Microsoft AI-200 - Developing AI Cloud Solutions on Azure

The batch embedding workload is planned to run in Azure Kubernetes Service (AKS) , and Kubernetes workloads should be defined declaratively by using YAML manifest files . Microsoft states that a Kubernetes manifest defines the desired state of cluster resources, including Deployments, Jobs, Services, ConfigMaps, and other Kubernetes objects. These manifests can then be deployed by using kubectl apply -f < manifest > .yaml.

For this case, YAML is particularly appropriate because the technical requirements emphasize repeatable, version-controlled, automated deployments . A batch vector embedding regeneration process would typically be represented as a Kubernetes workload such as a Job or another batch-oriented resource, with its configuration committed to Git alongside the rest of the deployment definitions.

kubectl run and kubectl create are imperative commands and are less appropriate for a controlled Git-based deployment model because they do not inherently preserve the complete desired-state definition as source-controlled configuration. XML is not the standard Kubernetes manifest format. az aks commands are primarily used to provision and manage the AKS cluster itself rather than declaratively define application workloads running inside the cluster.

Therefore, the correct choice is YAML-formatted Kubernetes manifest files .

Study Guide references: Azure Kubernetes Service → Kubernetes manifests; declarative deployments; kubectl apply; Git-based workload configuration.

Comply with the authentication policy for database access: Use managed identity authentication.

Support high-concurrency requests with minimal latency: Use a connection pooling library.

Protect database stability during traffic spikes: Configure maximum pool size.

The Recommendation API should use managed identity authentication because Fabrikam explicitly requires service-to-database authentication without long-lived credentials. Azure Container Apps managed identities allow applications to obtain Microsoft Entra tokens at runtime and access supported Azure resources without storing usernames, passwords, or connection secrets in application configuration.

For high-concurrency PostgreSQL access, implement a connection pooling library . Azure Database for PostgreSQL documentation recommends connection pooling because repeatedly opening new database connections creates backend processes and consumes CPU and memory. Reusing existing connections reduces connection-establishment overhead and improves latency and throughput. Azure also provides PgBouncer as a built-in pooling option for Flexible Server.

To protect PostgreSQL during sudden traffic spikes, configure a maximum pool size . An unlimited pool can allow application demand to create excessive concurrent connections, which can exhaust database memory and increase contention. Microsoft warns that increasing connection counts can cause significant performance degradation and recommends conservative connection limits combined with pooling.

Therefore, the optimal configuration combines managed identity + connection pooling + bounded pool size .

Study Guide references: Azure Container Apps managed identities; Azure Database for PostgreSQL connection management; PgBouncer; connection-pool sizing and resource protection.

Verified Answer: 1) Instrument the application code with the OpenTelemetry SDK; 2) configure an Azure Monitor trace exporter in the OpenTelemetry SDK; 3) redeploy the instrumented services.

Detailed Explanation: Trace correlation requires the application to create OpenTelemetry spans and an exporter to send those spans to Azure Monitor. Instrumentation must therefore be present before the updated service is deployed. The previous Application Insights SDK instrumentation does not satisfy the case-study requirement that all tracing use OpenTelemetry. Configuring an OpenTelemetry view is unrelated to basic distributed-trace export, and calling TrackEvent is an Application Insights SDK pattern rather than the required OpenTelemetry approach.

Study Guide Alignment: Security and operations: Key Vault, App Configuration, managed identity, OpenTelemetry, Azure Monitor, and KQL-based troubleshooting.

Official Microsoft Learn References: AI-200 Study Guide | Enable Azure Monitor OpenTelemetry

Trigger scaling based on queue depth: Azure Service Bus scaler

Deallocate instances when the message queue is empty: Minimum replicas = 0

Prevent uncontrolled burst scaling: Maximum replicas = 10

The batch retraining workload must scale according to queue length , so the correct KEDA trigger is the Azure Service Bus scaler . KEDA integrates with Azure Service Bus and monitors queue or subscription metrics, allowing AKS workloads to scale according to the number of pending messages rather than CPU utilization or HTTP request concurrency. Microsoft documents KEDA as the event-driven autoscaling mechanism for AKS and specifically supports Azure-based scalers for queue-driven workloads.

To meet the requirement to deallocate workload instances when no messages remain, configure minimum replicas = 0 . KEDA supports scale-to-zero for event-driven workloads, which is particularly appropriate for intermittent batch or machine-learning processing because no pods need to remain running while the queue is empty. Microsoft specifically identifies scale-to-zero as a KEDA capability and recommends it for sporadic ML workloads where cost optimization is important.

Finally, configure maximum replicas = 10 . A finite maximum replica count establishes an explicit upper scaling boundary, preventing large queue bursts from creating an uncontrolled number of worker replicas and protecting both infrastructure cost and downstream services.

Study Guide references: AKS → KEDA; Azure Service Bus event-driven scaling; minReplicaCount; maxReplicaCount; scale-to-zero; autoscaling limits.

