Rightsizing, spot and reserved
Match each machine to the load it carries, then buy the steady part on a commitment and the interruptible part at a deep discount.
Shrink machines to the load they carry, then pay the cheapest price each part of the fleet can accept.
Use it when#
Utilization graphs sit under about a third at peak, so you pay for idle capacity.
Load has a steady floor that never goes away, which is the part to commit to.
Some work can be interrupted and retried, such as batch jobs or stateless workers.
How it works#
Rightsizing comes first, because a discount on an oversized machine saves less than removing the oversize. Pick the instance family by the bottleneck: memory-heavy for caches, compute-heavy for CPU-bound code. ARM families, such as AWS Graviton, run many workloads for about 20% less if the software ports.
Then buy by predictability. On-demand has no commitment. Reserved instances or savings plans trade a one- or three-year commitment for roughly 30 to 70% off. Spot capacity, up to about 90% off, can be reclaimed with a two-minute warning on AWS.
Example: eight web servers where six are always busy and two handle peaks. Reserve the six, run the two on spot behind the same load balancer, and the loss of a spot box only trims capacity.
What it costs#
Commitments are paid whether the box is used or not, so reserving before rightsizing locks in waste. Spot fails at the worst moment unless the app shuts down cleanly and a non-spot copy stays up. Never put a primary database on spot.
In Quorum#
Pricing is a setting on compute Nodes from Series A, and the Design below splits one fleet three ways.
Drag to pan. Hold Ctrl or Cmd and scroll to zoom.
| Where | Setting | Value | Why |
|---|---|---|---|
| App Server | Instance size | medium | Sized for the steady load, which is what a commitment should cover. |
| App Server | Pricing model | reserved | The always-on box takes the reserved discount in return for a one-week term. |
| App Server | Term | 10080 | One calendar week, in sim minutes, so the discount applies. |
| App Server | Instance size | medium | The same size, so the Load Balancer treats the two alike. |
| App Server | Pricing model | spot | The box that can vanish takes the deepest discount, because the steady box still serves if it is reclaimed. |
| Database | Instance size | medium | Medium for the Database, which has no spot or reserved choice. |
Research this Design needs: Load balancing, Autoscaling, Reserved instances, Spot instances, Health checks, Load balancing.
Reserved instances and Spot instances both need Autoscaling. A reserved Node cannot go back to on-demand until its term ends. Rightsizing report marks Nodes that never get close to busy at peak, and ARM instances add a cheaper family. Savings plans discount every compute Node without a per-Node term. The Spot reclaim incident shows what an interruption costs a fleet with no on-demand copy. The spot instance glossary entry and Cut the cloud bill cover the rest.