I85PREP
⌘ K
AI-300 study materials
AI-300 study resources

AI-300 practice questions

One question at a time. Build your AI-300 understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Design and implement an MLOps infrastructure

A team needs one top-level Azure resource that holds models, environments, compute, and pipelines together. Which resource should you create?

  1. A compute cluster in the team's resource group
  2. An Application Insights instance for each model
  3. A model registry in each project subscription
  4. An Azure Machine Learning workspace Correct answer.

Explanation

The workspace is the top-level container in which models, environments, components, datastores, compute, and pipelines are created and shared. MLOps automation targets the workspace, so assets and jobs scoped to it can be versioned and referenced consistently.

Question 2Design and implement an MLOps infrastructure

You need a training cluster that can scale down to zero nodes when idle and grow to at most 10 nodes when jobs are queued. Which setting combination meets this?

  1. Maximum 0 nodes with autoscaling disabled
  2. Minimum 0 nodes and maximum 10 nodes Correct answer.
  3. Minimum 10 nodes and maximum 0 nodes
  4. Idle time of 0 seconds with 10 fixed nodes

Explanation

Autoscaling is bounded by the minimum and maximum node counts: a minimum of 0 lets the cluster release all nodes when no jobs are queued, and a maximum of 10 caps growth. Idle scale-down settings only control when nodes are released, not whether the cluster can shrink to zero.

Question 3Design and implement an MLOps infrastructure

A data scientist needs a persistent, single-node development environment with notebooks inside the workspace. Which compute target should you provision?

  1. Serverless compute for batch scoring
  2. A compute instance Correct answer.
  3. A compute cluster with four dedicated nodes
  4. A Kubernetes cluster attached to the workspace

Explanation

A compute instance is a managed, single-node virtual machine intended for interactive authoring, notebooks, and development work. It stays available for the assigned user rather than being created per job.

Question 4Design and implement an MLOps infrastructure

You create a datastore that authenticates to a storage account with an account key. Where does the platform keep that credential?

  1. In the environment image as a build argument
  2. In the datastore YAML committed to source control
  3. In a blob in the workspace's default container
  4. In the workspace's associated Azure Key Vault Correct answer.

Explanation

Secrets referenced by a datastore are stored in the Azure Key Vault associated with the workspace, and the datastore record holds only a reference to that secret. This keeps credentials out of code, YAML, and images.

Question 5Design and implement an MLOps infrastructure

Why should a custom environment be registered with an explicit version instead of relying on a floating latest tag?

  1. So the environment can skip private endpoint rules
  2. So compute clusters can autoscale without idle time
  3. So the exact dependency set is pinned for every run Correct answer.
  4. So the image builds faster on the first job run

Explanation

Environments are versioned assets, and a job that references a specific version resolves to the same dependency set every time it runs. That is what makes a run reproducible when a base image or package dependency changes later.

Question 6Design and implement an MLOps infrastructure

What does registering a trained model in the workspace model registry accomplish?

  1. It trains the model on a recurring schedule
  2. It exposes the model through an anonymous public URL
  3. It converts the model artifacts to ONNX format
  4. It assigns a versioned name that deployments reference Correct answer.

Explanation

Registration stores the model artifacts under a name and version so that deployments, pipelines, and downstream jobs reference an immutable asset. Each new registration of the same name creates the next version.

Question 7Design and implement an MLOps infrastructure

Which object defines a recurring trigger for a published pipeline in Azure Machine Learning?

  1. An endpoint deployment
  2. An environment version
  3. A datastore
  4. A schedule Correct answer.

Explanation

A schedule binds a pipeline to a recurrence or cron expression and can be enabled or disabled independently of the pipeline definition. It is the object that submits the pipeline automatically.

Question 8Design and implement an MLOps infrastructure

Training code running on workspace compute must read data from a storage account without any credential appearing in code or configuration. What should you configure?

  1. A shared access signature pasted into the training script
  2. The storage account key stored in the environment file
  3. A workspace managed identity with a data-reader role Correct answer.
  4. Anonymous public read access on the data container

Explanation

A managed identity gives the compute an identity in Microsoft Entra ID, and a role assignment on the storage account authorizes data access without any secret being handled by the code. Authorization is then performed through that identity at runtime.

Question 9Design and implement an MLOps infrastructure

You want to send 10 percent of live inference traffic to a new model version while the current version serves the remaining 90 percent. Which approach fits?

  1. Deploy the model to a second endpoint and use DNS round robin between them
  2. Reduce the endpoint to a single instance while testing
  3. Switch the endpoint's authentication mode from key to token
  4. Route 10 percent of endpoint traffic to a new deployment Correct answer.

Explanation

A managed online endpoint can host several deployments and distributes traffic among them by weight, so a small percentage can be directed to the new model. This provides a controlled canary rollout that can be adjusted or reversed without redeploying.

Question 10Design and implement an MLOps infrastructure

Your pipeline must register a new model version only when its accuracy on a holdout set beats the current production model. What should you add?

  1. An evaluation step that gates the registration job Correct answer.
  2. An autoscale rule on the training cluster
  3. A schedule that runs evaluation after registration
  4. A canary deployment on the batch endpoint

Explanation

The gate must run before registration, so the evaluation step produces the metric and the registration job is made conditional on that output. This keeps unqualified models out of the registry and prevents them from being deployed.

Question 11Design and implement an MLOps infrastructure

A managed online deployment must absorb bursty request volumes and release capacity when traffic falls. Which configuration provides this?

  1. An autoscale rule with minimum and maximum instance counts Correct answer.
  2. The endpoint's traffic weights across deployments
  3. The model's registered version number
  4. The cluster's idle seconds before scale-down

Explanation

Online deployments scale by instance count, driven by autoscale rules that watch metrics such as CPU utilization within a minimum and maximum range. The minimum keeps baseline capacity while the maximum caps cost during spikes.

Question 12Design and implement an MLOps infrastructure

Which workload is best served by a batch endpoint rather than an online endpoint?

  1. Streaming inference from an event hub
  2. Interactive testing inside a notebook
  3. Sub-second scoring for a mobile app
  4. Scheduled scoring of a large file of records Correct answer.

Explanation

Batch endpoints run long-lived scoring jobs over files or folders and write results to a specified output, which suits large scheduled scoring runs. Online endpoints are optimized for low-latency single-request inference.

Question 13Design and implement an MLOps infrastructure

Logging from MLflow in a script that runs on your machine should appear as a run in the workspace. What must the script be given?

  1. A registered component version identifier
  2. A batch deployment for the script
  3. An attached inference cluster
  4. A tracking URI that points at the workspace Correct answer.

Explanation

MLflow writes metrics, parameters, and artifacts to the tracking URI it is configured with, so pointing it at the workspace routes those records into the workspace's experiment tracking. Authentication is then handled by the caller's credentials.

Question 14Design and implement an MLOps infrastructure

A training pipeline must run whenever a new data file lands in a storage container. Which mechanism should you implement?

  1. A cron schedule that runs every minute
  2. A longer idle timeout on the cluster
  3. A canary deployment on the scoring endpoint
  4. An Event Grid trigger that submits the pipeline Correct answer.

Explanation

A blob-created event from the storage account can be delivered through Event Grid to a handler that submits the pipeline, which reacts to the data arrival itself. This avoids both polling and wasted runs when no new data exists.

Question 15Design and implement an MLOps infrastructure

The same data-preparation logic must be reused across several pipelines and versioned on its own. What should you create?

  1. A copy of the script in every pipeline
  2. A dedicated compute cluster per script
  3. A registered component with typed inputs and outputs Correct answer.
  4. A separate workspace for each pipeline

Explanation

A component packages code, environment, and interface into a versioned asset with typed inputs and outputs, so pipelines reference the same logic and pick up updates only when the component version changes. Copies of scripts drift apart over time.

Question 16Design and implement an MLOps infrastructure

Which capability notifies you that production input distributions have shifted away from the training baseline?

  1. Compute cluster autoscale activity logs
  2. A model monitoring signal for data drift Correct answer.
  3. Endpoint traffic mirroring between deployments
  4. A comparison of model registry versions

Explanation

Model monitoring compares production inputs against a baseline and raises a data drift signal when distributions diverge, which often precedes accuracy loss. It is set up on the deployment or model and runs on a schedule.

Question 17Design and implement an MLOps infrastructure

A custom model needs OS-level packages installed when the deployment image is built. Where should this be declared?

  1. In the endpoint's primary and secondary keys
  2. In the environment definition the deployment references Correct answer.
  3. In the deployment's traffic weight percentages
  4. In the model's registered tag values

Explanation

The environment definition carries the base image, conda or pip dependencies, and Docker build steps, so packages installed at image build time belong there. The deployment references an environment version, and the built image inherits those instructions.

