← Design Cost-Optimized Architectures

▶ Play from here · 60 min

Starts this lesson and continues through 5 more to the end of certification prep.

Compute cost — commit, interrupt, right-size, or go serverless

Why this lesson matters

Task statement 4.2 is "Design cost-optimized compute solutions". Its most specific knowledge item names the purchasing options outright: "AWS purchasing options (for example, Spot Instances, Reserved Instances, Savings Plans)". Its skills include:

• "Determining cost-effective AWS compute services with appropriate use cases (for example, AWS Lambda, Amazon EC2, AWS Fargate)"
• "Determining the required availability for different classes of workloads (for example, production workloads, non-production workloads)"
• "Selecting the appropriate instance family for a workload" and "… instance size …"
• "Determining appropriate scaling methods and strategies for elastic workloads (for example, horizontal compared with vertical, EC2 hibernation)"

The distinguishing question is: can this workload be interrupted, and can you predict it?

   steady, predictable, runs all year          →  Savings Plans (or Reserved Instances)
   interruptible, flexible on timing           →  Spot
   short, spiky, unpredictable, can't stop     →  On-Demand
   per-socket / per-core licences, host control →  Dedicated Hosts
   event-driven, idle most of the time         →  Lambda

The purchasing options, in AWS's words

From Amazon EC2 billing and purchasing options, fetched 2026-09-25:

And AWS's own decision paragraph, verbatim — it's the exam's decision tree:

"If you can't make a commitment to a specific instance configuration, but you can commit to a usage amount, purchase Savings Plans to reduce your On-Demand Instance costs. If you require a capacity reservation, purchase Reserved Instances or Capacity Reservations for a specific Availability Zone. … Spot Instances are a cost-effective choice if you can be flexible about when your applications run and if they can be interrupted. Dedicated Hosts or Dedicated Instances can help you address compliance requirements and reduce costs by using your existing server-bound software licenses."

On-Demand — the baseline

From On-Demand Instances: "You pay only for the seconds that your On-Demand Instances are in the running state, with a 60-second minimum." AWS's recommendation: "applications with short-term, irregular workloads that cannot be interrupted."

Savings Plans — commit to dollars per hour

From What are Savings Plans? and Savings Plans types, fetched 2026-09-25:

"Savings Plans provide savings beyond On-Demand rates in exchange for a commitment of using a specified amount of compute power (measured per hour) for a one or three year period."

Four plan types, verbatim discounts:

• Payment: "All upfront, Partial upfront, or No upfront".
• ⚠️ Commitment is fixed: "The terms of the commitment can't be changed after purchase."
• ⚠️ "Dedicated Instances are charged $2/hour in every Region you have at least one Dedicated Instance running. These dedicated fees are not discounted by Savings Plans."
• ⚠️ EKS: "Amazon EKS charges will not be covered by Savings Plans, but the underlying EC2 instances will be."

The trade-off in one line: EC2 Instance Savings Plans give the bigger discount (72%) for less flexibility; Compute Savings Plans give 66% but follow you across families, Regions, and onto Fargate and Lambda. "We're migrating from EC2 to containers on Fargate next year" → Compute Savings Plan.

Reserved Instances — commit to a configuration

From Reserved Instances for Amazon EC2 overview, fetched 2026-09-25. AWS's own recommendation comes first on the page:

"We recommend Savings Plans over Reserved Instances. Savings Plans are the easiest and most flexible way to save money on your AWS compute costs and offer lower prices (up to 72% off On-Demand pricing), just like Reserved Instances."

"Reserved Instances are not physical instances, but rather a billing discount applied to the use of On-Demand Instances in your account."

⚠️ "After you purchase a Reserved Instance, you cannot cancel your purchase." You might be able to modify, exchange, or sell it on the Reserved Instance Marketplace.

⚠️ "Reserved Instances do not renew automatically; when they expire, you can continue using the EC2 instance without interruption, but you are charged On-Demand rates."

Regional vs zonal — the capacity question

From Regional and zonal Reserved Instances:

⚠️ This is the only reason to pick an RI over a Savings Plan on the exam: a stem that needs a capacity guarantee in a specific AZ plus a discount. Savings Plans are pricing only. (A Capacity Reservation also reserves capacity, but on its own is not a discount.)

Spot — the interruption deal

From Spot Instances and the Spot product page, fetched 2026-09-25:

"A Spot Instance is an instance that uses spare EC2 capacity that is available for less than the On-Demand price."

• Discount: "up to a 90% discount compared to On-Demand prices" (product page).
• Price: "set by Amazon EC2, and is adjusted gradually based on the long-term supply of and demand for Spot Instances."
• Fit: "Spot Instances are a cost-effective choice if you can be flexible about when your applications run and if your applications can be interrupted. For example, … data analysis, batch jobs, background processing, and optional tasks." The product page adds "stateless, fault-tolerant, or flexible applications such as big data, containerized workloads, CI/CD, web servers, high-performance computing (HPC), and test & development workloads".
• ⚠️ No stacking: "Spot Instances are not covered by Savings Plans. … your spend on Spot Instances does not apply the commitments in your Compute Savings Plans."

