AWS Certified Solutions Architect - Associate (SAA-C03) #9 Domain 3-1 High-Performing Architectures — Choosing Compute
With #8 we finished the resilience domain (26%). The third domain is High-Performing (24%). The first topic is compute. SAA’s compute questions ask, on cost and performance grounds, “which instance family, with which purchasing option, or on serverless, should this workload run on?”
EC2 instance families #
EC2 instances are split into families by the nature of the workload. Rather than asking you to memorize the machines in each family, the exam asks “which family is this workload?”
| Family | Representative | Suitable workloads |
|---|---|---|
| General Purpose | T, M | Web servers, small DBs, balanced general workloads |
| Compute Optimized | C | Batch processing, high-performance computation, game servers |
| Memory Optimized | R, X | In-memory DBs, large caches, real-time analytics |
| Storage Optimized | I, D | High-IOPS local storage, large datasets |
| Accelerated Computing | P, G | ML training/inference, GPU graphics |
If “in-memory database” or “large memory” is the clue, it’s memory optimized (R/X); “machine learning/GPU” means accelerated computing (P/G); and “high-performance computation/batch” means compute optimized (C).
Purchasing options #
Even for the same instance, cost varies greatly depending on the purchasing method. This choice is the heart of the cost questions.
| Option | Commitment | Cost | Suitable workloads |
|---|---|---|---|
| On-Demand | None | Baseline price | Short-term , unpredictable , testing |
| Reserved Instances | 1 yr / 3 yr | Up to ~72% savings | Steady workloads |
| Savings Plans | 1 yr / 3 yr (hourly commitment) | Reserved-level savings | Steady but needs flexibility |
| Spot | None (can be interrupted) | Up to ~90% savings | Batch , stateless that tolerate interruption |
| Dedicated Host/Instance | Varies | Expensive | Need physical isolation for licensing/compliance |
- On-Demand — pay for what you use with no commitment. For when traffic can’t be predicted or for short-term jobs.
- Reserved / Savings Plans — large discounts for a 1–3 year commitment. Suitable for a steady 24/7 base load. Savings Plans commit you to “N dollars per hour,” giving the flexibility to keep the discount even when you change instance families.
- Spot — use spare capacity up to 90% cheaper, but you get interrupted when AWS reclaims it. Suitable for batch, big data, and stateless workloads that tolerate interruption.
- Dedicated Host — occupy a physical server exclusively. Used for per-server licensing or regulatory isolation requirements.
Serverless compute #
There are also options where you don’t manage the servers yourself.
- Lambda — runs code on an event basis with no server management at all. You’re billed for execution time and it scales automatically. A single invocation lasts up to 15 minutes, so it suits short, event-driven jobs.
- Fargate — a serverless container runtime. It runs ECS/EKS containers without you having to manage EC2. It’s the answer to “I use containers but don’t want to operate instances.”
To sum up the selection criteria:
- Short and event-driven → Lambda
- Containers but avoid server management → Fargate
- Long-running , fine-grained OS control → EC2
Selection flow by workload #
- Steady 24/7 load → EC2 + Reserved/Savings Plans (base), On-Demand for the variable portion
- Interruption-tolerant batch/bulk processing → Spot
- Unpredictable , intermittent events → Lambda (if short) or On-Demand + Auto Scaling
- Containers → Fargate (avoid server management) or ECS/EKS on EC2 (fine-grained control)
A common combination is to build an ASG with Reserved for the base load and On-Demand or Spot for the variable load.
Exam question patterns #
- “Cost savings for a steady 24/7 workload.” → Reserved Instances / Savings Plans
- “Batch/bulk processing that can be interrupted.” → Spot
- “Event-driven short execution, no server management.” → Lambda
- “Containers but avoid operating instances.” → Fargate
- “Memory-intensive in-memory DB.” → memory optimized (R/X)
- “ML/GPU workload.” → accelerated computing (P/G)
- “Per-server licensing / regulatory isolation.” → Dedicated Host
Common traps #
1) Using Spot for stateful/critical workloads #
Spot can be reclaimed at any time. It’s unsuitable for critical services or stateful jobs that can’t tolerate interruption.
2) Recommending Reserved for highly variable workloads #
Reserved/Savings Plans are for steady loads. If variability is high, On-Demand, Spot, or serverless is the fit.
3) Handling long jobs with Lambda #
Lambda has a 15-minute maximum per invocation. Jobs longer than that should go to Fargate, EC2, or Batch.
4) Concluding “serverless = always cheaper” #
For continuously high load, Reserved EC2 can be cheaper. It depends on the traffic pattern.
Wrap-up #
What this post locked in:
- EC2 families — general (T/M) , compute (C) , memory (R/X) , storage (I/D) , accelerated (P/G)
- Purchasing options — On-Demand (short-term) , Reserved/Savings (steady) , Spot (interruption-tolerant , up to 90% savings) , Dedicated (isolation)
- Serverless — Lambda (event-driven , up to 15 min) , Fargate (serverless containers)
- Standard combination — build an ASG with Reserved for the base and On-Demand/Spot for the variable
Next — Domain 3-2 Caching #
Now that we’ve nailed compute, we move on to the next means of boosting performance: caching.
In #10 Domain 3-2 Caching we’ll cover the difference between ElastiCache’s Redis and Memcached, DAX for accelerating DynamoDB, CloudFront for caching content close to users, and cache strategies (lazy loading , write-through).