Question 18Design and implement an MLOps infrastructure

Two teams work in separate workspaces and must consume the same model and environment assets. What should you use?

  1. A workspace registry that both workspaces can use Correct answer.
  2. A single datastore in the first workspace
  3. A YAML export of each asset sent by email
  4. A shared compute cluster in the first workspace

Explanation

A registry is a top-level asset store that holds models, environments, and components for consumption across many workspaces in the organization. Teams share curated versions instead of copying files between workspaces.

Question 19Design and implement an MLOps infrastructure

Which model artifact can be deployed to a managed online endpoint without writing a scoring script?

  1. An MLflow model with a logged signature Correct answer.
  2. A SavedModel directory without metadata
  3. A pickle file without an input example
  4. A zip archive of training CSVs

Explanation

An MLflow model that includes a signature describes its expected inputs and outputs, so the platform can generate the inference entry point automatically. Without that metadata the deployment needs an explicit scoring script.

Question 20Design and implement an MLOps infrastructure

A managed online endpoint serves blue at 90 percent and green at 10 percent. Green starts returning server errors, and you want it to stop serving requests while keeping its configuration for later. What should you do?

  1. Rotate the endpoint's primary and secondary keys
  2. Scale the blue deployment to one instance
  3. Set the green deployment's traffic weight to 0 Correct answer.
  4. Delete the green deployment and recreate it later

Explanation

Traffic is allocated by weight, so a weight of 0 removes the deployment from serving while keeping its image, environment, and settings intact. Re-enabling it later is a single weight change rather than a full redeploy.

Question 21Design and implement an MLOps infrastructure

You want an auditable record of pipeline changes and the ability to rebuild any previously released pipeline. What should the CI process rely on?

  1. Pipeline definitions stored in Git and deployed by the pipeline Correct answer.
  2. A manual export of the workspace every quarter
  3. Screenshots of each pipeline graph in a wiki
  4. A list of pipeline run IDs kept in a spreadsheet

Explanation

Keeping pipeline definitions as code in a repository gives each change a reviewed commit and lets CI recreate the exact released definition. Run IDs and exports describe history but cannot rebuild a definition.

Question 22Design and implement an MLOps infrastructure

A workspace must block outbound data exfiltration from its compute while still reaching an approved private storage account. Which configuration should you apply?

  1. Key-based datastore authentication over the public internet
  2. A managed virtual network with outbound rules limited to approved targets Correct answer.
  3. Public network access with an IP allowlist on the workspace
  4. A compute instance with a public IP and no network security group

Explanation

A managed virtual network places compute in an isolated network where outbound rules restrict egress to approved destinations, and private endpoints let the workspace reach storage without public exposure. Together these prevent data from leaving to unapproved endpoints.

Question 23Design and implement an MLOps infrastructure

You want to measure how a candidate deployment performs on real request payloads without changing what clients receive. Which feature should you use?

  1. Enabling data collection on the endpoint
  2. Enabling mirrored traffic on the candidate deployment Correct answer.
  3. Deploying the candidate to a batch endpoint
  4. Setting the candidate's traffic weight to 50 percent

Explanation

Mirrored traffic sends a copy of live requests to the candidate deployment and discards its responses, so clients still get answers from the production deployment. You obtain realistic latency and error data for the candidate with zero user impact.

Question 24Design and implement an MLOps infrastructure

Training features must be computed once, joined point-in-time correctly for training, and served with low latency at inference. Which design supports this?

  1. A datastore per feature plus a nightly CSV export
  2. A feature store with feature sets and materialization Correct answer.
  3. A notebook that rebuilds the feature table each night
  4. One registered model per feature column

Explanation

A feature store keeps feature definitions in feature sets and materializes them into an offline store for point-in-time correct training joins and an online store for low-latency serving. This prevents training and serving code from computing the same feature differently.

Question 25Design and implement an MLOps infrastructure

Training runs on low-priority nodes keep failing partway through when capacity is reclaimed. Which design makes the pipeline reliable?

  1. Write checkpoints to a datastore so runs can resume Correct answer.
  2. Disable autoscaling and pin the node count
  3. Increase the idle timeout so nodes are never released
  4. Send all training traffic to a canary deployment

Explanation

Low-priority nodes can be preempted at any time, so the workload must assume interruption and persist progress to durable storage. Restarted runs then resume from the last checkpoint instead of starting over.

