Cost optimization

Cost Optimization Skill

Setup Check

Before loading context files, check if context/CONTEXT.md exists in the current directory.

If context/CONTEXT.md exists — read it and proceed normally.

If context/CONTEXT.md does not exist — this skill was installed standalone (e.g. via npx skills add). Ask the user these questions before proceeding:

1. Role — junior-sre / senior-sre / sre-manager (shapes output depth and tone)
2. Cloud provider — aws / gcp / azure / on-prem / hybrid
3. Observability stack — e.g. Datadog, Prometheus+Grafana, New Relic
4. Company name and primary services affected (if relevant to this task)

Use the answers inline for this session. For persistent setup across all skills, suggest:

pipx install sre-agent
sre-agent init

Instructions

Step 1: Load Context

Read context/company/tech-stack.md for cloud provider and platform.

Step 2: Identify Cost Drivers

Ask the user to share:

• Top cost categories from the billing dashboard (compute, storage, data transfer, etc.)
• Any recent cost spikes (when and how much)
• Current monthly total and recent trend

Classify costs into:

Category	Typical % of bill	Optimization potential
Compute (EC2/GCE/VMs)	40–60%	High
Managed services (RDS, ElastiCache)	15–25%	Medium
Storage (S3, EBS, GCS)	5–15%	Medium
Data transfer / egress	5–20%	Variable
Kubernetes (EKS/GKE)	10–20%	High
Logging / observability	5–15%	Medium

Step 3: Waste Identification

Idle resources:

• EC2/VM instances with CPU < 5% sustained
• Unattached EBS volumes
• Unused Elastic IPs
• Unused load balancers
• Old snapshots (> 90 days with no restore)
• Dev/staging resources running 24/7 (schedule to turn off nights/weekends)

Overprovisioned resources:

• Instances where CPU is consistently < 20% and memory < 40%
• Database instances with < 20% utilization
• Over-specified Kubernetes resource requests (pod requests >> actual usage)
• EBS volumes that are only 20–30% full but sized for 10x current data

On-demand vs. commitment opportunity:

• Instances running > 3 months on-demand → Reserved Instance or Savings Plan candidate
• Steady-state workloads → Compute Savings Plans (1-year, no-upfront)
• Flexible workloads → Spot Instances (batch jobs, CI/CD workers, stateless services)

Step 4: Rightsizing Analysis

For compute rightsizing:

Rightsizing recommendation logic:
  If avg CPU < 20% and avg memory < 40%:
    → Downsize to next smaller instance type
  If avg CPU > 70% or avg memory > 80%:
    → This is correctly sized or may need upsizing
  If CPU is high but memory is low:
    → Consider compute-optimized instance type
  If memory is high but CPU is low:
    → Consider memory-optimized instance type

Step 5: Kubernetes Cost Optimization

• Compare pod resource requests vs. actual usage (Vertical Pod Autoscaler recommendations).
• Identify namespaces or workloads with no resource limits (risk of resource starvation and overspend).
• Check if cluster autoscaler is configured and working.
• Evaluate whether Spot/Preemptible nodes can be used for non-critical workloads.
• Check node utilization: if < 40% average, the cluster may be over-provisioned.

Step 6: Storage Optimization

• S3: lifecycle rules to transition to Infrequent Access after 30 days, Glacier after 90 days.
• EBS: delete unattached volumes, rightsize over-provisioned volumes.
• Log retention: audit log retention periods (7 days for debug logs, 30 days for info, 1 year for audit).
• Old snapshots: automate deletion of snapshots older than retention policy.

Step 7: Generate Recommendations

Rank by: (monthly savings) × (implementation ease) / (reliability risk)

Recommendation Report — Cloud Cost Optimization

Current monthly spend: $XX,XXX
Potential savings identified: $X,XXX/month (XX%)

Priority 1: Reserved Instances for steady-state EC2 (save ~$1,200/mo)
  Risk: Low | Effort: Low (1-click purchase) | Payback: immediate
  Action: Purchase 1-year no-upfront Compute Savings Plan for $X commit/hr

Priority 2: Schedule dev/staging environment off-hours (save ~$800/mo)
  Risk: None | Effort: Low | Payback: immediate
  Action: Lambda function to stop instances 6pm–8am weekdays, all weekend

Priority 3: Rightsize 8 overprovisioned RDS instances (save ~$600/mo)
  Risk: Medium (test in staging first) | Effort: Medium
  Action: See rightsizing table below

Priority 4: S3 lifecycle rules for logs bucket (save ~$200/mo)
  Risk: None | Effort: Low
  Action: Add lifecycle rule: IA after 30d, Glacier after 90d

Total addressable: ~$2,800/month (~28% reduction)

---

Guidelines

• Never recommend a cost saving that introduces a reliability risk without explicitly stating the trade-off and getting confirmation.
• Spot instances are only appropriate for stateless, fault-tolerant workloads.
• Reserved Instances are a financial commitment — confirm with finance/management before purchasing.
• Persona (sre-manager): frame the savings as ROI and include the time to implement each item. Calculate cost of engineering time to implement vs. expected savings.