Night 31 — Lab 4B: Fargate right-sizing (CPU/memory, Spot vs On-Demand, Compute Optimizer)

~2 hr. $0 AWS spend — read-only audit of live Iceland task definition, cost comparison table, CloudWatch utilization sample, and Compute Optimizer API. No new ECS deploy tonight — compares ecs-task-pipeline-sizing.json scenarios against production globalskiatlas-backend-k8s-iceland.

Plain language first

How Fargate bills

Fargate charges per vCPU-second and GiB-second while the task is RUNNING — from image pull through container stop. You do not pay for EC2 instances, but you pay for every second the task exists.

Analogy (GSA Iceland): Renting a moving truck by the minute. A 2 vCPU / 4 GiB task is a bigger truck; if your boxes fit in a van (1 vCPU / 2 GiB), you overpay for empty cargo space every monthly run.

Right-sizing vs purchase optimization

Tonight focuses on the first two for Iceland; Savings Plans are the exam comparison when spend is predictable.

Fargate CPU / memory rules (exam)

• CPU in units: 256 = 0.25 vCPU, 1024 = 1 vCPU, 2048 = 2 vCPU, etc.
• Memory must be valid for the CPU tier (see Fargate task size table).
• Ephemeral storage defaults 20 GiB; extra GiB is a separate line item.

Block 1 (~40 min) — Sizing scenarios + GSA map

Read gsa-fargate-rightsizing-map.md

Pipeline path (unchanged since Night 12):

EventBridge cron ──► ecs:RunTask ──► Fargate (private) ──► S3 iceland/YYYY-MM/

Compare ecs-task-pipeline-sizing.json

Open the backend repo aws/ task JSON files when available — tonight’s study copy lives in study-lab/ecs-task-pipeline-sizing.json for offline comparison.

Exam minimums (write in notes)

• Fargate On-Demand = no upfront, highest $/second.
• Fargate Spot = discount vs Fargate OD; not the same as EC2 Spot on raw instances.
• Compute Savings Plans apply to Fargate, Lambda, EC2 — flexible spend commitment.
• Reserved Instances lock EC2 instance attributes — poor fit for pure Fargate RunTask.
• Monthly scheduled batch: total cost ≈ (per-run $) × runs/month — EventBridge is not the dominant line item.

Decision tree (memorize)

Fargate task cost too high?
│
├─ CloudWatch shows low CPU/memory p95?
│     → Register smaller task definition revision
│
├─ Job retryable, 2-min interrupt OK?
│     → Fargate Spot capacity provider on RunTask / service
│
├─ Steady hours across Fargate + Lambda + EC2?
│     → Compute Savings Plan (exam: vs RI flexibility)
│
├─ NAT Gateway dominates bill?
│     → S3 gateway + ECR endpoints (Night 9) — not smaller CPU
│
└─ Durable output on Fargate disk?
      → Wrong — S3 (Night 27/30); disk is scratch only

Block 2 (~50 min) — Lab: live task def + cost table + Compute Optimizer

Prerequisites

Spend note: This lab only calls describe/list/get APIs — $0 unless you optionally run Iceland via Night 10 script.

What the lab collects

Does not change: Task definition, EventBridge rule, cluster capacity, or Spot settings.

Run the lab

# PowerShell — from repo root
.\HTML\study-lab\night-31-lab-fargate-rightsizing-setup.ps1
.\HTML\study-lab\night-31-lab-fargate-rightsizing-setup.ps1 -PrintDecisionTree

# Git Bash / WSL
bash HTML/study-lab/night-31-lab-fargate-rightsizing-setup.sh
bash HTML/study-lab/night-31-lab-fargate-rightsizing-setup.sh --print-decision-tree

End-to-end validation

1. Result file: night-31-fargate-rightsizing-result.json — live sizing, cost rows, optimizer status.

2. ECS console: Task definitions → globalskiatlas-backend-k8s-iceland → compare CPU/memory to iceland-rightsized scenario.

3. CloudWatch: ECS → Clusters → globalskiatlas-backend-k8s → Tasks → Metrics → CPUUtilization / MemoryUtilization for last Iceland run.

4. Compute Optimizer: Console → Compute Optimizer → ECS services — if empty, note “insufficient history” (RunTask-only workloads may have limited recommendations).

5. Pricing Calculator (optional): Model 12 monthly runs at current vs rightsized vs Fargate Spot.

Troubleshoot in this order:

1. AccessDenied on ECS? Need ecs:Describe* and cloudwatch:GetMetricStatistics.
2. No stopped tasks listed? Run Night 10 once or wait for Night 12 schedule — lab still prints sizing JSON comparison.
3. CloudWatch metrics empty? Task may be older than retention — check log group for run timestamps.
4. Compute Optimizer empty? Enroll account; wait 24–48 h after 30 h of utilization.
5. Cost table differs from console bill? Script uses public list rates — Savings Plans and Spot market change effective $.

Console reading (10 min)

• ECS → Task definition → JSON → cpu, memory, ephemeralStorage.
• Billing → Cost Explorer → filter Service = Amazon Elastic Container Service → last 30 days (should be small for monthly batch).
• CloudWatch → Metrics → ECS/ContainerInsights or AWS/ECS → CPUUtilization.

Reference files

• night-31-lab-fargate-rightsizing-setup.ps1 / .sh — read-only audit
• ecs-task-pipeline-sizing.json — three sizing scenarios
• gsa-fargate-rightsizing-map.md — GSA resource map
• iceland-pipeline-compute.md — Night 27 cost context

Teardown

.\HTML\study-lab\night-31-lab-fargate-rightsizing-teardown.ps1

bash HTML/study-lab/night-31-lab-fargate-rightsizing-teardown.sh

Removes only night-31-fargate-rightsizing-result.json. No AWS resources were created.

Block 3 (~30 min) — Quiz + cost pillar notes

File: night-31-quiz.json — 15 timed scenario questions (~24 min at 96 sec each).

Score on the study site: /aws-solutions-architect-study/quiz/31/.

Write one cost win per tier:

Week 5 map (continued)

5 flashcards (write tonight)

1. Fargate billing — When does the meter start?
   (From task provisioning/RUNNING until stopped — per vCPU-second and GiB-second.)

2. Spot vs Savings Plan — Iceland monthly, retry OK?
   (Fargate Spot per run — Savings Plan when steady committed hours across compute.)

3. Right-sizing signal — CPU p95 25% for a month?
   (Candidate to reduce CPU/memory in new task definition revision.)

4. Invalid task def — 512 MiB with 2 vCPU?
   (Violates Fargate CPU/memory pairing — task fails registration.)

5. Durable output — Shrink Fargate disk to save $?
   (No — artifacts in S3; ephemeral disk is scratch only.)

Night 32 preview

Lab 4C — Step Functions capstone — state machine: ECS RunTask → wait → SQS → Lambda → SNS failure path; deploy sam-pipeline-orchestrator per gsa-integration-map.md.

Reference links

• Fargate pricing
• Fargate task CPU and memory
• Fargate Spot capacity providers
• Compute Savings Plans
• AWS Compute Optimizer