Question 26Design and implement an MLOps infrastructure

Which declarative template format can deploy an Azure Machine Learning workspace and its dependent resources repeatably across environments?

  1. A run history JSON document
  2. An MLflow model signature
  3. Bicep Correct answer.
  4. A Conda environment YAML file

Explanation

Bicep is an infrastructure-as-code language that declares Azure Resource Manager resources, including a workspace, storage account, key vault, and container registry. Conda files, model signatures, and run histories describe runtime dependencies or metadata rather than deployable Azure resources.

Question 27Design and implement an MLOps infrastructure

A development compute instance is often left running overnight. Which configuration stops it automatically after a period with no interactive activity?

  1. Idle shutdown after a set number of minutes Correct answer.
  2. A quota increase for the subscription
  3. Auto-scaling with a minimum node count of zero
  4. Low-priority scheduling with preemption

Explanation

A compute instance supports an idle shutdown setting that stops the instance after a configured period without activity, which prevents unnecessary charges. Auto-scaling applies to compute clusters, preemption applies to low-priority jobs, and quota changes affect subscription limits.

Question 28Design and implement an MLOps infrastructure

Where should the connection details for an Azure Storage account be defined so that Azure Machine Learning jobs can reference the storage by a friendly name?

  1. In a workspace datastore Correct answer.
  2. In a model's version tags
  3. In a job's metric log
  4. In the environment's Conda dependencies

Explanation

A datastore stores connection information for an Azure Storage account and exposes it to jobs through a workspace-scoped name. Environments, model tags, and metrics are unrelated to storage connection configuration.

Question 29Design and implement an MLOps infrastructure

Which Azure Machine Learning CLI v2 command submits a command job defined in a YAML specification file?

  1. az ml job show --name nightly-train
  2. az ml job list --workspace-name mlw-prod
  3. az ml job create --file job.yml Correct answer.
  4. az ml job stream --name nightly-train

Explanation

The `az ml job create` command accepts a YAML job definition through the `--file` parameter and submits it to the workspace. The other commands list, display, or stream an existing job rather than create one.

Question 30Design and implement an MLOps infrastructure

A GitHub Actions workflow must submit jobs to an Azure Machine Learning workspace without storing a client secret in the repository. Which authentication approach should you use?

  1. A workspace primary key stored in a workflow variable
  2. A federated credential trusting GitHub's OIDC provider Correct answer.
  3. The subscription's built-in administrator account
  4. Anonymous access enabled on the workspace endpoint

Explanation

Workload identity federation lets GitHub Actions exchange a short-lived OIDC token for an Azure access token, so no long-lived secret is stored. Workspace keys are secrets, anonymous access is not an appropriate or generally available option, and built-in administrator credentials are not used for automated workspace authentication.

Question 31Design and implement an MLOps infrastructure

A custom environment is built from a Dockerfile stored alongside the project. Which section of the environment YAML specifies where the Dockerfile and its build context are located?

  1. `build` Correct answer.
  2. `conda_file`
  3. `tags`
  4. `inference_config`

Explanation

The `build` section of an environment definition specifies the build context path and the Dockerfile path used to build the image. The `conda_file` section lists packages, `inference_config` describes deployment behavior, and `tags` are metadata.

Question 32Design and implement an MLOps infrastructure

In an Azure Machine Learning pipeline YAML, a second step needs a file that the first step wrote as a named output. How should the second step reference that file?

  1. Embed the storage account key in the downstream step's command
  2. Reference it with `${{parent.jobs.step1.outputs.model}}` Correct answer.
  3. Create a separate datastore for each step's output folder
  4. Read the upstream step's run history and copy the file manually

Explanation

Pipeline expressions expose an upstream step's named output through the `parent.jobs` hierarchy, and Azure Machine Learning resolves that reference to the output location. This wiring avoids manual copying and does not require storage credentials in the command.

Question 33Design and implement an MLOps infrastructure

Which workspace configuration sends resource and audit logs to Azure Monitor so they can be queried and retained?

  1. A diagnostic setting routing logs to Log Analytics Correct answer.
  2. A compute cluster auto-scaling policy with a maximum node count
  3. A data drift monitor configured with a training baseline
  4. A model registry tag applied to each model version

Explanation

A diagnostic setting on the workspace routes platform logs and metrics to destinations such as a Log Analytics workspace for querying and retention. Drift monitors, autoscaling policies, and model tags serve different purposes.