The correct solution is to implement connection pooling , ideally by using the built-in PgBouncer capability in Azure Database for PostgreSQL Flexible Server. Connection pooling allows many concurrent application requests to share and reuse a smaller set of established PostgreSQL server connections instead of repeatedly opening and closing database sessions. Microsoft notes that creating a new PostgreSQL connection for each operation consumes server resources because each connection involves its own backend process. Reusing pooled connections therefore reduces connection overhead, lowers CPU and memory consumption, and improves throughput for high-concurrency workloads.

This directly matches Fabrikam ' s stated requirement that database connections support high concurrency with minimal latency through connection optimization . PgBouncer can efficiently multiplex large numbers of client connections onto fewer PostgreSQL connections, particularly in transaction pooling mode.

Increasing max_connections is not the preferred solution because each additional PostgreSQL connection consumes memory and other resources; Microsoft explicitly recommends PgBouncer instead when more concurrent connections are required. Increasing shared_buffers addresses caching and memory allocation rather than connection-management overhead. Read replicas can scale read workloads, but they do not optimize the application ' s connection lifecycle.

Study Guide references: Azure Database for PostgreSQL Flexible Server → connection pooling; PgBouncer; connection management and high-concurrency performance.

Detailed Explanation: Azure Key Vault references in App Service settings satisfy the requirement to keep secrets out of container images, source control, and directly stored application configuration. App Service resolves the referenced secret at runtime by using the app identity, so the application can consume the value as a normal setting without embedding credentials in the image. GitHub secrets are build/deployment secrets rather than a runtime App Service secret-delivery mechanism. Dockerfile ENV instructions would place secret material in the image configuration and violate the case requirements.

Study Guide Alignment: Security and operations: Key Vault, App Configuration, managed identity, OpenTelemetry, Azure Monitor, and KQL-based troubleshooting.

Official Microsoft Learn References: AI-200 Study Guide | Use Key Vault references for App Service and Functions | Managed identities for Azure resources

Verified Answer: 1) Create the Service Bus client; 2) create the queue processor/receiver; 3) register message and error handlers; 4) start the message processor.

Detailed Explanation: The processor depends on a namespace client/connection, so the client is created first. Next, the queue-specific processor or receiver is created from that client. Message and error callbacks must be registered before processing starts so incoming messages and failures have defined handlers. Starting the processor is therefore the final step. Dead-lettering failed messages is an optional application settlement decision and is not a prerequisite for constructing and starting the processor.

Study Guide Alignment: Azure service integration: Service Bus, Event Grid, Azure Functions triggers/bindings, and event-driven processing.

Official Microsoft Learn References: AI-200 Study Guide | Service Bus message transfers, locks, and settlement

Meet the 200-ms semantic search latency requirement: Increase compute vCores.

Optimize the environment for high-dimensional pgvector index residency: Increase memory allocation.

Support the continuous ingestion of transaction-based embeddings: Enable storage autoscale.

For the strict sub-200-ms vector-search latency target, increasing compute vCores is the appropriate choice. Vector similarity operations are computationally intensive, and additional CPU capacity improves mathematical throughput and parallel query execution. Microsoft’s pgvector guidance emphasizes query-plan optimization, ANN indexes such as HNSW, and sufficient compute resources when optimizing vector workloads.

For high-dimensional pgvector index residency , increase memory allocation . HNSW provides strong query-performance characteristics but consumes more memory than IVFFlat. Keeping frequently accessed vector index structures in memory minimizes disk access and materially improves latency. Microsoft explicitly notes that HNSW requires more memory while providing a better speed/recall tradeoff.

For the continuous ingestion of millions of embeddings, enable storage autoscale . Azure Database for PostgreSQL Flexible Server can automatically increase allocated storage as capacity approaches configured thresholds, avoiding an out-of-storage condition as data volumes grow. Microsoft recommends storage autogrow for workloads whose storage demand can increase dynamically.

Increasing max_connections does not directly improve vector computation or index residency, read replicas primarily scale reads, and backup retention does not address ingestion capacity.

Study Guide references: Azure Database for PostgreSQL Flexible Server → pgvector performance optimization; compute and memory sizing; HNSW indexing; storage autogrow.

Verified Answer: Identify document changes: Change feed processor. Scale out vectorization processing: Lease container.

Detailed Explanation: The change feed processor is designed to react to inserts and updates in a monitored Cosmos DB container, which directly matches the requirement to generate embeddings for new or changed documents. Its lease container stores processing state and coordinates work across multiple workers, allowing the processing workload to scale out without duplicating ownership of change-feed ranges. A periodic full scan would consume unnecessary RUs, and neither strong consistency nor container RU throughput is the coordination mechanism for distributed change-feed workers.

Study Guide Alignment: AI data-management workloads: Cosmos DB, PostgreSQL, caching, vector storage, vector retrieval, consistency, and connection optimization.

Official Microsoft Learn References: AI-200 Study Guide | Azure Cosmos DB change feed processor