AI-200 Hard Questions — 25 Difficult Questions

Question 1

You are designing an enterprise cloud solution that integrates Azure Event Grid, Azure Service Bus, and Azure Event Hubs. You need to select the appropriate service for three specific event-driven scenarios. Which THREE pairings of Azure messaging service to scenario are correct? (Choose three)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Azure Service Bus for managing financial transactions requiring strict ordering, sessions, and dead-lettering.. Option A is correct because Azure Service Bus is the enterprise brokered-messaging service that natively supports FIFO ordering via sessions, message sessions for correlated/ordered processing, and built-in dead-letter queues for poison messages, which are exactly the requirements for financial transactions. Option B is correct because Azure Event Hubs is designed for massive-scale event ingestion (millions of events per second, with throughput units or processing units) and is the right fit for 500,000 telemetry events per second from smart grid sensors. Option C is correct because Azure Event Grid is a reactive, push-based event routing service that natively emits blob-created events from Azure Storage and can deliver them to serverless endpoints such as Azure Functions for AI processing. Option D is not correct because Event Hubs is a high-throughput streaming ingestion service, not a low-latency pub/sub webhook trigger for discrete user profile updates; Event Grid is the appropriate service for that scenario. Option E is not correct because Event Grid is not built for continuous high-volume telemetry streaming into analytics engines like Databricks; Event Hubs (often with Capture or Stream Analytics) is the correct ingestion path for that workload.

Answer

Azure Event Hubs for triggering real-time serverless webhooks upon user profile updates.

Answer

Azure Event Grid for streaming continuous IoT sensor telemetry into Azure Databricks.

Question 2

You are integrating Azure Event Hubs with Apache Spark on Azure Databricks for real-time AI model scoring. You need to configure the Spark connector to read from all partitions of the Event Hub efficiently without missing records or creating duplicate processing bottlenecks. Which consumer property is critical to configure correctly?

Accepted Answer

A dedicated Consumer Group and correct starting offset configuration (e.g., startingPosition).. When reading from Azure Event Hubs with the Spark connector, the critical consumer-side configuration is a dedicated Consumer Group plus a correct starting offset (e.g., startingPosition set to earliest, latest, or a specific offset). This ensures each Spark job reads all partitions exactly once without duplicating another consumer's progress or skipping records.

Answer

Setting the Event Grid webhook handshake timeout to 120 seconds.

Answer

Enabling AMQP 1.0 web socket transport mode across all Spark worker nodes.

Answer

The Service Bus shared access policy connection string with Manage rights.

Question 3

You are configuring a vector index in Azure AI Search and want to apply scalar quantization to reduce the memory footprint. What is the primary effect of scalar quantization on float32 vector data?

Accepted Answer

It compresses float32 vectors to int8 integers, reducing memory usage by approximately 75%. Scalar quantization in Azure AI Search compresses each float32 component of a vector into an 8-bit integer (int8), reducing storage from 4 bytes to 1 byte per dimension — roughly a 75% memory reduction. This allows larger vector indexes to fit in memory and lowers cost, with only a small recall degradation that is typically acceptable for most RAG workloads.

Answer

It increases vector dimensions from 1536 to 3072 to improve accuracy

Answer

It eliminates the need for an HNSW graph structure

Answer

It converts vectors into plain text strings for keyword indexing

Question 4

You are troubleshooting a slow vector query performance issue in Azure Cosmos DB for NoSQL. Which THREE actions can you take to improve vector search execution times? (Choose three)

Accepted Answer

Use the diskANN indexing type instead of flat indexes for large collections. Option A is correct because diskANN is the approximate nearest neighbor (ANN) index type in Cosmos DB for NoSQL designed for large-scale vector search; it dramatically reduces latency and RU cost compared to flat (exact, brute-force) indexes, which scan every vector. Option B is correct because effective partition key choice spreads vector data and query load across physical partitions, enabling parallel execution and avoiding hot partitions that throttle vector query throughput. Option E is correct because applying filter conditions (e.g., WHERE on scalar properties) narrows the candidate set before vector distance ranking, so fewer vectors are compared and execution time drops. Option C is not correct because base64 encoding is merely a serialization format for storing vectors; it does not improve search execution time and can add conversion overhead. Option D is not correct because disabling all non-vector indexing policies removes indexes that support the filters and predicates used to pre-filter vector queries, which typically degrades rather than improves performance.

Answer

Convert all vector embeddings to base64 string format

Answer

Disable all non-vector indexing policies on the container

Question 5

You are designing an enterprise event-driven architecture where microservices across different Azure subscriptions need to subscribe to domain events published by a central Order Service. Which Azure Event Grid architecture should you implement?

Accepted Answer

Deploy an Azure Event Grid Event Domain and create event subscriptions under the domain for each consuming team.. Azure Event Grid Event Domains are purpose-built for multi-tenant, multi-subscription event distribution scenarios. An Event Domain acts as a namespace that can host thousands of topics, allowing each consuming team or subscription to create its own event subscriptions under the domain without the publisher needing to manage individual topics per consumer. This decouples the central Order Service from consumer onboarding and scales cleanly across subscriptions.

Answer

Deploy an Azure Service Bus relay across all subscriptions.

Answer

Create individual Event Hubs namespaces in every consumer subscription and peer them via ExpressRoute.

Answer

Use Azure Storage Queues shared via SAS URLs across subscriptions.

Question 6

Your AI processing pipeline requires high-throughput streaming ingestion and retention of telemetry data. Which THREE features or capabilities are native to Azure Event Hubs? Each correct answer represents a complete capability.

Accepted Answer

Event Hubs Capture for automatically batching and saving stream data to Azure Storage. Option A is correct because Event Hubs Capture is a native feature that automatically batches streaming data and writes it to Azure Blob Storage or Azure Data Lake Storage (in Avro format) based on a time or size window, satisfying the retention requirement without custom code. Option C is correct because Consumer Groups are a native Event Hubs construct that lets multiple independent applications each maintain their own offset and read the same event stream independently, which is essential for multiple downstream consumers. Option D is correct because Event Hubs partitions the stream and allows multiple receivers to process partitions in parallel, which is the foundation of its high-throughput streaming ingestion and parallel processing model. Option B is not correct because Event Hubs does not natively push events via webhooks for serverless triggers; that pattern belongs to Event Grid, whereas Event Hubs uses AMQP/Kafka pull-based consumption. Option E is not correct because automatic dead-lettering based on a MaxDeliveryCount threshold is a Service Bus queue/subscription feature, not a native Event Hubs capability.

Answer

Native webhook push delivery for discrete serverless event triggers

Answer

Automatic dead-lettering of messages exceeding a MaxDeliveryCount threshold

Question 7

You are investigating an incident where an AI model endpoint experienced a spike in memory consumption, resulting in Out-Of-Memory (OOM) container crashes. You want to query Log Analytics to find the exact container memory working set percentage across all replicas during the past 2 hours. Which table and KQL query should you use?

Accepted Answer

InsightsMetrics | where TimeGenerated > ago(2h) | where Namespace == 'container' and Name == 'MemoryWorkingSetBytes' | summarize avg(Val) by Computer. Container metrics in Azure Monitor are stored in the `InsightsMetrics` table where `Namespace == 'container'` and `Name == 'memoryWorkingSetBytes'` or similar performance counters.

Answer

SecurityEvent | where TimeGenerated > ago(2h) | summarize count() by Account

Answer

AzureActivity | where TimeGenerated > ago(2h) | summarize count() by OperationName

Answer

AppTraces | where TimeGenerated > ago(2h) | summarize count() by Message

Question 8

You are designing a data layer architecture in Azure Cosmos DB for NoSQL for an AI application that performs high-frequency writes and real-time vector searches. You need to understand how vector indexing impacts write operations. What is the primary trade-off?

Accepted Answer

Writing documents consumes more Request Units (RUs) because vector indexes must be updated. Writing documents consumes more Request Units (RUs) because vector indexes must be updated on every write. In Azure Cosmos DB for NoSQL, vector indexing adds overhead to write operations since the index must be maintained as documents are inserted or updated. This is the primary trade-off: faster vector search at the cost of higher write RU consumption.

Answer

Vector properties cannot be updated after initial document creation

Answer

Writes are completely blocked until the vector index reaches 100% synchronization

Answer

Writes require explicit distributed transactions across regions

Question 9

You are debugging a Python application integrated with Application Insights. You notice that custom telemetry events generated in a background worker thread are not appearing in Application Insights. What is the most likely cause of this issue?

Accepted Answer

The application terminated or the worker thread exited before the telemetry client buffer was flushed. The Application Insights Python SDK buffers telemetry in memory and sends it asynchronously in batches. If the background worker thread exits or the application terminates before the buffer is flushed, the queued telemetry events are lost and never reach Application Insights. This is the most common cause of missing custom events from short-lived threads.

Answer

Log Analytics ingestion rate limiting always drops multi-threaded telemetry

Answer

Application Insights automatically blocks all telemetry generated outside of HTTP request threads

Answer

The instrumentation key was incorrectly formatted as a GUID

Question 10

You are managing an Azure Cosmos DB for NoSQL database containing vector embeddings. You need to perform a migration of embeddings from one container to another without taking the application offline. Which Azure feature should you use?

Accepted Answer

Azure Cosmos DB Change Feed processor. The Azure Cosmos DB Change Feed processor provides an ordered, continuous stream of inserts and updates from a source container, which a consumer can read and write into a target container in near real time. This enables online migration of vector embeddings without taking the application offline, since the source container keeps serving reads and writes while the change feed replicates data. It is the canonical pattern for live container-to-container migration in Cosmos DB for NoSQL.

Answer

Azure Data Factory tumbling window triggers over Blob Storage

Answer

SQL Server Integration Services (SSIS) file import

Answer

Taking a manual offline portal backup and restoring to a new account

Question 11

You are optimizing an enterprise Azure Cosmos DB for NoSQL vector database. Which THREE administrative or architectural practices help ensure high availability and predictable performance? (Choose three)

Accepted Answer

Choose a partition key that evenly distributes data and query load across physical partitions. Option A is correct because a well-chosen partition key that evenly distributes data and query load across physical partitions prevents hot partitions, which is essential for predictable performance and scalability in Cosmos DB for NoSQL. Option B is correct because enabling multi-region write replication (multi-master) provides global low-latency writes and improves availability by allowing the account to continue serving requests even if a region becomes unavailable. Option E is correct because provisioning adequate Request Units (RUs) or configuring autoscale throughput ensures the database has enough throughput capacity to handle workload demands, preventing throttling (429 errors) and maintaining predictable performance. Option C is incorrect because storing all documents in a single logical partition without a partition key creates a hot partition and limits scalability, harming both availability and performance. Option D is incorrect because disabling automatic indexing entirely removes the default indexing policy that supports efficient queries, which would degrade query performance rather than improve it.

Answer

Store all documents in a single partition without a partition key

Answer

Disable automatic indexing entirely across all properties

Question 12

You are configuring semantic ranking in Azure AI Search to improve search relevance for your enterprise application. Which tier of Azure AI Search is required to enable semantic ranking?

Accepted Answer

Standard (S1 or higher). Semantic ranking in Azure AI Search requires the Standard tier (S1) or higher — it is not available on Free or Basic tiers. Semantic ranking uses Microsoft's language models to rerank the top results from BM25 or vector search, improving relevance for natural-language queries. Enabling it on a supported tier is done by adding a semantic configuration to the index and setting queryType to 'semantic'.

Answer

Storage Optimized (L1)

Answer

Free

Answer

Basic

Question 13

You are troubleshooting a latency issue with an AI model deployed in a Virtual Machine. You suspect the issue is related to storage IOPS limits. Which tool should you use to check the storage performance metrics?

Accepted Answer

Azure Monitor Metrics for the Virtual Machine.. Azure Monitor 'Metrics' for the Virtual Machine resource provides 'Disk Read Operations/sec' and 'Disk Write Operations/sec', which are key for diagnosing IOPS bottlenecks.

Answer

Azure Network Watcher.

Answer

Azure Resource Graph Explorer.

Answer

Application Insights Dependency logs.

Question 14

Your containerized AI application requires a specific version of a kernel driver that is not installed on the AKS nodes. What is the most efficient way to manage this?

Accepted Answer

Use a privileged InitContainer to load the driver.. A privileged InitContainer runs before the main application container and, with privileged security context, can load kernel modules (e.g., via modprobe or insmod) onto the node. This is the standard Kubernetes pattern for node-level setup that must complete before the workload starts. It is more efficient than rebuilding the cluster and avoids granting the application container itself elevated privileges.

Answer

Mount the /proc filesystem as read-write.

Answer

Run the application as root.

Answer

Rebuild the AKS cluster.

Question 15

You are designing an AI architecture that stores vector embeddings in Azure Cosmos DB for NoSQL. You need to implement role-based access control (RBAC) so that specific application microservices can only execute read-only queries against the database without having permission to write or delete documents. How should you configure this?

Accepted Answer

Create a custom SQL role definition with read-only data actions and assign it to the microservice's managed identity. Azure Cosmos DB for NoSQL has a native RBAC system built on SQL role definitions and role assignments. A custom role definition can be scoped to specific data-plane actions such as 'Microsoft.DocumentDB/databaseAccounts/readMetadata' and 'Microsoft.DocumentDB/databaseAccounts/sqlDatabases/containers/items/read', granting read-only access. Assigning that role to the microservice's managed identity enforces least privilege without exposing keys.

