Compute and Application Services
This domain tests selecting a compute primitive and a purchasing model that fit a workload's traffic pattern, runtime limit, and cost profile. Most questions reduce to two decisions: instance or serverless, and commitment or on-demand.
EC2 purchasing options
- On-Demand: no commitment, billed per second with a 60-second minimum on Linux; use for spiky or short-lived work.
- Reserved Instances: 1- or 3-year term, Standard (largest discount, fixed family) or Convertible (trade discount for flexibility); zonal RIs can reserve capacity in an AZ.
- Savings Plans: Compute Savings Plans apply across EC2, Fargate, and Lambda in any Region, family, or OS; EC2 Instance Savings Plans trade scope for a deeper rate on one family in one Region.
- Spot Instances: up to 90% off, interruptible with a two-minute warning; right for batch, CI, and stateless workers.
- Dedicated Hosts give per-socket visibility for bring-your-own-license; Dedicated Instances give hardware isolation without host-level control.
- Capacity Reservations hold capacity in an AZ without a discount commitment.
Serverless and containers
AWS Lambda runs event-driven code with a 15-minute maximum execution time, per-millisecond billing, and concurrency limits. AWS Fargate runs containers without managing servers, behind ECS or EKS. Elastic Beanstalk is a PaaS layer that still provisions EC2 instances for you. AWS Batch handles queue-based batch jobs; Lightsail offers fixed monthly bundles for simple workloads.
Scaling and traffic distribution
Auto Scaling groups use launch templates to add or remove instances on target-tracking or step policies. Application Load Balancer handles HTTP/HTTPS with host- and path-based routing; Network Load Balancer handles TCP/UDP at layer 4 with static IPs and extreme throughput; Gateway Load Balancer fronts third-party appliances.
Worked example
An image-thumbnail service runs at 5,000 requests an hour during the day and almost zero at night, with each job finishing in under 30 seconds. Reasoning: the runtime fits Lambda's limit and the workload is spiky and event-driven, so S3 event notifications triggering Lambda plus DynamoDB for state beat a fleet of always-on EC2 instances. If a job needed 40 minutes of CPU, Lambda would be disqualified and a Spot-backed ECS or Batch worker fleet would be the answer.
Traps that cost marks
- Assuming Spot capacity is guaranteed, or that stopping an instance releases its EBS volume.
- Believing Savings Plans are Region-locked - only EC2 Instance Savings Plans are.
- Describing Elastic Beanstalk as serverless.
- Forgetting that an ALB requires subnets in at least two AZs.