AWS compute

Choose a compute model from workload duration, event/retry semantics, resource requirements, isolation, network access and operating responsibility. “Managed” is not the same as no security or application ownership.

Learning route

Requirement	Lesson	Operating boundary
Guest OS/device control	EC2	Patch OS/software and manage storage
Event-driven function	Lambda	Invocation/retry/runtime/dependency ownership
Native container orchestration	Fargate	Tasks, services, roles and capacity model
Kubernetes API	EKS	Cluster/workload lifecycle and access
Queued batch	Batch	Compatible job resources, retries and dependencies
Simplified selected plans	Lightsail	Product-specific constraints and VM operations

Then study fleet lifecycle, image publishing and delivery pipelines.

A short job can run on Batch; long-lived containers may run on ECS. Standard Lambda limits are not a universal duration rule for every advanced Lambda mode. Predictable demand does not automatically justify a commitment before measuring eligible usage.

Selection exercise

For an API, an overnight data job and a device-driver workload, identify the required resources/control, failure model and operations owner. Compare two candidates at equal workload/availability and include storage, network, logs and idle capacity.

Build one sandbox checkpoint through the selected lesson, test an intentional failure and perform its cleanup. Do not treat a product matrix as proof of performance.

Compute decision tree and common patterns

Need guest OS, driver or device control? ── yes → EC2
                 │ no
Need Kubernetes APIs/ecosystem? ──────── yes → EKS
                 │ no
Is work a queue/scheduled batch? ─────── yes → Batch or an ECS job
                 │ no
Is it an event-driven bounded function? yes → Lambda
                 │ no
Need a continuously serving container?  yes → ECS/Fargate

This is a starting route, not an eligibility contract. A workload can require several services: a web API may use an ALB, ECS service, SQS queue, Batch workers and S3. Decide based on latency, state, task duration, GPU/CPU/memory shape, identity and network requirements. A simple VM website may fit Lightsail when its plan and operational boundaries are enough.

Pattern	Control/data path	Failure question
Auto-scaled web tier	DNS → load balancer → EC2 Auto Scaling or ECS service	How is an unhealthy member replaced and traffic drained?
Event processing	Queue/event → Lambda or worker → durable store	How are duplicates, poison messages and retries handled?
Batch pipeline	Schedule/input → queue → Batch job → durable output	Can work checkpoint after interruption?
Container platform	Registry → revision → ECS task or EKS pod	Who patches hosts and updates the running revision?

For instance selection, compare burstable, general-purpose, compute-, memory-, storage- and accelerator-optimized families against measured use. The family label alone does not establish price or availability in a Region. For cost, count compute plus persistent disks, snapshots, egress, public addresses, load balancers and telemetry; include idle capacity and failure-domain requirements.

References

• Compute services.
• EC2 operating responsibilities.
• ECS architecture and operating choices.
• Pricing calculator.