Skip to content
intermediate Phase 17 · AWS Cost Management

Savings Plans & Reserved

Save costs with Savings Plans, Reserved Instances, and Spot.

1h
0 problems
Topic Progress 0%

The Commitment Menu

The Commitment Menu

Four levers cut compute cost, each with a different flexibility/ discount trade:

  • EC2 Instance Savings Plans — commit to a specific instance family in a region for 1 or 3 years; deepest discount (AWS advertises up to 72%) but tied to the family.
  • Compute Savings Plans — commit to a $/hour amount usable by any EC2 instance family/size/OS/region, plus Fargate and Lambda; more flexible, somewhat smaller discount.
  • Reserved Instances — the older sibling of Instance SPs: zonal RIs include a capacity reservation, regional RIs give flexibility across sizes in the family.
  • Spot — bid on spare capacity at up to ~90% off with interruption anytime (two-minute notice when available); for fault-tolerant, flexible workloads only.

Payment options — all upfront, partial, none — trade cash flow against discount depth.

Matching Workload to Commitment

Matching Workload to Commitment

Walk the decision:

  1. Steady baseline, known family (databases, always-on fleets) → EC2 Instance SP or zonal RI — you get discount and, for zonal RIs, capacity certainty.
  2. Mixed/unknown compute, multi-account → Compute SP: one commitment covers EC2 across families, Fargate tasks, and Lambda.
  3. Interruptible, elastic work (CI, batch, ML training, workers) → Spot, diversified across instance types inside an Auto Scaling group or Spot Fleet so one pool drying up does not halt you.
  4. Spiky but latency-sensitive → keep the baseline on SP, burst on-demand.

Verify with Cost Explorer reports: utilization (did you use what you bought?) and coverage (how much of total spend is discounted?).

Pitfalls and Interview Focus

Pitfalls and Interview Focus

The expensive mistakes: committing to an experimental workload (two-year SP on a service you sunset in six months — sell or exchange before it rots); buying zonal capacity you never consume; mixing Spot into synchronous user paths without checkpointing and graceful fallback to on-demand. Spot interruptions are normal operations — handle the rebalancing signal and make workers idempotent.

Remember discounts apply only to compute, not data transfer, support, or third-party licenses; and Savings Plans apply automatically across the account family, so "which instance does my SP cover" is answered by coverage reports, not spreadsheets.

Interview: state flexibility vs discount as the axis, pick per workload, and mention how you monitor utilization after the purchase.

Quiz

1. Which commitment covers Lambda and Fargate as well as EC2?

Question 1 options

2. What does a zonal Reserved Instance provide beyond discount?

Question 2 options

3. Where should interruptible ML batch training run?

Question 3 options

Flashcards

Question

Compute vs Instance SP

Answer

Compute: any family + Fargate/Lambda, flexible. Instance SP: family-locked, deeper discount.

Question

Zonal vs regional RI

Answer

Zonal = discount + capacity reservation; regional = discount, flexible across sizes.

Question

Spot interruption

Answer

Capacity reclaimed anytime, usually with a 2-minute notice — design for rebalancing.

Question

Utilization vs coverage

Answer

Utilization: bought vs used. Coverage: share of total spend that is discounted.

Question

Payment options

Answer

All/partial/none upfront — more upfront buys a deeper discount.

Revision Notes

Key Takeaways

  • 1. Flexibility and discount depth are inversely related — pick deliberately per workload
  • 2. Spot is a reliability-architecture decision, not just a price
  • 3. Utilization/coverage reports are the post-purchase health check

Interview Tips

  • • Take a workload description and choose SP type/term, citing flexibility needs and stability
  • • Mention monitoring after purchase — interviewers hear enough about buying, not managing

Cheat Sheet

Stable family → Instance SP/RI (zonal=+capacity); mixed → Compute SP (EC2+Fargate+Lambda); interruptible → Spot+checkpoints. Track utilization+coverage; commit only to proven baselines.

Continue Learning