SAA-C03
Deep Dive
24% of the exam (D3) evaluates high-performance design. For compute, you need to pick the right service (EC2, Lambda, ECS/EKS) and the right instance type / purchasing option for each scenario.
Contents
EC2 instance types are organized into families by use case. The prefix identifies the family, the number the generation, and the suffix any additional characteristics (a=AMD,g=GPU,n=networking,i=NVMe SSD,d=HDD).
General Purpose (T, M, A1)
Balance of CPU, memory and networking. T = burstable (CPU credits). M = steady baseline.
Compute Optimized (C)
High CPU-to-memory ratio. For workloads that need intensive processing.
Memory Optimized (R, X, High Memory)
Lots of RAM. For in-memory databases, real-time analytics, big data.
Storage Optimized (I, D, H)
High local I/O throughput. NVMe SSD (I), high-density HDD (D, H).
Accelerated Computing (P, G, Inf, Trn)
GPUs (P, G) for ML training/inference, 3D rendering. Inferentia/Trainium for ML on AWS.
On-Demand
0% (baseline)Full price per hour/second
No commitment. Pay only for what you use. Maximum flexibility.
Reserved Instances (RI)
Up to 72% with 3-year All UpfrontUp to 72% discount vs On-Demand
1- or 3-year commitment. Standard RI (bigger discount), Convertible RI (can change family).
Savings Plans
EC2 Savings Plans: up to 72%. Compute: up to 66% but also applies to Lambda/Fargate.Up to 66% discount (Compute) / 72% (EC2)
$/hour commitment for 1 or 3 years. Applies automatically to EC2, Lambda, Fargate.
Spot Instances
Up to 90% — the biggest discountUp to 90% discount vs On-Demand
Uses unused AWS EC2 capacity. Can be reclaimed with 2 minutes of notice.
Dedicated Hosts
No price savings — value comes from compliance and existing licensesThe most expensive (an entire physical server)
Dedicated physical server. Lets you use existing software licenses (BYOL: Windows Server, SQL Server).
Spot Fleet — smart mixing
Spot Fleet can combine multiple instance types and AZs to maximize available capacity. If a Spot instance gets reclaimed, it automatically launches another from the available pool. For batch jobs or ML training: Spot Fleet + instance diversification.
Lambda runs code without managing servers. Scales automatically from 0 to thousands of concurrent executions. You only pay for execution time (milliseconds).
Limits that matter for the exam
Common event sources
Cold Start
First invocation or after inactivity: Lambda has to initialize the environment (~100ms-1s). Mitigate with Provisioned Concurrency(keeps N instances pre-initialized) or SnapStart for Java.
| Service | Orchestration | Infrastructure management | When to use |
|---|---|---|---|
| ECS + EC2 | ECS (AWS native) | You manage the EC2 instances | Full control over the host. Spot instances to cut costs. |
| ECS + Fargate | ECS (AWS native) | AWS manages the infrastructure (serverless) | No server management. Pay per container vCPU/memory. |
| EKS + EC2 | Kubernetes | You manage the nodes | Kubernetes portability. Teams with K8s experience. |
| EKS + Fargate | Kubernetes | AWS manages the nodes (serverless) | Kubernetes without managing nodes. More expensive than ECS Fargate. |
ECS vs EKS signals in the exam
Choose ECS when:
Choose EKS when:
| Scenario | Service | Reason |
|---|---|---|
| Processing images when uploaded to S3 | Lambda | Event-driven, short-lived, scales automatically |
| High-concurrency REST API without managing servers | Lambda + API Gateway | Serverless, pay per request, automatic scaling |
| ML training batch job that takes 2 hours | EC2 Spot (C or P family) | Lambda caps at 15 min. Spot cuts cost by up to 90% |
| Containerized microservices from the AWS team | ECS + Fargate | AWS native, no infra to manage, native integration |
| Migrating a Kubernetes app from on-prem to AWS | EKS | Same Kubernetes API, no rewriting manifests |
| Web app with predictable 24/7 load | EC2 Reserved Instances | Up to 72% savings vs On-Demand for steady load |
| Overnight batch data analysis | EC2 Spot Fleet | Interruption-tolerant, maximum savings |
Did you get this topic?
Test what you just learned
A company runs machine learning training jobs that take between 2 and 8 hours. The jobs are interruption-tolerant (they can restart from checkpoints). The company wants to minimize cost. Which EC2 configuration is most appropriate?
Put what you learned to the test with original SAA-C03 exam questions.
Start SAA-C03 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.