism or proof is: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically. The condition asked about here is: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. A suitable correction is Choose a supported GPU compute target with the necessary GPU quota and runtime; the targeted justification or result is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity.

C:

Increase the cluster's maximum node count within quota and workload capacity requirements. That operation would be relevant if: A team wants predictable concurrency for 35 simultaneous jobs but its compute cluster maximum nodes is fixed at one. Its corresponding reason or verification is: Autoscaling capacity limits determine how much work can execute concurrently; a larger job queue alone cannot increase throughput. However the problem in this question is: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. The requirement instead calls for Choose a supported GPU compute target with the necessary GPU quota and runtime; the question-specific check is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity.

Question 26

A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. What should be fixed?

  1. Increase the cluster's maximum node count within quota and workload capacity requirements.
  2. Choose a supported GPU compute target with the necessary GPU quota and runtime.
  3. Request the correct VM-family core quota or select another permitted SKU.
  4. Select a compute SKU with sufficient memory per node for the job's actual footprint.

Explanation

Correct Answer

C:

Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically.

Incorrect Answers

A:

Increase the cluster's maximum node count within quota and workload capacity requirements. This is useful for the distinct issue: A team wants predictable concurrency for 35 simultaneous jobs but its compute cluster maximum nodes is fixed at one. The relevant technical rationale or evidence for that issue is: Autoscaling capacity limits determine how much work can execute concurrently; a larger job queue alone cannot increase throughput. Here the reported issue is: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. The relevant response is: Request the correct VM-family core quota or select another permitted SKU. The first result cannot verify this different correction: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically.

B:

Choose a supported GPU compute target with the necessary GPU quota and runtime. It addresses another case: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. For that case, the distinguishing justification or acceptance check is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity. The present case instead states: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. It calls for Request the correct VM-family core quota or select another permitted SKU. Showing success for the other case would not establish: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically.

D:

The alternative, Select a compute SKU with sufficient memory per node for the job's actual footprint, belongs to a different technical condition: A large training run repeatedly exhausts memory but CPU utilization remains low on current nodes. There the source-specific mechanism or proof is: Memory pressure depends on VM characteristics and job data; increasing CPU utilization settings does not add physical RAM. The condition asked about here is: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. A suitable correction is Request the correct VM-family core quota or select another permitted SKU; the targeted justification or result is: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically.

Question 27

A shared batch-processing cluster frequently evicts interactive notebook users. What architectural change separates development from batch execution?

  1. Point the job to an existing registered compute target or provision its required cluster.
  2. Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs.
  3. Assign and grant data access to the appropriate managed identity for the compute execution.
  4. Provision a user-dedicated Azure Machine Learning compute instance for interactive development.

Explanation

Correct Answer

B:

These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes.

Incorrect Answers

A:

Point the job to an existing registered compute target or provision its required cluster. A separate source case supports that approach when: A cluster was deleted after a completed experiment; a new run still references its name and fails before training starts. The supporting rationale or observable result is: Compute target names must resolve to available resources; a preserved experiment record cannot execute on deleted infrastructure. This question describes a different issue: A shared batch-processing cluster frequently evicts interactive notebook users. Address it through Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs, with the required justification or evidence: These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes.

C:

Assign and grant data access to the appropriate managed identity for the compute execution. This is useful for the distinct issue: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. The relevant technical rationale or evidence for that issue is: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context. Here the reported issue is: A shared batch-processing cluster frequently evicts interactive notebook users. The relevant response is: Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs. The first result cannot verify this different correction: These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes.

D:

Provision a user-dedicated Azure Machine Learning compute instance for interactive development. It addresses another case: A data scientist wants a persistent managed notebook development machine for interactive experiments. For that case, the distinguishing justification or acceptance check is: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads. The present case instead states: A shared batch-processing cluster frequently evicts interactive notebook users. It calls for Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs. Showing success for the other case would not establish: These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes.

Question 28

A compute instance remains billed overnight even when no notebook kernel is active. Which operating policy directly reduces idle waste?

  1. Configure idle shutdown or automatic start/stop schedules for the compute instance.
  2. Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs.
  3. Assign and grant data access to the appropriate managed identity for the compute execution.
  4. Provision a user-dedicated Azure Machine Learning compute instance for interactive development.

Explanation

Correct Answer

A:

Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks.

Incorrect Answers

B:

Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs. That operation would be relevant if: A shared batch-processing cluster frequently evicts interactive notebook users. Its corresponding reason or verification is: These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes. However the problem in this question is: A compute instance remains billed overnight even when no notebook kernel is active. The requirement instead calls for Configure idle shutdown or automatic start/stop schedules for the compute instance; the question-specific check is: Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks.

