Course outline for Cloud-Native Engineering for AI Workloads

Module 1: Package a mock AI gateway and worker

• Containers: Package mock model gateway and agent-worker dependencies with separate runtime settings and disposable workspaces; distinguish packaging from a complete execution-security boundary.
• Docker: Build reproducible mock AI service images; keep provider configuration, dummy credentials, and mutable task state outside image layers.
• Kubernetes: Declare prepared gateway and worker workloads with resource limits; distinguish replica reconciliation from recovery or completion of an agent task.

Practical outcome: Build the supplied service image and deploy prepared gateway/worker manifests with resource limits; compare desired and observed replicas.

Module 2: Configure AI workload boundaries

• Service discovery: Connect workers to a mock model gateway through declared discovery; verify request routing without relying on fixed pod addresses.
• Configuration: Manage provider endpoints, timeouts, concurrency, and queue limits outside images; diagnose bad settings and verify protected input mounts reject writes.
• Secrets: Supply dummy model-provider credentials outside images and logs; distinguish credential storage from least-privilege access and complete encryption guarantees.

Practical outcome: Connect workers to a mock model gateway, mount read-only inputs, and supply dummy provider credentials and bounded queue settings outside images and logs.

Module 3: Investigate worker capacity and lifecycle

• Scaling: Compare worker throughput, throttling, and queue depth at two replica counts; choose bounded concurrency from downstream model-service capacity rather than worker availability.
• Networking: Map gateway, worker, and mock-tool connectivity; document remaining outbound authority without assuming read-only files also restrict network actions.
• Observability: Record replica, queue, throttling, cancellation, and workspace-cleanup evidence; distinguish stopping pending tasks from reversing already committed effects.

Practical outcome: Compare two worker replica counts against a rate-limited mock model service; inspect throttling, queue depth, cancellation, and supplied workspace-cleanup evidence.

Practical exercise

• Produce a service image, local gateway/worker manifests, configuration inventory, and timestamped operational evidence using the supplied scaffold.
• Verify connectivity; inject bad provider configuration and demonstrate a denied write to a read-only input mount without exposing dummy credentials.
• Remove a worker replica and observe reconciliation; inspect cancellation before and after a mock task effect, and verify the scaffold's disposable-workspace cleanup.
• Compare two replica counts under bounded load; measure throughput, queue depth, throttling, and errors and justify a concurrency limit rather than assuming more replicas help.