Question 320
A production system needs to keep inference traffic on approved private networks and avoid public access. Which endpoint property needs review?
Explanation
Correct Answer
Serving interface exposure is independent from model quality; network controls must enforce the application's required perimeter.
Incorrect Answers
The alternative, Assign an approved identity and grant the online deployment's runtime access to the data, belongs to a different technical condition: A model needs to read secure reference data during prediction, but the deployment container cannot access the storage account. There the source-specific mechanism or proof is: Serving identities need explicit data-plane authorization rather than inheriting the data scientist's interactive workspace rights. The condition asked about here is: A production system needs to keep inference traffic on approved private networks and avoid public access. A suitable correction is Configure supported private network access and authentication for the managed serving endpoint; the targeted justification or result is: Serving interface exposure is independent from model quality; network controls must enforce the application's required perimeter.
Use online inference for synchronous responses or treat batch endpoint calls as asynchronous jobs. That operation would be relevant if: A batch endpoint processes files from a protected data store but users expect an immediate JSON prediction response for each file. Its corresponding reason or verification is: Batch endpoints execute workloads asynchronously and return job status, not interactive predictions in the immediate request response. However the problem in this question is: A production system needs to keep inference traffic on approved private networks and avoid public access. The requirement instead calls for Configure supported private network access and authentication for the managed serving endpoint; the question-specific check is: Serving interface exposure is independent from model quality; network controls must enforce the application's required perimeter.
Select a serving instance type with sufficient memory for the loaded model and request concurrency. A separate source case supports that approach when: An engineer deploys a large model with an instance SKU whose memory is smaller than its runtime footprint. The supporting rationale or observable result is: A healthy endpoint needs compute resources that can load its model and serve expected traffic without memory exhaustion. This question describes a different issue: A production system needs to keep inference traffic on approved private networks and avoid public access. Address it through Configure supported private network access and authentication for the managed serving endpoint, with the required justification or evidence: Serving interface exposure is independent from model quality; network controls must enforce the application's required perimeter.