What's changed: Added per-section figures (cert-figure-retrofit). New AI-200 Chapter 2 (containers = ACR/Container Apps/AKS selection, scaling & serverless = KEDA event-driven autoscaling/Azure Functions triggers & bindings/containers vs functions)
2.1Running AI Workloads in Containers
Understand running/distributing AI back-ends in containers—container basics, image management in Azure Container Registry (ACR), and choosing between Azure Container Apps and Azure Kubernetes Service (AKS)—from a developer’s view.
AI back-ends (model-calling APIs, RAG pipelines, agent tool execution) are commonly run and distributed as containers. A container bundles the app and its dependencies into one image for reproducible, run-anywhere behavior. On Azure, store images in Azure Container Registry (ACR) and distribute them to container runtimes.
2.1.1Runtime options
- Azure Container Registry (ACR): a private registry for container images—the hub for build/store/distribute.
- Azure Container Apps: serverless-oriented container hosting with easy scaling (incl. scale-to-zero) and event-driven triggers—good for microservices/AI APIs.
- Azure Kubernetes Service (AKS): managed Kubernetes—for advanced control, scale, and special needs (e.g., GPUs), with more operational overhead.
2.1.2Choosing between them
The rule of thumb: "start with Container Apps; move to AKS when needed." Container Apps hides Kubernetes complexity and makes scaling and event-driven triggers easy—enough for many AI APIs/workers. Choose AKS for fine-grained Kubernetes control, custom networking/scheduling, or special hardware. Both consume images from ACR.
Common: (1) "private registry for container images" = ACR. (2) "serverless-oriented container hosting with easy scaling/event triggers" = Azure Container Apps. (3) "managed Kubernetes for advanced control/special needs" = AKS. (4) the rule: "start with Container Apps, move to AKS when needed."
Watch out: (1) ACR (image store) and the runtime (Container Apps/AKS) are separate layers. (2) Choose Container Apps (simple, serverless-oriented) vs AKS (powerful, more ops) by requirements—AKS is not always superior. (3) Containerization gives reproducibility; auth/networking (Chapter 4) still need separate design.
When you need to train and manage models, use Azure Machine Learning. It provides training jobs, pipelines, a model registry, and managed endpoints, serving as the foundation for AI solutions that run containerized custom models on AKS or Container Apps.
2.1.3Section summary
- Run/distribute AI back-ends as containers for reproducibility; store images in ACR
- Container Apps = simple, serverless-oriented / AKS = managed Kubernetes for advanced needs
- Rule: start with Container Apps, move to AKS when needed
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which Azure private registry stores and distributes container images?
Q2. Which serverless-oriented container runtime makes scaling (incl. scale-to-zero) and event triggers easy?
Q3. Which runtime suits advanced Kubernetes control or special hardware needs (e.g., GPUs)?
Q4. Which is the most appropriate policy for choosing a container runtime?
Q5. Which correctly relates ACR and the runtime (Container Apps/AKS)?