Domain 3: Cost-Optimized Architecture Design (15%)¶

Overview¶

Cost is now its own domain at 15 percent. In C02 cost was spread across "Design for New Solutions" and "Continuous Improvement for Existing Solutions" (including a cost control task). Four tasks: 3.1 architectures based on AWS pricing models, 3.2 data transfer and network costs, 3.3 cost allocation and visibility, 3.4 storage and database costs. The recurring exam skill is choosing the cheapest option that still meets the stated performance, availability and durability requirements.

📖 Exam guide: Content Domain 3 - Official tasks and skills

Task 3.1: Pricing models and cost-optimized architectures¶

Purchase options (3.1.1). Match the commitment to the predictability of the usage:

Architecture choices that cut cost (3.1.2). Graviton (AWS Arm-based processors) typically provides better price-performance when the software supports Arm. Pick the Step Functions workflow type by pattern: Express for high-volume short workflows is billed by requests and duration and is far cheaper per execution than Standard, which is billed per state transition and suits long, auditable workflows. Serverless (Lambda) removes idle cost for spiky workloads but can cost more than reserved instances for sustained high-throughput work.

Rightsizing and scaling (3.1.3). AWS Compute Optimizer recommends instance types, EBS volumes and Lambda memory from measured utilization. Scale to demand with scheduled scaling (known patterns), target tracking (hold a metric at a target) and Fargate Spot for interruptible containers.

AI/ML cost (3.1.4, new). Select the right service tier for the workload, batch requests that do not need real-time responses, and cache repeated prompts or responses. In Bedrock terms this means matching on-demand versus provisioned throughput and batch inference to the traffic shape; in SageMaker AI it means choosing among real-time, serverless, asynchronous and batch inference.

Task 3.2: Data transfer and network costs¶

Data transfer is a classic hidden cost. The principles: data entering AWS from the internet is free; data leaving AWS to the internet, moving between Regions, and in many cases moving between Availability Zones is charged.

• Content delivery (3.2.1): CloudFront caches at the edge, reducing origin load and egress; Origin Shield adds a central caching layer to reduce origin requests; origin failover supports availability.
• Private connectivity and Regional traffic (3.2.2): gateway VPC endpoints for S3 and DynamoDB are free and avoid NAT gateway data processing charges; interface endpoints (PrivateLink) avoid internet egress for AWS services. A NAT gateway charges per GB processed, so heavy S3 traffic from private subnets through a NAT gateway is a typical waste. Keep chatty components in the same Availability Zone where availability allows (AZ-aware architectures).
• Replication (3.2.3): S3 Cross-Region Replication and DynamoDB global tables multiply storage and transfer cost; replicate only what recovery or latency requires.
• Network architecture pricing (3.2.4): Transit Gateway charges per attachment hour and per GB processed; PrivateLink charges per endpoint hour and per GB. At low volume and few VPCs, peering can be cheaper; at scale, Transit Gateway's operational benefits win.
• AI/ML transfer (3.2.5, new): cross-Region inference routing lets a Bedrock request be served from another Region for capacity; consider its transfer and compliance implications.

Task 3.3: Cost allocation and visibility¶

• Tagging strategy (3.3.1): tag by business unit, project, environment and owner; activate cost allocation tags in Billing so they appear in cost data.
• Reporting (3.3.2): Cost Explorer for interactive analysis, Cost and Usage Reports for detailed line-item data (delivered to S3 and queried with Athena).
• Enforce tagging (3.3.3): tag policies in Organizations, SCPs that deny creating resources without required tags, and AWS Config rules that flag non-compliant resources.
• Chargeback and showback (3.3.4): showback reports cost to teams for awareness; chargeback bills them. Cost Categories group costs by rules into business-meaningful buckets; split cost allocation data attributes shared-cluster costs (for example EKS) to workloads; for generative AI, monitor Bedrock token usage per application or team.
• Monitoring and alerting (3.3.5): AWS Budgets alerts on actual or forecast thresholds; Cost Anomaly Detection uses machine learning to flag unusual spend.

