Cost Optimization — Basics

Cloud Cost Optimization — Basics

The 3 levers

1. Use less (efficiency).
2. Pay less per unit (commitments / spot).
3. Stop using what you don’t need (cleanup).

Most “cost reductions” are #3 first, then #1, then #2.

Common waste sources (AWS but generally applicable)

• Idle / oversized EC2 — instances ~10% utilized.
• Old snapshots & unattached EBS volumes.
• Idle ELBs / NAT Gateways (NAT GW is hourly + per-GB; surprisingly costly).
• Untagged resources — no one knows who owns them.
• Cross-AZ traffic — between AZs has cost (often surprising).
• NAT egress for AWS service calls — solve with VPC endpoints.
• Public egress — biggest line item often. Cache via CloudFront, compress, batch.
• CloudWatch log retention — 30d default → forever if not set.
• Pre-prod environments running 24/7 — schedule shut-down.
• Lambda cold start with provisioned concurrency at low traffic.
• Old machine images / ECR images.
• Unused RDS read replicas.
• DynamoDB provisioned, not on-demand at low or spiky usage.
• S3 versions and incomplete multipart uploads.
• GPU sitting idle.

Pricing patterns

• On-demand — full price.
• Spot — 60-90% off, can be interrupted (good for batch, ML training, stateless workers).
• Savings Plans / Reserved — 30-60% off for 1-3y commitment.
• Committed-use discounts (GCP) similar.
• Volume discounts by tier.
• Egress is where cloud providers make money — avoid where possible.

Right-sizing

• Look at p95 CPU / memory utilization.
• Drop to next instance size if < 50% utilized at peak.
• ARM (Graviton on AWS, Ampere on GCP) is ~20% cheaper for many workloads.
• Use Compute Optimizer / Recommender.

Storage optimization

• S3 lifecycle: Standard → IA → Glacier → Delete.
• Block storage: gp3 over gp2 (cheaper, faster).
• Cold data → Glacier Deep Archive (~$1/TB/mo).
• Delete old snapshots, AMIs.
• Compress logs (gzip / zstd).

Network optimization

• VPC Endpoints for S3, DynamoDB, SSM, ECR — avoid NAT egress fees.
• CloudFront / CDN for content — cheaper outbound and faster.
• Co-locate services in same AZ where possible.
• Use PrivateLink / VPC peering for cross-account/VPC traffic.
• Compression (gzip/br) at LB.

Containers / K8s

• Cluster autoscaler / Karpenter — provision nodes only when needed.
• HPA for pods.
• Right-size requests — over-requesting wastes capacity.
• Spot node pools for batch, fault-tolerant.
• Bin-packing — small workloads share nodes.
• PDB + anti-affinity carefully — over-spreading wastes capacity.

Lambda / serverless

• ARM (cheaper).
• Reduce package size (faster cold starts, faster invocation = cheaper).
• Memory tuning — sometimes more memory = faster = cheaper net.
• Provisioned concurrency only for hot paths.
• Step Functions for orchestration vs chained Lambdas.

Database

• Aurora Serverless v2 for spiky low-volume workloads.
• DynamoDB on-demand for unpredictable; provisioned + auto-scaling for steady.
• Read replicas only if read load needs them.
• Backup retention review.
• Reserved instances on RDS for steady workloads.

Tagging discipline

Without tags, you can’t allocate cost or accountability.

• Environment (prod / staging / dev)
• Owner (team or email)
• Project / Service
• CostCenter

Enforce via SCP / Organization Policy: deny resource creation without required tags.

Visibility

• AWS: Cost Explorer, Cost & Usage Reports → Athena, Cost Anomaly Detection.
• GCP: Billing reports, Recommender, BigQuery billing export.
• Cross-cloud: Vantage, CloudHealth, Apptio Cloudability, Infracost.
• Per-PR: Infracost estimates Terraform diff.

FinOps practices

• Show team / service-level cost dashboards.
• Monthly cost reviews per service.
• Per-feature unit economics ($/request, $/customer).
• Budget alerts at 50/80/100%.
• Anomaly alerts (sudden spikes).
• Quarterly right-sizing review.

Common interview Qs

1. Bill went up 30% this month — investigate. Cost Explorer by service; correlate with deploys / traffic; identify top contributors.
2. Reduce Lambda cost. Right-size memory, ARM, batch invocations, reduce concurrency, async path where possible.
3. DynamoDB cost spike. Check capacity mode, hot partitions, scan vs query, TTL configured.
4. Cross-AZ traffic high. Co-locate components, single-AZ for non-critical, gateway endpoints.
5. Should we use spot? For batch, CI runners, dev clusters, ML training, stateless workers — yes. For stateful single-replica DBs, no.
6. NAT Gateway cost. Use VPC endpoints, reduce egress, consider single shared NAT vs per-AZ depending on resilience need.