Question 34Design and implement an MLOps infrastructure

A training pipeline must always read the curated dataset version that was approved for the current release. How should the pipeline reference the data asset?

  1. By the datastore connection string alone, omitting the asset name
  2. By the storage account's public blob URL without a version
  3. By asset name and version, such as azureml:curated_data:3 Correct answer.
  4. By a relative path on the compute node's local disk

Explanation

Referencing a data asset by name and version pins the pipeline to the approved immutable snapshot, such as `azureml:curated_data:3`. An unversioned URL, a connection string without an asset name, or a local path does not guarantee which data contents are used.

Question 35Design and implement an MLOps infrastructure

Which property of an Azure Machine Learning command job specifies the local source directory that is snapshotted and made available to the running script?

  1. `compute`
  2. `environment`
  3. `limits`
  4. `code` Correct answer.

Explanation

The `code` property points to the source directory that Azure Machine Learning captures and mounts for the job, which supports reproducibility. The `environment` property defines dependencies, `compute` selects the target, and `limits` controls runtime constraints.

Question 36Design and implement an MLOps infrastructure

A pipeline step produces a large model file that the next step reads on the same compute target. Which output mode avoids an explicit copy to the datastore before the next step reads it?

  1. `copy` output mode
  2. `rw_mount` output mode Correct answer.
  3. `upload` output mode
  4. `download` output mode

Explanation

The `rw_mount` output mode exposes the output location through a read-write mount, so the step writes to the mounted path without a separate upload operation. The `upload` mode explicitly copies output to the datastore, and `download` is not the output-side mode that avoids that copy.

Question 37Design and implement an MLOps infrastructure

Which reference to a container base image best ensures that rebuilding an environment yields the same image content over time?

  1. A base image pinned by its SHA-256 digest Correct answer.
  2. A tag such as latest so the newest base is always used
  3. A Dockerfile that installs the newest packages at build time
  4. A conda file that lists packages without version pins

Explanation

A digest uniquely identifies an immutable image manifest, so the same base content is pulled on every rebuild. Mutable tags, unpinned packages, and build-time installation of the newest packages all allow content to change between builds.

Question 38Design and implement an MLOps infrastructure

A pipeline references a component that other teams update frequently. Which reference prevents their updates from silently changing your pipeline's behavior?

  1. `azureml:data_prep:2` Correct answer.
  2. `azureml:data_prep@latest`
  3. `azureml:data_prep`
  4. A local file path to the component's source folder

Explanation

A registered component reference that includes an explicit version pins the pipeline to that immutable component definition. A name-only reference resolves to the latest version, and a local source path does not use a governed registered version.

Question 39Implement machine learning model lifecycle and operations

You register a model named churn-model in an Azure Machine Learning workspace, and a model with that name already exists. What is the result of the new registration?

  1. A new model named churn-model-2 is created automatically.
  2. The registration fails with a name conflict error.
  3. A new version of churn-model is created with an incremented version number. Correct answer.
  4. The existing registration is overwritten and keeps its version number.

Explanation

The model registry treats a model name as a container and each registration under that name as an immutable version. Registering again with the same name adds the next version and leaves earlier versions intact, so old deployments keep working.

Question 40Implement machine learning model lifecycle and operations

Which setting connects an MLflow experiment run to an Azure Machine Learning workspace?

  1. The AZUREML_WORKSPACE_ID environment variable
  2. MLFLOW_REGISTRY_URI set to the workspace artifact store
  3. MLFLOW_TRACKING_URI set to the workspace tracking URI Correct answer.
  4. An AZUREML_DATAREFERENCE pointing to a workspace datastore

Explanation

MLflow clients send metrics, parameters, and artifacts to the URI in MLFLOW_TRACKING_URI. Pointing that variable at the workspace's Azure ML tracking URI makes runs appear in the workspace's experiments.

Question 41Implement machine learning model lifecycle and operations

An application must return predictions within 100 milliseconds per request for interactive users. Which Azure Machine Learning hosting option should serve the model?

  1. A batch endpoint
  2. A batch endpoint with a mini-batch size of one
  3. A command job on a compute cluster
  4. A managed online endpoint Correct answer.

Explanation

Online endpoints keep a model resident in a container behind an HTTPS scoring URI and respond synchronously, which suits low-latency interactive calls. Batch endpoints run asynchronous jobs over datasets and are not designed for per-request latency targets.

