Skip to content

Right-Sizing — Resources at the Right Size#

"A team that allocates 4 vCPU to an instance averaging 15% CPU is burning money on 75% of it every month. Right-sizing = downsize the instance based on the usage profile."

This guide covers right-sizing practices, tooling, and the "when to downsize" decision for AWS / GCP / K8s.

🎯 Right-Sizing Targets#

Resource Ideal utilization
CPU 50-70% average
Memory 60-80%
Disk 40-70% (growth buffer)
Network 30-60%

🔑 80%+ utilization = capacity risk. Below 20% = waste.

📊 Profile-Based Sizing#

Step 1: Collect current usage#

# CPU p95 (1 week)
quantile_over_time(0.95,
  rate(container_cpu_usage_seconds_total[5m])[7d:1h]
)

# Memory peak
max_over_time(
  container_memory_working_set_bytes[7d]
)

Step 2: Compute the right size#

New size = max(p95 × 1.2, p99 × 1.0)

Example: CPU p95 = 600m, p99 = 900m
New = max(720m, 900m) = 900m → 1000m (round up)

Step 3: Validate (1 week of observation)#

  • Any throttling during HPA scaling spikes?
  • OOMKilled?
  • Latency impact?

🛠️ Cloud-Side Right-Sizing#

AWS Compute Optimizer#

aws compute-optimizer get-ec2-instance-recommendations \
  --account-ids <ACCT>

→ A specific recommendation like "recommend i3.2xlarge → i3.large".

GCP Recommender#

gcloud recommender recommendations list \
  --project=<PROJECT> \
  --recommender=google.compute.instance.MachineTypeRecommender

Azure Advisor#

az advisor recommendation list --category Cost

🐳 K8s Right-Sizing#

VPA (Vertical Pod Autoscaler)#

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: payments-vpa
spec:
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: payments
  updatePolicy:
    updateMode: "Off"   # recommendation only

kubectl describe vpa payments-vpa
# Recommendation:
#   cpu: 234m
#   memory: 512Mi

→ A human reviews it and applies it manually to the deployment.

Kubecost#

helm install kubecost kubecost/cost-analyzer \
  -n kubecost --create-namespace \
  --set kubecostToken=<TOKEN>

UI: per-namespace + per-pod right-sizing recommendation. Kubecost cluster-wide cost dashboard.

Goldilocks (open-source)#

helm install goldilocks fairwinds-stable/goldilocks

→ Auto-generates a VPA recommendation per namespace + dashboard.

📦 Database Right-Sizing#

Postgres / MySQL#

-- Connection count
SELECT count(*) FROM pg_stat_activity;

-- Cache hit ratio
SELECT
  sum(blks_hit) / nullif(sum(blks_hit + blks_read), 0)::float
FROM pg_stat_database;
-- > 0.99 = optimal
-- < 0.95 = increase shared_buffers

Sizing rule#

  • Burstable workload → t-class (AWS) / e2 (GCP) / B-series (Azure)
  • Sustained → r-class (memory) or m-class (general)
  • CPU-heavy → c-class

🔑 Burstable instances choke under 24/7 high CPU. Choose based on the profile.

🚧 Right-Sizing Risks#

1. Over-downsizing → crash under spike#

  • Leave 20-30% headroom
  • Validate with a load test

2. Memory under-provision → OOMKilled#

  • Base the memory peak on max, not p99

3. Insufficient storage IOPS#

  • Watch the IOPS limit when switching gp3 → gp2

4. Reduced network throughput#

  • Smaller instance = less bandwidth

🚫 Anti-Pattern Table#

Anti-pattern Why it's bad Do this instead
Guess "make it big, we'll downsize later" $X wasted monthly Profile-based
Right-size once and forget Workload changes Quarterly review
VPA Auto + HPA Auto on the same resource Race VPA Off mode
Downsize memory → OOM Restart cycle Peak buffer
Downsize disk Migration is hard Generous + alarm at 80%
Reserved + right-size afterwards Wasted commitment Right-size first, reserved after
Same size for production and dev Dev waste Per-env sizing

📋 Right-Sizing Checklist#

[ ] Quarterly right-sizing review
[ ] Cloud Compute Optimizer / Advisor / Recommender
[ ] K8s: VPA Off mode + Goldilocks dashboard
[ ] Kubecost (per-namespace cost)
[ ] CI gate: Kubecost diff on the PR (PR-Cost-Diff.md)
[ ] Per-env sizing (dev small, prod profile-based)
[ ] Correct class for burstable vs sustained workload
[ ] Memory: p99 + buffer (OOM prevention)
[ ] Storage: 20%+ headroom + alarm
[ ] DB connection pool (PgBouncer)
[ ] Quarterly cost report to management

📚 References#

"Right-sizing isn't a 'one-time cleanup' — it's a quarterly discipline. Profiles change, traffic grows and shrinks. An undisciplined team's AWS bill balloons 20% every 6 months."

🎓 Learning Path: This document is used as a "Read first" resource in the F1 module.