C:

Assign and grant data access to the appropriate managed identity for the compute execution. A separate source case supports that approach when: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. The supporting rationale or observable result is: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context. This question describes a different issue: A compute instance remains billed overnight even when no notebook kernel is active. Address it through Configure idle shutdown or automatic start/stop schedules for the compute instance, with the required justification or evidence: Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks.

D:

Provision a user-dedicated Azure Machine Learning compute instance for interactive development. This is useful for the distinct issue: A data scientist wants a persistent managed notebook development machine for interactive experiments. The relevant technical rationale or evidence for that issue is: Compute instances provide managed interactive workstations; shared autoscaling job clusters are designed for submitted execution workloads. Here the reported issue is: A compute instance remains billed overnight even when no notebook kernel is active. The relevant response is: Configure idle shutdown or automatic start/stop schedules for the compute instance. The first result cannot verify this different correction: Instance billing continues while running; an idle-stop schedule reduces running hours without deleting code and notebooks.

Question 29

A restricted workspace prohibits public IP addresses on compute hosts, but an engineer selected the default network option. What should be specified?

  1. Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs.
  2. Choose a supported GPU compute target with the necessary GPU quota and runtime.
  3. Assign and grant data access to the appropriate managed identity for the compute execution.
  4. Deploy compute through a supported private or managed network configuration without public IP.

Explanation

Correct Answer

D:

Network isolation must be designed into compute connectivity; a notebook setting alone does not remove external host exposure.

Incorrect Answers

A:

The alternative, Use dedicated compute instances for notebooks and a managed job compute cluster for batch jobs, belongs to a different technical condition: A shared batch-processing cluster frequently evicts interactive notebook users. There the source-specific mechanism or proof is: These compute types serve different lifecycles; isolating workloads prevents interactive sessions from relying on autoscaled batch nodes. The condition asked about here is: A restricted workspace prohibits public IP addresses on compute hosts, but an engineer selected the default network option. A suitable correction is Deploy compute through a supported private or managed network configuration without public IP; the targeted justification or result is: Network isolation must be designed into compute connectivity; a notebook setting alone does not remove external host exposure.

B:

Choose a supported GPU compute target with the necessary GPU quota and runtime. That operation would be relevant if: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. Its corresponding reason or verification is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity. However the problem in this question is: A restricted workspace prohibits public IP addresses on compute hosts, but an engineer selected the default network option. The requirement instead calls for Deploy compute through a supported private or managed network configuration without public IP; the question-specific check is: Network isolation must be designed into compute connectivity; a notebook setting alone does not remove external host exposure.

C:

Assign and grant data access to the appropriate managed identity for the compute execution. A separate source case supports that approach when: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. The supporting rationale or observable result is: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context. This question describes a different issue: A restricted workspace prohibits public IP addresses on compute hosts, but an engineer selected the default network option. Address it through Deploy compute through a supported private or managed network configuration without public IP, with the required justification or evidence: Network isolation must be designed into compute connectivity; a notebook setting alone does not remove external host exposure.

Question 30

A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. Which setting should be inspected?

  1. Request the correct VM-family core quota or select another permitted SKU.
  2. Use an autoscaling compute cluster with minimum node count set to zero.
  3. Assign and grant data access to the appropriate managed identity for the compute execution.
  4. Choose a supported GPU compute target with the necessary GPU quota and runtime.

Explanation

Correct Answer

C:

Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context.

Incorrect Answers

A:

The alternative, Request the correct VM-family core quota or select another permitted SKU, belongs to a different technical condition: A newly created high-memory training cluster stays at zero capacity after all jobs enter the queue. The quota page reports zero remaining cores in that VM family. There the source-specific mechanism or proof is: Cluster scaling is bounded by regional and VM-family quota; queued work cannot create unavailable cores automatically. The condition asked about here is: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. A suitable correction is Assign and grant data access to the appropriate managed identity for the compute execution; the targeted justification or result is: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context.

B:

Use an autoscaling compute cluster with minimum node count set to zero. That operation would be relevant if: A team submits many short training jobs and wants nodes to scale down to zero when no jobs remain. Its corresponding reason or verification is: Managed compute clusters can provision capacity for jobs and scale back when idle, reducing unnecessary compute charges. However the problem in this question is: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. The requirement instead calls for Assign and grant data access to the appropriate managed identity for the compute execution; the question-specific check is: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context.

D:

Choose a supported GPU compute target with the necessary GPU quota and runtime. A separate source case supports that approach when: A deep-learning job cannot start on the selected CPU-only cluster because CUDA acceleration is required. The supporting rationale or observable result is: GPU training needs compatible hardware and drivers; changing the model display name cannot provide CUDA capacity. This question describes a different issue: A job logs under an employee's identity on a local compute instance but fails on a managed cluster when reading storage. Address it through Assign and grant data access to the appropriate managed identity for the compute execution, with the required justification or evidence: Job identity can differ from interactive user identity; runtime access must be granted explicitly to the executing context.

Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests