Microsoft AI-300 (Operationalizing Machine Learning and Generative AI Solutions) Practice Test 25

Question 240

A data-processing team scores millions of stored records overnight rather than serving interactive requests. An operator is tasked to recover transient failures in independent scoring work units without rerunning an entire successful workload. Which mechanism provides that capability?

  1. Batch endpoint
  2. Batch job output location
  3. Batch mini-batch tuning
  4. Batch retry policy

Explanation

Correct Answer

D:

Retry settings can reattempt failed units under supported batch execution behavior. Persistent schema defects should be fixed rather than retried indefinitely.

Incorrect Answers

A:

A batch endpoint initiates batch inference jobs and produces output artifacts; it is not the same low-latency request-response path as online inference; the outstanding need is to recover transient failures in independent scoring work units without rerunning an entire successful workload.

B:

Batch execution persists result artifacts rather than returning every prediction inline to the initiating client; the output location must be accessible to the job identity; this approach leaves the team unable to recover transient failures in independent scoring work units without rerunning an entire successful workload.

C:

Mini-batch sizing affects worker overhead and resource consumption; it is separate from the number of model versions in a registry; the needed capability is to recover transient failures in independent scoring work units without rerunning an entire successful workload.

Question 241

An online deployment fails or gives incorrect responses despite a valid registered model. The immediate design goal is to load or initialize the model and other reusable serving resources when the inference worker starts. Which approach addresses that goal?

  1. Health probe readiness
  2. Scoring run function
  3. Scoring init function
  4. Input schema validation

Explanation

Correct Answer

C:

An initialization routine prepares the model once per worker lifecycle. Placing heavyweight loading inside every request handler increases latency.

Incorrect Answers

A:

Readiness checks help determine whether a worker can accept traffic; they do not certify predictive accuracy or fairness; the requirement, by contrast, is to load or initialize the model and other reusable serving resources when the inference worker starts.

B:

The request handler implements per-call inference logic and output formatting; it must respect the client's payload contract and expected response schema; the team would still need to load or initialize the model and other reusable serving resources when the inference worker starts.

D:

Early validation checks required fields, data types, and allowed ranges; it improves error clarity but does not fix missing model dependencies; the implementation under review needs to load or initialize the model and other reusable serving resources when the inference worker starts.

Question 242

An online deployment fails or gives incorrect responses despite a valid registered model. The team is required to parse each incoming prediction request and produce a valid response using the initialized model. Which option is the best fit?

  1. Custom environment packaging
  2. Input schema validation
  3. Scoring run function
  4. Scoring init function

Explanation

Correct Answer

C:

The request handler implements per-call inference logic and output formatting. It must respect the client's payload contract and expected response schema.

Incorrect Answers

A:

Serving images must contain libraries compatible with the saved estimator; a training environment is not automatically reused unless specified and compatible; the implementation under review needs to parse each incoming prediction request and produce a valid response using the initialized model.

B:

Early validation checks required fields, data types, and allowed ranges; it improves error clarity but does not fix missing model dependencies; the requested task remains to parse each incoming prediction request and produce a valid response using the initialized model.

D:

An initialization routine prepares the model once per worker lifecycle; placing heavyweight loading inside every request handler increases latency; the actual design criterion is to parse each incoming prediction request and produce a valid response using the initialized model.

Question 243

An online deployment fails or gives incorrect responses despite a valid registered model. An operator is tasked to reject malformed payloads before they trigger obscure model prediction errors. Which mechanism provides that capability?

  1. Inference request serialization
  2. Scoring run function
  3. Health probe readiness
  4. Input schema validation

Explanation

Correct Answer

D:

Early validation checks required fields, data types, and allowed ranges. It improves error clarity but does not fix missing model dependencies.

Incorrect Answers

A:

Serialization and feature ordering preserve the model's input contract; a successful HTTP request can still yield wrong predictions if columns are misaligned; the actual design criterion is to reject malformed payloads before they trigger obscure model prediction errors.

