4. Compute Selection

EC2

Full control, custom configs

ECS/Fargate

Containers, minimal ops

EKS

Kubernetes compatibility

Lambda

Event-driven, <15 min

Batch

Large-scale batch jobs

On-Demand

Short-term, unpredictable

Reserved 1y/3y

Up to 72% off

Savings Plans 1y/3y

Up to 72%, flexible

Spot

Up to 90% off, fault-tolerant (2-min warning)

Dedicated Host

BYOL, compliance

Standard RI

locked, sellable on Marketplace

Convertible RI

change family/OS, NOT sellable

Compute SP

most flexible (EC2 + Fargate + Lambda)

EC2 Instance SP

1 family + 1 region, 72% max

Savings Plans: SageMaker SP

4. Compute Selection

Target Tracking

maintain metric (recommended)

Step Scaling

granular response

Simple Scaling

legacy

Scheduled

time-based

Predictive

ML-based 48h forecast

Spot Allocation: capacity-optimized (recommended, lowest interruption)

4. Compute Selection

Spot Allocation: price-capacity-optimized (balance)

4. Compute Selection

Spot Allocation: lowest-price (cheap, more interruption)

4. Compute Selection

Spot Allocation: diversified

4. Compute Selection

Cluster

low latency, 1 AZ

Spread

max 7/AZ, isolation

Partition

distributed systems (HDFS)