The interruption notice — verified

From Spot Instance interruption notices, fetched 2026-09-25:

"A Spot Instance interruption notice is a warning that is issued two minutes before Amazon EC2 stops or terminates your Spot Instance. If you specify hibernation as the interruption behavior, you receive an interruption notice, but you do not receive a two-minute warning because the hibernation process begins immediately."

• Delivered "as an EventBridge event and as items in the instance metadata" (spot/instance-action).
• AWS recommends checking "every 5 seconds".
• ⚠️ "Interruption notices are emitted on a best effort basis." Design for interruption, don't depend on the warning.
• Interruption behaviours: terminate, stop, or hibernate.

And the earlier signal: the rebalance recommendation "notif[ies] you that a Spot Instance is at an elevated risk of interruption", "without having to wait for the two-minute Spot Instance interruption notice." That's what Auto Scaling's Capacity Rebalancing (SAA2 lesson 2) acts on.

⚠️ Spot is wrong for: a single stateful database, anything that "must not be interrupted", or a steady 24/7 baseline (that's a Savings Plan). The common right answer mixes them: Savings Plan for the baseline, Spot for the burst, On-Demand for what can't be interrupted.

Dedicated Hosts vs Dedicated Instances

From Amazon EC2 Dedicated Hosts, fetched 2026-09-25:

"There are no performance, security, or physical differences between Dedicated Instances and instances on Dedicated Hosts."

The difference is visibility and licensing:

AWS's rule: "If you require your instances to run on dedicated hardware, but you do not need visibility or control over instance placement, and you do not need to use per-socket or per-core software licenses, you can consider using Dedicated Instances instead."

Exam signal: "existing per-core SQL Server / Windows Server licences" → Dedicated Hosts. "Compliance says single-tenant hardware, no licensing angle" → Dedicated Instances. Dedicated Host Reservations give "a discount of up to 70 percent compared to On-Demand Dedicated Host pricing".

Right-sizing — family, size, and generation

From the instance types guide, fetched 2026-09-25, the current-generation categories are General purpose, Compute optimized, Memory optimized, Storage optimized, Accelerated computing, and High-performance computing. The EC2 User Guide's instance types page adds: "Names are based on instance family, generation, processor family, capabilities, and size."

⚠️ The per-category descriptions and example workloads live on sub-pages I did not fetch (.../instancetypes/gp.html, co.html, mo.html, so.html, ac.html). SAA3 lesson 2 covers family selection for performance; here, the cost rule is simply match the family to the bottleneck, then pick the smallest size that meets it.

Three cost-relevant facts that are on the pages I fetched:

• Burstable (T) instances "provide a baseline level of CPU performance with the ability to burst above the baseline", for "web servers, small and medium databases, … development and test environments."
• Flex instances (C7i-flex, M7i-flex and others) "provide the most cost-effective way to run a broad spectrum of general purpose applications", with "a baseline CPU performance of 40 percent".
• Graviton is "a family of processors designed to deliver the best price performance".
• "Because you pay by the instance second, it's convenient and inexpensive to test multiple instance types before making a decision."

Compute Optimizer (lesson 1) is the right-sizing tool: 14 days of metrics by default, 93 with the paid enhanced option.

Virtualization (the knowledge item mentions it): "Current generation instance types support hardware virtual machine (HVM) only", and AWS recommends HVM AMIs "For best performance".

Scaling for cost — and hibernation

Scaling mechanics are SAA2 lesson 2. The cost angle: horizontal scaling lets capacity follow demand, so you stop paying for the peak all day.

Hibernation is named in the guide twice. From Hibernate your Amazon EC2 instance, fetched 2026-09-25:

"Hibernation saves the contents from the instance memory (RAM) to your Amazon Elastic Block Store (Amazon EBS) root volume."

"If an instance or application takes a long time to bootstrap and build a memory footprint to become fully productive, you can use hibernation to pre-warm the instance."

"You're not charged for instance usage for a hibernated instance when it is in the stopped state … You are charged for storage of any EBS volumes, including storage for the RAM contents."

Exam signal: "instances take 20 minutes to warm caches; we want to stop them overnight without losing the warm state" → hibernate.

⚠️ I did not fetch the hibernation prerequisites page, so this lesson makes no claim about the root volume encryption requirement, maximum RAM, or maximum hibernation duration. Read docs.aws.amazon.com/AWSEC2/latest/UserGuide/hibernating-prerequisites.html.

Production vs non-production

The skill "Determining the required availability for different classes of workloads" is a cost skill: you don't buy production resilience for a dev environment. Levers that fit non-production, all from pages cited above:

• Stop it when idle — On-Demand bills only "in the running state"; hibernation keeps warm state.
• Spot for "test & development workloads".
• Burstable T instances for "development and test environments".
• Single-AZ and One Zone storage classes where the data is reproducible (lesson 2).
• Aurora Serverless that scales down when idle (lesson 4).