Answer

Configure Azure Blob Storage RBAC permissions on the Cosmos DB container

Answer

Share the primary master master-key with all microservices

Answer

Disable authentication on the Cosmos DB account during query execution

Question 16

You are designing a secure enterprise event-driven architecture. Events published to an Azure Event Grid Domain must be partitioned based on a tenant ID. How should publishers send events to the Event Grid Domain to ensure proper routing to domain topics?

Accepted Answer

Publish events directly to the Event Grid Domain endpoint and specify the subject or topic name within the event structure.. Event Grid Domains are designed as a single publishing endpoint that fans out to thousands of domain topics. Publishers send events to the domain endpoint and set the 'subject' field (or use the topic name in the URL path) so Event Grid routes each event to the correct domain topic based on the tenant ID. This is the native partitioning mechanism for Event Grid Domains and avoids per-tenant endpoint sprawl.

Answer

Send all events to an Azure Service Bus Topic, then configure a logic app to forward them to Event Grid.

Answer

Use Azure Event Hubs capture keys to route events to tenant-specific storage containers.

Answer

Create separate Event Grid custom topics for each tenant and configure an Azure API Management gateway to multiplex the requests.

Question 17

Your enterprise architecture uses Azure Service Bus Topics and Subscriptions with geo-disaster recovery paired namespaces. A disaster occurs, and you trigger a failover to the secondary region. What happens to the messages currently in the subscriptions of the primary namespace?

Accepted Answer

Messages remaining in the primary namespace are not automatically replicated to the secondary namespace; failover relies on metadata synchronization.. Service Bus geo-disaster recovery (Geo-DR) pairs a primary and secondary namespace and continuously replicates only the namespace metadata — queues, topics, subscriptions, filters, and authorization rules. Message data itself is not replicated, so messages sitting in the primary namespace's subscriptions at the moment of failover are not available in the secondary namespace and are effectively stranded until the primary is restored.

Answer

All queued messages are automatically migrated to the secondary region within 5 seconds.

Answer

Azure Event Grid automatically replays all lost messages from the storage capture.

Answer

The primary namespace messages are permanently deleted and cannot be recovered.

Question 18

You are designing a high-availability RAG architecture where search queries must succeed even if an entire Azure region experiences an outage. You deploy Azure AI Search in a primary region and a secondary region. How should client applications handle failover?

Accepted Answer

Use Azure Traffic Manager or Azure Front Door in front of regional search service endpoints with health probes. Azure Traffic Manager and Azure Front Door are the correct global load-balancing services to place in front of regional Azure AI Search endpoints, using health probes to detect regional outages and route client traffic to a healthy secondary region. Traffic Manager operates at the DNS layer with priority or performance routing, while Front Door provides anycast-based Layer 7 routing with faster failover and additional WAF/TLS offload. Both support health probe configuration that automatically removes an unhealthy regional endpoint from rotation.

Answer

Hardcode the secondary search IP address in client application configuration files

Answer

Rely on Cosmos DB automatic failover to replicate search indices

Answer

Configure SQL replication between search indexes

Question 19

You are deploying a custom AI model container to Azure Container Apps. The container listens on port 8080 and requires health probes to verify readiness before routing traffic. How should you configure the probes?

Accepted Answer

Configure an HTTP readiness probe targeting port 8080 on your health check endpoint. Azure Container Apps supports HTTP and TCP probes defined at the container level, and an HTTP readiness probe is the correct mechanism to signal that the app is ready to receive traffic on its listening port. Since the container listens on port 8080 and exposes a health check endpoint, an HTTP readiness probe targeting that port lets the platform delay traffic routing until the endpoint returns a success status code. This matches the Kubernetes-style probe model that Container Apps implements, where readiness probes control traffic admission rather than process restart.

Answer

Configure an Azure Front Door health probe pointing to the internal container IP

Answer

Configure a TCP liveness probe pointing to port 22 for SSH management

Answer

Configure a startup probe that restarts the container if CPU usage drops below 5 percent

Question 20

You are running a distributed AI training job on AKS using MPI. Which network configuration is critical for performance?

Accepted Answer

Enable Azure Accelerated Networking on the node pool.. Azure Accelerated Networking uses SR-IOV to bypass the host virtual switch and give VMs (and AKS nodes) direct hardware access to the NIC, drastically reducing latency and CPU overhead. For MPI-based distributed AI training, which is highly sensitive to inter-node latency and bandwidth (e.g., AllReduce operations), Accelerated Networking is critical for performance. It is enabled on the AKS node pool via `--enable-accelerated-networking`.

Answer

Use an internal VNET peering.

Answer

Enable standard load balancing.

Answer

Disable the pod network policy.