Free AI-200 practice test: 30 questions with answers

Vector search in Cosmos DB for NoSQL requires a vector embedding policy that declares the path, dimensions, and distance function, plus a matching vector index. Multi-region writes affect availability, not similarity search, and the change feed only streams changes.

Session consistency gives read-your-writes and monotonic reads within a client session at a lower cost than strong consistency. Eventual has no ordering guarantees within a session, and strong consistency is the most expensive option for reads.

The change feed processor reads changes in order and stores its progress in a lease container, so it can resume after a restart. Analytical store is aimed at large-scale analytics, and a trigger inside the database does not give a durable resumable consumer.

pgvector supports approximate nearest-neighbour indexes such as HNSW and IVFFlat, and the index must use the same distance operator as the query for it to be used. Converting to text or JSON removes the vector operators, and disabling autovacuum leads to bloat rather than faster search.

Pooling reuses established server connections, both at an external pooler and in an application-side pool, so the setup cost is paid once rather than per request. Raising max_connections adds memory pressure without removing setup cost, and the timeout and dimension settings do not affect connection establishment.

Azure Managed Redis supports key expiration and explicit deletion, which covers both time-based and event-driven invalidation of cached answers. Blob lifecycle rules manage object tiers and deletion, and Front Door and Azure Files do not provide application-level cache invalidation.

Similarity lookup needs vector fields with an index that supports nearest-neighbour search, so the application can find cached embeddings close to the new query. Pub/Sub and keyspace notifications deliver messages about changes and do not search by similarity.

Point reads are the cheapest operation in Cosmos DB because they go straight to one item, and an indexing policy that excludes unused paths lowers the write and storage cost of index maintenance. Cross-partition queries fan out and cost more, and removing the partition key breaks scalability.

Key Vault stores the secret centrally, supports rotation, and the app reads the current version at runtime using a managed identity without credentials in code. Hard-coded or image-embedded secrets cannot be rotated without redeploying.

Azure App Configuration stores key-value settings and feature flags centrally and supports dynamic refresh. Key Vault is for secrets and certificates, not general configuration, and the other two services have unrelated purposes.

OpenTelemetry provides distributed tracing, and propagating trace context through HTTP headers and message properties links spans across services into one trace. Separate log files or resource-level metrics do not connect a single request across components.

KQL filters the time range and failure status, then summarizes counts into five-minute bins. The other options use SQL syntax, a non-query command, or a PowerShell cmdlet that does not query telemetry.

Kubernetes events show reasons such as OOMKilled or liveness probe failures, and logs from the previous container instance show the error that caused the crash. Billing, DNS zones, and branch rules do not explain container restarts.

When a vault uses Azure RBAC for data access, the identity must hold a data-plane role such as Key Vault Secrets User. The old access policy model no longer grants access in that mode.

App Configuration feature flags can be evaluated at runtime, and the app can refresh settings without a restart or redeploy. Key Vault, Blob versioning, and registry webhooks serve different purposes.

A managed identity lets the function obtain Entra ID tokens without storing credentials, and access can be granted per resource. Stored passwords and account keys are the credentials the question is trying to avoid.

ACR Tasks builds and pushes container images inside the registry, and can be triggered by a source code commit or a base image update. Azure Image Builder produces virtual machine images, not container images, so it is the main distractor.

A Key Vault reference in an app setting is resolved by App Service at runtime using the app's managed identity, so the value never sits in the app configuration. Embedding the secret in the image or in wwwroot exposes it to anyone who can read the image or the files.

Container Apps implements event-driven scaling with KEDA scalers, and the Service Bus scaler reads queue length to set the replica count, including scale to zero. CPU-based autoscaling does not react to queue backlog, and a Kubernetes HPA manifest is not how Container Apps scale rules are declared.

Multiple revision mode keeps several revisions active so traffic weights can be split between them, and revision labels give a stable URL to a specific revision for targeted testing. Single revision mode replaces the active revision, and changing a value in the template creates a new revision rather than being a way to skip one.

A Service gives a stable virtual IP and DNS name and load balances across the pods selected by its label selector. A ReplicaSet maintains pod count but provides no stable endpoint, and a ConfigMap only holds configuration data.

az containerapp logs show streams console logs for a named container, and --follow keeps the stream open. The activity log records control plane operations, not container output, and the other two commands do not return container logs.

Unique immutable tags, such as the commit SHA or build number, make each deployment traceable to a specific build. A mutable latest tag can point to different images over time, which breaks reproducibility.

A topic with subscriptions gives each consumer its own durable copy, and messages that exceed the max delivery count move to that subscription's dead-letter queue for inspection. Event Hubs is built for streaming telemetry, and Storage queues lack topic subscriptions and built-in dead-letter handling of that kind.

Dead-lettering moves the message out of the active queue and lets you attach a reason and description for review. Abandoning returns the message for redelivery, and completing it removes the message without preserving it for inspection.

Event Grid event subscriptions support subject and event type filters, so non-matching events are never delivered to the handler. host.json controls function host behaviour, and lifecycle policies and Key Vault policies do not filter Event Grid events.

An HTTP trigger starts the function on incoming requests, and the response can be returned as the function output. Timer and blob triggers start on a schedule or blob changes, not on HTTP requests.

Triggers and bindings let the runtime handle the Service Bus receive and Cosmos DB write, so the function focuses on business logic. Calling ARM is for management operations and is not how data bindings are used.

Service Bus subscriptions support SQL filters that evaluate message properties, so only matching messages are delivered to the subscription. Sessions group related messages for ordered processing and do not select messages by property.

Peek-lock keeps the message locked until the consumer completes it, so an abandoned or crashed lock returns the message for redelivery. Receive-and-delete removes the message as soon as it is read, which risks losing it.