Deploy and retrain a model
Free Microsoft Certified: Azure Data Scientist Associate practice — 6 questions on Deploy and retrain a model, with explanations. No sign-up. Full 12-question mixed test →
You have a managed online endpoint named 'fraud-endpoint' with a live deployment 'blue' receiving 100% of traffic. You want to test a new deployment 'green' against real production traffic for the next 24 hours, but green's predictions must NEVER be returned to callers, and blue must continue serving 100% of live results. Which configuration achieves this?
Mirror traffic (shadow testing) lets a deployment receive a copy of live requests for evaluation without its results ever being returned to the caller. The maximum supported mirror_traffic percentage is 50%, while the live traffic split stays 100/0 so blue continues serving all real responses.
Your managed online endpoint deployment 'green' is throttling during peak load. You want it to automatically scale from 2 to 10 instances when average CPU utilization exceeds 70%. Which approach correctly implements this using CLI v2/SDK v2?
Autoscaling for managed online endpoint deployments is implemented through Azure Monitor autoscale settings applied to the deployment resource (Microsoft.MachineLearningServices/workspaces/onlineEndpoints/deployments), with rules based on metrics like CPU utilization percentage and configured min/max/default instance counts.
Data scientists need a retraining pipeline that automatically starts whenever a new labeled dataset file lands in an Azure Blob Storage container, without waiting on a fixed schedule. Which solution best meets this requirement in Azure Machine Learning?
Azure ML SDK v2 JobSchedule supports only time-based triggers (cron/recurrence), not native blob-event triggers. To react immediately to new files, you wire Event Grid's BlobCreated event to an Azure Function (or Logic App) that calls the published pipeline's REST endpoint to start a run — this is the standard event-driven retraining pattern.
A batch endpoint deployment scores 2 million individual image files. Throughput is far below expectations because each file is processed as its own call to the scoring script's run() function, causing high per-call overhead. Which deployment setting should you increase to improve throughput?
mini_batch_size controls how many input files (or rows, for tabular data) are grouped together and passed to a single run() invocation. Increasing it reduces the number of run() calls needed, cutting per-call overhead and improving throughput when scoring many small files.
A client application calls a managed online endpoint every few seconds via a long-running background service. Security requires that credentials automatically expire and cannot be used indefinitely if leaked; the client already uses the SDK to refresh credentials programmatically. Which auth_mode should the endpoint use, and what must the client account for?
aml_token issues short-lived, workspace-scoped tokens with a default expiry of roughly one hour. Clients must proactively refresh the token before expiry using the SDK, satisfying the requirement for credentials that expire automatically rather than remaining valid indefinitely.
How does traffic routing differ between managed online endpoints and batch endpoints when multiple deployments exist under the same endpoint?
Managed online endpoints route live traffic based on configurable percentage allocations across deployments (blue/green). Batch endpoints have no traffic-splitting concept — every invocation goes to whichever deployment is marked 'default' unless the caller explicitly names a deployment in the invoke call.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →