Compute is usually the biggest line item
Architects choose instance type, purchase option, and architecture (x86 vs Graviton). SAA-C03 expects you to match workload tolerance to pricing model.
EC2 pricing models
| Model | When | Trade-off |
|---|---|---|
| On-Demand | Unpredictable, short-term, no upfront | Highest hourly, zero commitment |
| Spot | Fault-tolerant, flexible, stateless batch | Up to ~90% discount; can interrupt with 2-min notice |
| Reserved Instances (RI) | Steady state 1–3 years | Capacity reservation in AZ (Standard) or flexibility (Convertible) |
| Savings Plans | Steady $/hour commitment (1–3 yr) | Applies across instance families/Region (Compute SP) or EC2-specific SP |
Exam trend: Savings Plans often preferred over legacy RI wording — flexible discount for consistent usage.
Spot best practices
- Use Spot Fleet or Auto Scaling with mixed instances policy.
- Checkpoint long jobs; handle Spot interruption notice via instance metadata.
- Good for: rendering, CI runners, Kafka consumers, EMR, EKS worker nodes with diversification.
- Bad for: single-instance databases, non-interruptible synchronous API only on Spot.
Right-sizing
- Downsize after Compute Optimizer or CloudWatch CPU/RAM metrics show headroom.
- T instances — cheap dev; production sustained load needs non-burstable families.
Graviton (ARM)
- m7g, c7g, r7g etc. — often 20–40% better price/performance for supported stacks.
- Recompile or use multi-arch containers; verify library compatibility.
Exam: "Reduce cost without changing architecture much" + Linux open-source stack → Graviton same size family.
Lambda and Fargate cost angles
- Lambda — pay per invoke + GB-second; optimize memory (also scales CPU), reduce cold starts only if worth provisioned concurrency cost.
- Fargate — pay per vCPU/memory per task second; right-size task definitions; consider EC2 launch type if always-on high utilization cheaper.
Licensing and hybrid
- BYOL — License Manager tracks bring-your-own Windows/SQL licenses on dedicated hosts or eligible instances.
- Dedicated Hosts — compliance/license isolation; usually more expensive than default tenancy.
Scheduling and hibernation
- Instance Scheduler (Solutions Library / custom Lambda) — stop dev instances nights/weekends.
- Hibernate — save RAM to EBS, faster boot than cold start; still pay EBS storage.
Decision flow (exam)
- Prod baseline 24/7 → Savings Plan or RI after 30–60 days stable usage (Cost Explorer recommendation).
- Batch / workers → Spot with fallback On-Demand.
- Spiky unknown → On-Demand + Auto Scaling; maybe Lambda.
- Dev/test → Spot + schedule stop.
Exam traps
- Spot without interruption handling → wrong for critical single instance.
- Convertible RI — exchange for different instance family; Standard RI — lower discount, less flexible.
- Savings Plans discount applies to usage, not arbitrary flat bill — need matching compute spend.
Official reference
SAA-C03 exam guide — cost-optimized compute.