Developing AI Cloud Solutions on Azure

Learn the artifact lifecycle before orchestration. One tested image should move through environments without hidden rebuilds or embedded environment secrets.

• Build, tag, version, retain, and inspect images in Azure Container Registry
• Distinguish repositories, manifests, layers, tags, and immutable digests
• Run ACR Tasks manually and from source, schedule, or base-image triggers
• Use managed identity and AcrPull instead of registry passwords where supported
• Configure App Service custom containers, listening ports, app settings, and logs
• Inject environment-specific settings at runtime rather than rebuilding images
• Protect secrets with identity and Key Vault references or retrieval
• Practice startup, health, image pull, registry, port, and configuration diagnosis

Deploy immutable application versions and connect scaling to real demand while protecting downstream quotas and database connections.

• Create Container Apps environments, apps, ingress, identities, probes, secrets, and storage
• Understand revision-scope changes and immutable revision snapshots
• Use single- and multiple-revision modes intentionally
• Canary, split traffic, label revisions, deactivate candidates, and roll back
• Configure HTTP, CPU, memory, and custom KEDA scale rules
• Scale Service Bus workers to zero with managed-identity scaler authentication
• Bound minimum, maximum, concurrency, cooldown, and polling behavior
• Deploy AKS workloads with manifests for Deployment, Service, ConfigMap, Secret, and probes
• Troubleshoot AKS and Container Apps logs, events, revisions, replicas, DNS, and end-to-end connectivity

Spend the largest block on the 25–30% data domain. Measure each store according to its own consistency, query, indexing, vector, connection, and cache model.

• Connect to Cosmos DB for NoSQL with SDK clients and identity where supported
• Use point reads and partition-scoped parameterized queries
• Interpret RU charge, query metrics, indexing policy, partition distribution, and consistency
• Store embeddings with a vector policy and query VECTORDISTANCE with TOP N
• Combine vector retrieval with trusted partition and metadata filters
• Implement a change feed processor with stable name, lease container, and idempotency
• Model PostgreSQL tables with relational metadata and pgvector columns
• Compare exact search, HNSW, IVFFlat, and supported alternatives from measured recall and latency
• Optimize PostgreSQL compute, memory, storage, filter indexes, query plans, pools, and PgBouncer
• Use Azure Managed Redis for caching, expiration, invalidation, semantic cache, and vector search patterns

Build asynchronous paths that acknowledge quickly, survive restarts, contain poison work, and preserve trace and business identity across boundaries.

• Build and deploy HTTP, Service Bus, Event Grid, and Cosmos DB-triggered Functions
• Know what triggers, input bindings, output bindings, and SDK clients each provide
• Choose Service Bus queues for point-to-point work and topics for durable fan-out
• Use topic subscription filters and sessions only when requirements justify them
• Design peek-lock settlement, lock renewal, bounded retry, timeouts, and idempotency
• Inspect DeadLetterReason and DeadLetterErrorDescription and implement controlled replay
• Use Event Grid event type, subject, and advanced filters to reduce irrelevant deliveries
• Configure Event Grid retries and dead-letter destinations and expect at-least-once delivery
• Align KEDA backlog scaling with Service Bus lock, concurrency, AI quota, and data-tier capacity

Finish by proving that the system can rotate credentials, refresh configuration, expose useful telemetry without leaking data, and recover from realistic failures.

• Prefer managed identity and DefaultAzureCredential over stored Azure service keys
• Use least-privilege Azure RBAC and separate runtime, deployment, and replay duties
• Store unavoidable secrets in Key Vault and test rotation and refresh behavior
• Use App Configuration labels, prefixes, feature flags, caching, and controlled refresh for non-secrets
• Instrument traces, metrics, and logs with Azure Monitor OpenTelemetry
• Propagate W3C trace context over HTTP and messaging and avoid sensitive/high-cardinality attributes
• Use Application Insights transaction diagnostics and application map
• Write KQL for p95 latency, failures, dependency duration, queue age, retries, DLQ reasons, revisions, and RUs
• Run duplicate, poison, quota, cache-failure, connection-pressure, rotation, network, and rollback drills
• Complete timed original questions and explain every distractor from official documentation