Cost Optimization and FinOps

Right-size instances, spot capacity, storage tiers and the metrics that keep cloud bills predictable.

ExpertOps11 min read

Read these first

Security Defence in Depth

Why this matters

StreamHub's monthly AWS bill exceeded $2.4M before FinOps practices — 30% was waste from over-provisioned instances, orphaned resources, and un-tiered storage.
Without cost visibility per team, no one owns the bill and optimisation becomes nobody's job.
Spot instances, reserved capacity, and storage lifecycle policies can cut compute costs 40–70% with minimal reliability trade-offs.

FinOps practices

Cost allocation tags — tag every resource with team, service, environment for chargeback.
Right-sizing — match instance type and count to actual CPU/memory utilisation (target 60–70%).
Reserved / savings plans — commit to 1-year usage for 30–40% discount on predictable workloads.
Spot / preemptible instances — 60–90% cheaper for fault-tolerant batch jobs (transcoding, ML training).
Storage tiering — lifecycle policies move cold data to cheaper tiers automatically.

FinOps observability loop

Tagged resources → Cost Explorer + Compute Optimizer → right-sizing actions.

Cost visibility dashboard

Service	Monthly cost	% of total	Optimisation opportunity
EKS compute	$980K	41%	Right-size pods; move batch to spot
RDS / Aurora	$420K	17%	Reserved instances; read replica audit
S3 storage	$310K	13%	Lifecycle to Glacier; delete incomplete uploads
CloudFront CDN	$280K	12%	Already optimised; cache hit ratio 94%
Kafka (MSK)	$190K	8%	Reduce retention from 7d to 3d for non-critical topics
Other	$220K	9%	Orphaned EBS volumes, unused Elastic IPs

EKS compute
Monthly cost$980K
% of total41%
Optimisation opportunityRight-size pods; move batch to spot
RDS / Aurora
Monthly cost$420K
% of total17%
Optimisation opportunityReserved instances; read replica audit
S3 storage
Monthly cost$310K
% of total13%
Optimisation opportunityLifecycle to Glacier; delete incomplete uploads
CloudFront CDN
Monthly cost$280K
% of total12%
Optimisation opportunityAlready optimised; cache hit ratio 94%
Kafka (MSK)
Monthly cost$190K
% of total8%
Optimisation opportunityReduce retention from 7d to 3d for non-critical topics
Other
Monthly cost$220K
% of total9%
Optimisation opportunityOrphaned EBS volumes, unused Elastic IPs

Example StreamHub cost breakdown — EKS compute is the largest optimisation target.

Right-sizing Kubernetes workloads

Most pods are over-provisioned. Use Vertical Pod Autoscaler (VPA) recommendations or Grafana metrics.

Java

# Before: over-provisioned
resources:
  requests:
    cpu: 2000m
    memory: 4Gi
  limits:
    cpu: 4000m
    memory: 8Gi

# After: right-sized based on p95 utilisation
resources:
  requests:
    cpu: 500m
    memory: 1Gi
  limits:
    cpu: 1000m
    memory: 2Gi

Spot instances for batch workloads

StreamHub's video transcoding pipeline runs on spot instances with fallback.

Java

apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
  name: transcoder-spot
spec:
  template:
    spec:
      requirements:
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot"]
        - key: node.kubernetes.io/instance-type
          operator: In
          values: ["c6i.4xlarge", "c6i.8xlarge"]
      taints:
        - key: workload
          value: transcoding
          effect: NoSchedule
  limits:
    cpu: 500
    memory: 1000Gi
  disruption:
    consolidationPolicy: WhenUnderutilized

Aspect	Strategy	Savings
Reserved instances (1yr)	30–40% off on-demand	Predictable baseline workloads
Spot instances	60–90% off on-demand	Fault-tolerant batch, stateless workers
Savings Plans	Flexible commitment across instance families	Mixed workload environments
Graviton (ARM)	20% better price-performance	Java/Go services with ARM builds
S3 Intelligent-Tiering	Auto-moves to cheaper tier	Unpredictable access patterns

Reserved instances (1yr)
Strategy30–40% off on-demand
SavingsPredictable baseline workloads
Spot instances
Strategy60–90% off on-demand
SavingsFault-tolerant batch, stateless workers
Savings Plans
StrategyFlexible commitment across instance families
SavingsMixed workload environments
Graviton (ARM)
Strategy20% better price-performance
SavingsJava/Go services with ARM builds
S3 Intelligent-Tiering
StrategyAuto-moves to cheaper tier
SavingsUnpredictable access patterns

Combine strategies: reserved for baseline, spot for burst, Graviton where compatible.

Storage lifecycle automation

Java

{
  "Rules": [
    {
      "ID": "tier-streamhub-logs",
      "Filter": { "Prefix": "logs/" },
      "Status": "Enabled",
      "Transitions": [
        { "Days": 30, "StorageClass": "STANDARD_IA" },
        { "Days": 90, "StorageClass": "GLACIER_IR" }
      ],
      "Expiration": { "Days": 365 }
    },
    {
      "ID": "cleanup-temp-uploads",
      "Filter": { "Prefix": "uploads/temp/" },
      "Status": "Enabled",
      "Expiration": { "Days": 7 }
    }
  ]
}

Eliminating waste

Common waste sources

Orphaned EBS volumes — detached volumes from terminated instances; automate cleanup with Lambda.
Over-retained logs — 7-day Kafka retention for debug topics that nobody reads after 24 hours.
Idle load balancers — NLBs/ALBs with zero traffic still cost $20/month each.
Uncompressed data — enable S3 compression for JSON logs; 5x storage reduction.
NAT gateway charges — $0.045/GB processed; use VPC endpoints for S3/DynamoDB to bypass NAT.

Quick recall

Everything you need if you only revisit this box.

FinOps = cost visibility + accountability + continuous optimisation, not one-time audits.
Tag every resource for chargeback; right-size pods to 60–70% utilisation.
Reserved instances for baseline; spot for fault-tolerant batch; Graviton for compatible workloads.
S3 lifecycle policies auto-tier cold data; delete temp uploads after 7 days.
Weekly cost reviews and waste audits (orphaned volumes, idle LBs, over-retained logs).

Test yourself

Answer these before moving on — recall is what makes it stick.