Task 3.4: Storage and database costs¶

• Storage tiering (3.4.1): S3 Intelligent-Tiering moves objects between access tiers automatically (good for unknown or changing access patterns); S3 Lifecycle policies transition and expire on a schedule; S3 Glacier storage classes for archives (retrieval time and minimum duration are the trade-offs); EFS lifecycle moves infrequently accessed files to cheaper classes.
• Block storage (3.4.2): gp3 volumes decouple IOPS and throughput from size and are usually cheaper than gp2 for the same performance; use Provisioned IOPS (io1/io2) only when the workload needs sustained IOPS beyond what gp3 provides.
• Databases (3.4.3): DynamoDB on-demand capacity for unpredictable traffic versus provisioned with auto scaling (and reserved capacity) for predictable traffic; RDS reserved DB instances; Aurora Serverless for variable load; Aurora I/O-Optimized for I/O-heavy workloads where I/O charges dominate; Database Savings Plans.
• Backup and DR retention (3.4.4): AWS Backup plans with retention matched to requirements, and Amazon Data Lifecycle Manager for snapshot schedules and expiry. Keeping backups forever is a cost; keeping too few is a risk.
• Observability cost (3.4.5, new): control log volume with filtering, aggregate metrics, set retention periods and archive older logs to cheaper storage.

aws s3api put-bucket-lifecycle-configuration --bucket example-logs --lifecycle-configuration '{
  "Rules": [{
    "ID": "tier-then-expire",
    "Status": "Enabled",
    "Filter": {"Prefix": "logs/"},
    "Transitions": [{"Days": 30, "StorageClass": "GLACIER"}],
    "Expiration": {"Days": 365}
  }]
}'

📖 AWS Cost Management documentation - Cost Explorer, Budgets, Cost Categories and anomaly detection 📖 Amazon S3 Intelligent-Tiering - Access tiers and when to use them

What changed from C02 in this domain¶

• Structure: C02 had no standalone cost domain; cost lived in "Design for New Solutions" (2.5, 2.6) and "Continuous Improvement" (3.5 cost control). C03 gathers it into four tasks.
• New: AI/ML cost optimization (3.1.4), AI/ML data transfer (3.2.5) and cost-aware observability (3.4.5); Bedrock token-usage monitoring for chargeback.
• Services added to scope: Reserved Instances (named), AWS Cost Categories, AWS Data Transfer Terminal, Amazon Data Lifecycle Manager.

Key terms¶

• Savings Plan - a pricing model that gives lower rates in exchange for a committed dollars-per-hour spend for one or three years.
• Spot Instance - spare EC2 capacity offered at a discount that AWS can reclaim with short notice.
• AWS Graviton - AWS-designed Arm-based processors that often provide better price-performance.
• Compute Optimizer - a service that recommends resource configurations based on measured utilization.
• Gateway VPC endpoint - a no-charge endpoint type for S3 and DynamoDB that keeps traffic on the AWS network.
• NAT gateway data processing charge - a per-GB fee for traffic that passes through a NAT gateway.
• CloudFront Origin Shield - an additional caching layer that reduces the number of requests reaching the origin.
• Cost allocation tag - a tag activated in Billing so its values appear in cost reports.
• Showback - reporting costs to teams for visibility without billing them.
• Chargeback - billing teams or business units for the costs they incur.
• Cost Categories - rules that group AWS costs into custom business-meaningful categories.
• AWS Budgets - a service that alerts when actual or forecast spend crosses thresholds.
• Cost Anomaly Detection - a service that uses machine learning to identify unusual spending.
• S3 Intelligent-Tiering - an S3 storage class that moves objects between access tiers automatically based on usage.
• gp3 - an EBS general purpose SSD volume type that lets IOPS and throughput be set independently of size.
• Aurora I/O-Optimized - an Aurora configuration with no separate I/O charges, suited to I/O-intensive workloads.