AI Cloud Developer Notes: ship the call that fits

Domain 1: Containerized solutions

Start with the image, then pick the host. Azure Container Registry stores the artifact. App Service, Container Apps, and AKS run it. Mixing those jobs is the usual distractor pattern on this exam.

Azure Container Registry

ACR holds container images and tags. Build and push from a Dockerfile, then version with tags that name the release. ACR Tasks can rebuild when a base image updates or on a schedule, so you do not wait for a manual push after a security patch in the base. Managing repositories, tags, and retention keeps stale images from becoming the default pull.

The registry is not a compute service. A successful push does not mean anything is running. The next stem usually asks which host pulls that tag.

App Service containers

App Service can run a container from ACR. Set environment variables for non-secret configuration. Inject secrets from Key Vault or App Service secret settings rather than baking them into the image. App Service fits a web or API container when you want platform scaling and slots without writing Kubernetes manifests.

Azure Container Apps

Container Apps sits between App Service simplicity and full Kubernetes. An environment groups apps that share networking and observability. A revision is an immutable deploy of the app; traffic can split across revisions while you validate a change. KEDA-based scaling reacts to HTTP load, queue depth, or other event sources so replica count can drop toward zero when idle.

Choose Container Apps when revisions and event-driven scale matter more than managing nodes. Choose AKS when the stem requires Kubernetes manifests, custom controllers, or cluster troubleshooting the guide names.

Azure Kubernetes Service

AKS runs your workloads from manifests: Deployments, Services, ConfigMaps, and Secrets. You own the YAML shape, readiness probes, and how the Service exposes the pod. When something fails, read pod logs, describe events, and check connectivity between pods or to a private dependency. A CrashLoopBackOff is an application or probe problem until the events say otherwise. A timeout to a database is often a network path or DNS name, not a larger node pool by default.

Extra Designing a full multi-region cluster topology is broader than the developer focus here. Using manifests and reading logs for a failing pod is in scope.

Azure Database for PostgreSQL

Connect with the SDK or a standard driver. Design schemas and indexes for the queries you run, not for a generic dump of every column. The pgvector extension stores embeddings and supports similarity search. RAG patterns usually retrieve top-k neighbors, then apply metadata filters so the wrong tenant or document type does not enter the prompt context.

Connection pooling and reuse matter when Functions or containers open many short connections. For vector workloads, size compute, memory, and storage for the index and working set. Under-provisioned memory shows up as slow similarity queries long before row count alone explains the lag.

Azure Managed Redis

Redis is the hot path: cache a repeated read, set an expiration, and invalidate when the source changes. Stale embeddings or feature flags in cache are a correctness bug, not only a performance miss. Azure Managed Redis also supports vector indexing when the latency budget needs an in-memory similarity path.

Paired choice: Redis for ephemeral speed versus Cosmos or PostgreSQL as the durable store. Do not treat Redis as the only system of record for documents you cannot rebuild.

Azure Event Grid

Event Grid delivers events to handlers with subject and type filters so subscribers skip noise. Custom events let your app publish into the same model. When delivery fails, Event Grid retries according to its policy; plan a dead-letter destination when the handler stays down. Filters belong on the subscription, not only in application code after every event arrives.

Azure Functions

A trigger starts the function: HTTP, queue, Event Grid, timer, and others. Bindings inject inputs and outputs so you write less glue code around Service Bus, storage, or Cosmos. Use Functions for serverless APIs and event handlers that scale with load. Configure app settings and Key Vault references outside the deployment package. Deploy the function app through your usual pipeline; a console-only change is hard to reproduce.

Extra Designing every enterprise API solely through API Management is not the center of the AI-200 midpoint map. Knowing Functions triggers and bindings is.

Azure App Configuration

App Configuration holds application settings and feature flags that are not secrets. Dynamic refresh lets you change behavior without a new container revision when the app is written to poll or listen for updates. Paired choice: Key Vault for passwords and keys; App Configuration for toggles and non-secret endpoints. Putting a database password in App Configuration fails the distinction the exam tests.

Extra Deep Entra ID token design and custom authentication middleware beyond “use identity to reach Key Vault or App Configuration” sit outside the skill list emphasis here.

OpenTelemetry and KQL

OpenTelemetry SDKs instrument code so traces follow one request across containers, Functions, and downstream calls. Spans name the slow hop; baggage and attributes help you filter later. Export to Azure Monitor or another backend your environment uses.

KQL queries logs and metrics after the fact. Start from the time window and the resource, then filter on severity, operation id, or custom dimensions that match the OpenTelemetry attributes you emitted. A trace answers “where did this request go.” A KQL log query answers “how often did this error appear” or “what did the process print.”

Extra Classic Blob Storage lifecycle-only scenarios from older developer exams are useful history, but AI-200’s data domain centers Cosmos, PostgreSQL with pgvector, and Managed Redis.

Frequently asked questions