Question 42Implement machine learning model lifecycle and operations

You want 5 percent of live requests to reach a new deployment while the current deployment serves the rest. What should you configure on the online endpoint?

  1. A rollback rule on the new deployment
  2. Traffic allocation that assigns 5 percent to the new deployment and 95 percent to the existing one Correct answer.
  3. A mirroring policy between the two deployments
  4. An autoscale rule that limits the new deployment to 5 percent of instances

Explanation

An online endpoint routes requests across its deployments according to per-deployment traffic percentages. Setting 5 and 95 implements a canary release without changing client URLs.

Question 43Implement machine learning model lifecycle and operations

A client application must authenticate to a managed online endpoint by using a credential that expires automatically after a short period. Which authentication method should it use?

  1. Key-based authentication with the endpoint primary key
  2. Token-based authentication with a Microsoft Entra ID token Correct answer.
  3. A workspace connection secret stored in Azure Key Vault
  4. A shared access signature issued by the workspace storage account

Explanation

Endpoints can require either keys or tokens. Keys are static credentials that remain valid until regenerated, whereas tokens are short-lived Microsoft Entra ID credentials suited to automatic expiry.

Question 44Implement machine learning model lifecycle and operations

What does an Azure Machine Learning environment define for a model deployment?

  1. The number of instances and autoscale rules for the endpoint
  2. The runtime image and dependencies used to run the scoring script Correct answer.
  3. The model version and traffic percentage for the deployment
  4. The TCP port and hostname of the scoring container

Explanation

An environment captures the base image, Python packages, and environment variables needed to execute code. Deployments bind a model, a scoring script, and an environment together so the container is reproducible.

Question 45Implement machine learning model lifecycle and operations

Which Azure CLI command creates a new deployment on a managed online endpoint from a YAML specification?

  1. az ml endpoint create
  2. az ml job create
  3. az ml model deploy
  4. az ml online-deployment create Correct answer.

Explanation

The Azure ML CLI v2 groups endpoint resources under online-endpoint and online-deployment commands. The deployment create command accepts a YAML file that describes the model, environment, scoring script, and instance settings.

Question 46Implement machine learning model lifecycle and operations

In the scoring script used by a managed online endpoint, which function is invoked to load the model when the container starts?

  1. predict(data)
  2. score(data)
  3. init() Correct answer.
  4. run(data)

Explanation

The init function runs once when the container starts and is the place to load the model into memory. The run function is then called for each request to produce predictions.

Question 47Implement machine learning model lifecycle and operations

You want Azure ML model monitoring to compare production inference data against the training dataset. Which configuration element supplies the reference distribution?

  1. An alert email action group
  2. A baseline dataset for the data drift signal Correct answer.
  3. The endpoint traffic allocation
  4. The deployment environment version

Explanation

The data drift signal compares the recent production data with a baseline dataset that represents the training distribution. The baseline is a required input for the drift calculation.

Question 48Implement machine learning model lifecycle and operations

Which setting must be enabled on a managed online deployment so that Azure ML model monitoring can analyze the data sent to it?

  1. Token authentication on the endpoint
  2. Data collection on the deployment Correct answer.
  3. A batch endpoint with a default deployment
  4. Autoscaling on the endpoint

Explanation

Model monitoring for online endpoints relies on inference data captured by the model data collector. Enabling data collection stores request and response payloads that monitoring signals later analyze.

Question 49Implement machine learning model lifecycle and operations

You need to view request latency and failed request traces for a managed online endpoint. What should you enable?

  1. A model data collector with a 100 percent sampling rate
  2. An Event Grid subscription on the workspace
  3. Application Insights integration for the endpoint Correct answer.
  4. Diagnostic settings that send job logs to a storage account

Explanation

Managed online endpoints emit request metrics and traces to Application Insights when integration is enabled. That telemetry supports latency charts and failed request investigation in the Azure portal.

Question 50Implement machine learning model lifecycle and operations

You must run a retraining pipeline every Monday at 02:00 UTC. What should you create?

  1. A compute cluster with a minimum of one node
  2. An event-driven trigger on the workspace blob container
  3. A deployment with a traffic rule
  4. A job schedule with a cron expression for Monday 02:00 UTC Correct answer.

Explanation

Azure ML schedules run jobs or pipelines on a recurrence defined by a cron expression. A weekly cron entry targeting Monday 02:00 UTC produces the required cadence.