B:

The request handler implements per-call inference logic and output formatting; it must respect the client's payload contract and expected response schema; the outstanding need is to reject malformed payloads before they trigger obscure model prediction errors.

C:

Readiness checks help determine whether a worker can accept traffic; they do not certify predictive accuracy or fairness; this approach leaves the team unable to reject malformed payloads before they trigger obscure model prediction errors.

Question 244

An online deployment fails or gives incorrect responses despite a valid registered model. The immediate design goal is to include native and Python dependencies needed by the model's inference code. Which approach addresses that goal?

  1. Inference request serialization
  2. Custom environment packaging
  3. Scoring init function
  4. Input schema validation

Explanation

Correct Answer

B:

Serving images must contain libraries compatible with the saved estimator. A training environment is not automatically reused unless specified and compatible.

Incorrect Answers

A:

Serialization and feature ordering preserve the model's input contract; a successful HTTP request can still yield wrong predictions if columns are misaligned; the needed capability is to include native and Python dependencies needed by the model's inference code.

C:

An initialization routine prepares the model once per worker lifecycle; placing heavyweight loading inside every request handler increases latency; the team would still need to include native and Python dependencies needed by the model's inference code.

D:

Early validation checks required fields, data types, and allowed ranges; it improves error clarity but does not fix missing model dependencies; the specified outcome instead requires the team to include native and Python dependencies needed by the model's inference code.

Question 245

An online deployment fails or gives incorrect responses despite a valid registered model. The team is required to convert incoming JSON fields to correctly ordered feature values before calling the estimator. Which option is the best fit?

  1. Inference request serialization
  2. Scoring run function
  3. Input schema validation
  4. Custom environment packaging
  5. Health probe readiness

Explanation

Correct Answer

A:

Serialization and feature ordering preserve the model's input contract. A successful HTTP request can still yield wrong predictions if columns are misaligned.

Incorrect Answers

B:

The request handler implements per-call inference logic and output formatting; it must respect the client's payload contract and expected response schema; the implementation under review needs to convert incoming JSON fields to correctly ordered feature values before calling the estimator.

C:

Early validation checks required fields, data types, and allowed ranges; it improves error clarity but does not fix missing model dependencies; the requested task remains to convert incoming JSON fields to correctly ordered feature values before calling the estimator.

D:

Serving images must contain libraries compatible with the saved estimator; a training environment is not automatically reused unless specified and compatible; this alternative does not achieve the requirement to convert incoming JSON fields to correctly ordered feature values before calling the estimator.

E:

Readiness checks help determine whether a worker can accept traffic; they do not certify predictive accuracy or fairness; the actual design criterion is to convert incoming JSON fields to correctly ordered feature values before calling the estimator.

Question 246

An online deployment fails or gives incorrect responses despite a valid registered model. An operator is tasked to keep a new inference replica out of normal service until initialization is complete and it is ready. Which mechanism provides that capability?

  1. Inference request serialization
  2. Custom environment packaging
  3. Health probe readiness
  4. Scoring init function
  5. Scoring run function
  6. Input schema validation

Explanation

Correct Answer

C:

Readiness checks help determine whether a worker can accept traffic. They do not certify predictive accuracy or fairness.

Incorrect Answers

A:

Serialization and feature ordering preserve the model's input contract; a successful HTTP request can still yield wrong predictions if columns are misaligned; this alternative does not achieve the requirement to keep a new inference replica out of normal service until initialization is complete and it is ready.

B:

Serving images must contain libraries compatible with the saved estimator; a training environment is not automatically reused unless specified and compatible; the actual design criterion is to keep a new inference replica out of normal service until initialization is complete and it is ready.

D:

An initialization routine prepares the model once per worker lifecycle; placing heavyweight loading inside every request handler increases latency; this approach leaves the team unable to keep a new inference replica out of normal service until initialization is complete and it is ready.

E:

The request handler implements per-call inference logic and output formatting; it must respect the client's payload contract and expected response schema; the needed capability is to keep a new inference replica out of normal service until initialization is complete and it is ready.

