Skip to content
All case studies

Case study 082022

Cloud Cost Program

A 41% cut in AWS spend with no loss of reliability, by making cost visible to the teams who create it.

Role
Program lead
Team
Platform team + 6 service teams
Timeline
5 months
Focus
Platform

−41%

AWS spend

$1.2M

saved per year

40% → 3%

untagged spend

Billing data is attributed to teams and services, which drives weekly reports, budget alerts and a ranked savings backlog.Billing data to Cost attribution; Cost attribution to Team reports (showback); Cost attribution to Savings backlog (top spend); Team reports to Budget alerts; Savings backlog to Changes shippedBilling dataAWS CUR · tagsCost attributionper team · per serviceTeam reportsweekly spend vs SLOsBudget alertsanomaly detectionSavings backlogranked by $ per dayChanges shippedGraviton · spot · sizingshowbacktop spend
Billing data is attributed to teams and services, which drives weekly reports, budget alerts and a ranked savings backlog.

The problem

The bill was one number that finance reviewed once a quarter. Untagged resources made up 40% of spend, staging ran around the clock, and every team assumed the cost belonged to someone else.

What I owned

  • A tagging policy enforced in Terraform and CI, and cost attribution per team and per service.
  • Weekly cost reports and anomaly alerts for every team.
  • The savings backlog, ranked by dollars saved per engineering day.
  • The largest changes: the Graviton migration, spot nodes for batch work and Kubernetes rightsizing.

Key decisions

  1. 1

    Show cost to the people who create it

    Once each team saw its weekly spend next to its SLOs, most of the waste was removed by its owners without being asked.

  2. 2

    Rank work by dollars per day

    Every idea was estimated for savings and effort. The top ten items delivered most of the reduction in the first two months.

  3. 3

    Reliability is the constraint, not the target

    Every change had to hold existing SLOs. Spot instances went to batch and CI work only, never to paths that serve requests.

Outcome

The bill fell 41% in five months and has grown more slowly than usage since. Teams now review their own cost weekly, with alerts when something spikes.

  • AWS
  • Terraform
  • Kubernetes
  • Karpenter
  • Graviton
  • Grafana