Pricing Models: On-Demand, RIs (Reserved Instance), Savings Plans and Spot#
Almost every cost question is a matching game: workload pattern → purchase option. Get these five into your head and half of Domain 4 is free.
The options#
| Option | Discount | Commitment | Best for |
|---|---|---|---|
| On-Demand | 0% | None | Spiky, unknown, short-lived workloads |
| Savings Plans (Compute / EC2 (Elastic Compute Cloud) Instance / SageMaker) | up to 72% | 1 or 3 years, $/hour commitment | Steady usage, want flexibility |
| Reserved Instances (Standard / Convertible) | up to 72% | 1 or 3 years, specific instance family | Very steady, known capacity |
| Spot Instances | up to 90% | None - can be reclaimed (2-min notice) | Fault-tolerant, interruptible work |
| Dedicated Hosts / Instances | 0% (premium) | None/1-3y | Licensing, compliance, single-tenant |
Savings Plans vs Reserved Instances (a favorite comparison)#
- Savings Plans = commit to $X/hour of compute; covers any instance family, region (Compute SP (Savings Plan)), Fargate and Lambda too. More flexible → usually the better answer
- Reserved Instances = commit to a specific instance type/region; can be standard (most discount, resellable) or convertible (less discount, can change family/OS); zonal RI also reserves capacity in an AZ (Availability Zone)
- Scheduled RIs = recurring time windows (rarely tested)
Not: “surekli calisan ve tahmin edilebilir is yuk” = Savings Plans (esnek oldugu icin RI’ye tercih edilir). “Kesintiye dayanikli toplu is / ML egitim / test ortami” = Spot. “Windows/SQL lisansi bizim” = Dedicated Host.
Spot - details the exam tests#
- Spare capacity at up to 90% off; AWS can reclaim with a 2-minute interruption notice
- Combine with ASG (Auto Scaling Group) / Batch / ECS (Elastic Container Service) / EKS (Elastic Kubernetes Service) / Fargate Spot and graceful shutdown handling
- Hibernate / stop-interruptible options for checkpoint-friendly workloads
- Diversify instance types and AZs to reduce interruption probability (see my notes: spot-instances)
Never put Spot on: databases, long transactions, steady user-facing SLAs without fallback.
Free tier & misc#
- 12-month free tier (750 hrs t2/t3.micro), always-free products (Lambda 1M requests), 12-month trial offers
- Per-second billing for most compute (Linux) and per-GB data transfer pricing
flowchart LR
q{What does the workload look like?}
q -->|unknown/spiky| od[On-Demand]
q -->|steady 24/7| sp[Savings Plans / RI]
q -->|interruptible batch| spot[Spot]
q -->|license-bound| dh[Dedicated Host]Exam check#
Q1: A CI/CD build farm runs thousands of short-lived jobs that can tolerate failures and must minimize cost. What should be used?
- A) On-Demand instances
- B) Spot Instances
- C) Dedicated Hosts
- D) Standard RIs
Cevabı göster
Answer: B. Fault-tolerant, interruptible, cheapest = Spot.
Q2: A company runs 40 m5.large instances 24/7 for the next 3 years and wants the largest discount but may need to switch to c5 later. Which option?
- A) Compute Savings Plans (3-year)
- B) On-Demand
- C) Scheduled RIs
- D) Spot Instances
Cevabı göster
Answer: A. Steady usage + long horizon + flexibility to change families = Compute Savings Plans (a Convertible RI would also discount but Compute SP is the flexible, family-agnostic commitment).
Ozet#
- On-Demand = esnek, pahali; tahmin edilemeyen isler
- Savings Plans = $/saat taahhut, esnek (aile/bolge degisebilir), Lambda/Fargate dahil
- RI = instance’a ozel taahhut; standard = en ucuz, convertible = degistirilebilir
- Spot = %90’a varan indirim, 2 dakika onceden geri alinabilir
- Dedicated Host = lisans/uyumluluk icin tek kiracili
- Is deseni → secim eslesmesini ogren, gerisi detay