F:

Early validation checks required fields, data types, and allowed ranges; it improves error clarity but does not fix missing model dependencies; the requirement, by contrast, is to keep a new inference replica out of normal service until initialization is complete and it is ready.

Question 247

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. The immediate design goal is to supply an endpoint-issued key where key authentication has been selected for client requests. Which approach addresses that goal?

  1. Credential rotation
  2. Token-based invocation
  3. HTTP status investigation
  4. Key-based invocation

Explanation

Correct Answer

D:

Key authentication verifies possession of a configured secret for invocation. It is distinct from Azure RBAC authorization for managing the endpoint resource.

Incorrect Answers

A:

Key rotation limits continued use of a compromised credential; changing traffic weights alone does not revoke knowledge of the old key; this approach leaves the team unable to supply an endpoint-issued key where key authentication has been selected for client requests.

B:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; the needed capability is to supply an endpoint-issued key where key authentication has been selected for client requests.

C:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving; retrying every error without classification obscures root causes; the requirement, by contrast, is to supply an endpoint-issued key where key authentication has been selected for client requests.

Question 248

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. The team is required to acquire an appropriate access token when the endpoint is configured for token-based authentication. Which option is the best fit?

  1. Credential rotation
  2. Data plane versus control plane
  3. Endpoint scoring URI
  4. Token-based invocation

Explanation

Correct Answer

D:

Token invocation relies on supported token issuance and correct request headers. Sending a storage account key does not authenticate an ML endpoint.

Incorrect Answers

A:

Key rotation limits continued use of a compromised credential; changing traffic weights alone does not revoke knowledge of the old key; the team would still need to acquire an appropriate access token when the endpoint is configured for token-based authentication.

B:

Inference authorization and management privileges protect different operations; a user who can read endpoint metadata is not necessarily authorized to score requests; the specified outcome instead requires the team to acquire an appropriate access token when the endpoint is configured for token-based authentication.

C:

The scoring URI is the service data-plane interface for inference requests; resource-management endpoints control configuration and are not interchangeable prediction targets; the implementation under review needs to acquire an appropriate access token when the endpoint is configured for token-based authentication.

Question 249

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. An operator is tasked to send application predictions to the specific endpoint scoring address rather than an Azure management API path. Which mechanism provides that capability?

  1. Endpoint scoring URI
  2. HTTP status investigation
  3. Key-based invocation
  4. Token-based invocation

Explanation

Correct Answer

A:

The scoring URI is the service data-plane interface for inference requests. Resource-management endpoints control configuration and are not interchangeable prediction targets.

Incorrect Answers

B:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving; retrying every error without classification obscures root causes; this alternative does not achieve the requirement to send application predictions to the specific endpoint scoring address rather than an Azure management API path.

C:

Key authentication verifies possession of a configured secret for invocation; it is distinct from Azure RBAC authorization for managing the endpoint resource; the actual design criterion is to send application predictions to the specific endpoint scoring address rather than an Azure management API path.

D:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; the outstanding need is to send application predictions to the specific endpoint scoring address rather than an Azure management API path.

Question 250

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. The immediate design goal is to distinguish permission to invoke a model from permission to update its deployment configuration. Which approach addresses that goal?

  1. HTTP status investigation
  2. Data plane versus control plane
  3. Key-based invocation
  4. Token-based invocation

Explanation

Correct Answer

B:

Inference authorization and management privileges protect different operations. A user who can read endpoint metadata is not necessarily authorized to score requests.

Incorrect Answers

A:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving; retrying every error without classification obscures root causes; the outstanding need is to distinguish permission to invoke a model from permission to update its deployment configuration.

C:

Key authentication verifies possession of a configured secret for invocation; it is distinct from Azure RBAC authorization for managing the endpoint resource; the needed capability is to distinguish permission to invoke a model from permission to update its deployment configuration.

D:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; the requirement, by contrast, is to distinguish permission to invoke a model from permission to update its deployment configuration.

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