Containers and serverless — pay for what runs

The knowledge item is "Optimization of compute utilization (for example, containers, serverless computing, microservices)".

Lambda, from Lambda pricing, fetched 2026-09-25:

"Lambda Functions are priced based on the number of requests served and the duration your code runs, measured in GB-seconds."

Duration is "rounded up to the nearest 1ms". Free tier: "one million requests and 400,000 GB-seconds per month". ⚠️ The live price table on that page is filled in client-side and didn't come through in the raw page text; the page's worked examples use "$0.20 per million requests" and "$0.0000166667" per GB-second. Treat those as example figures and check the Region selector before quoting a rate.

Fargate, from Fargate pricing, fetched 2026-09-25: you pay "for the amount of vCPU, memory, and storage resources consumed", billed "per second with a 1-minute minimum" for Linux (5 minutes for Windows). Fargate Spot: "up to a 70% discount off the regular Fargate price" for interrupt-tolerant ECS tasks. SAA2 lesson 4 verified Fargate Spot's two-minute warning.

Compute Savings Plans cover both — the Savings Plans types page says they "also apply to your Fargate and Lambda usage". ⚠️ The Lambda pricing page, as retrieved on 2026-09-25, didn't restate this for standard Lambda functions (it mentioned Savings Plans only for Lambda Managed Instances). I'm relying on the Savings Plans guide; check docs.aws.amazon.com/savingsplans/latest/userguide/sp-services.html if a question turns on it.

⚠️ Lambda isn't automatically cheaper. A function busy around the clock pays for every GB-second. The cost case for Lambda is idle time you no longer pay for.

Hybrid and edge — Outposts

The knowledge item: "Hybrid compute options (for example, AWS Outposts)". From What is AWS Outposts?:

"AWS Outposts is a fully managed service that extends AWS infrastructure, services, APIs, and tools to customer premises … for lower latency and local data processing needs."

Form factors: Outposts racks ("an industry-standard 42U rack") and Outposts servers (1U or 2U, for "sites that have limited space or smaller capacity requirements"). Pricing: "When you place an order, you can choose from a variety of Outpost configurations … You also choose a contract term and a payment option." You are also "billed for … any data transfer from the AWS Region to the Outpost".

Exam signal: Outposts is chosen for latency, local processing, or data residency — not because it is cheaper. It's a term commitment on hardware you host.

⚠️ The guide also lists "Distributed compute strategies (for example, edge processing)". I did not fetch a page that frames edge compute in cost terms (CloudFront Functions, Lambda@Edge, Local Zones), so this lesson makes no claim about it. Read the CloudFront developer guide's edge-functions chapter.

⚠️ "AWS global infrastructure (for example, Availability Zones, AWS Regions)" matters for cost because prices differ by Region — every pricing page fetched for this module has a Region selector. I did not fetch a page that states which Regions are cheapest, and it would change anyway.

⚠️ The skill "Determining an appropriate load balancing strategy (ALB [Layer 7] … NLB [Layer 4] … Gateway Load Balancer)" is covered on function in SAA2 lesson 3. I did not fetch Elastic Load Balancing pricing, so this module makes no per-type cost claim. Read aws.amazon.com/elasticloadbalancing/pricing/.

Choosing, under exam conditions

Check yourself

1. How much warning does a Spot Instance get, and when do you get none?
2. EC2 Instance Savings Plan vs Compute Savings Plan — discount and flexibility?
3. When does the exam want a Reserved Instance over a Savings Plan?
4. Dedicated Host or Dedicated Instance for bring-your-own per-core licences?
5. Does a Compute Savings Plan discount Spot usage?

Answers

1. Two minutes — "issued two minutes before Amazon EC2 stops or terminates your Spot Instance." With hibernation as the interruption behaviour there's no two-minute warning, "because the hibernation process begins immediately." And notices are "best effort".
2. EC2 Instance: up to 72%, locked to "a specific instance family in a chosen AWS Region". Compute: up to 66%, any family, size, Region, OS, tenancy, plus Fargate and Lambda.
3. When the stem needs reserved capacity in a specific AZ — a zonal RI "reserves capacity in the specified Availability Zone"; Savings Plans don't reserve capacity.
4. Dedicated Host — "per-socket, per-core, or per-VM software licenses" and core/socket visibility. Dedicated Instances have only "Partial support" for BYOL.
5. No. "Spot Instances are not covered by Savings Plans."

Teaching this section

• Draw the four-way box first (steady / interruptible / spiky / licensed). Every purchasing question sorts into one cell.
• Demo: aws ec2 describe-spot-price-history for one instance type across AZs. Then show the spot/instance-action metadata path. The two-minute notice becomes concrete.
• Best discussion prompt: "We bought three-year Standard RIs for m5. Next year we move to Graviton and Fargate. What did we do wrong?" Walk them to Convertible RIs and Compute Savings Plans.
• Watch for: "Spot + Savings Plan stacks". It doesn't.
• Timing: 30 min presented, 20 on the choosing table.