cloud-tradeoffs

name	cloud-tradeoffs
description	Managed vs self-hosted, multi-cloud, egress costs

Cloud Architecture Tradeoffs

Managed vs Self-Hosted

The single most common architecture decision. Apply this heuristic:

Self-host when:

• The managed version costs >3x self-hosted at your scale
• You have the ops expertise to run it well
• Compliance/data residency requires it
• The tool is a core competency (you build on top of it)

Use managed when:

• The ops burden of self-hosting would distract from product work
• The team doesn't have deep expertise in that system
• You're pre-scale and need to move fast
• The managed version has SLAs you can't match yourself

Tool	Managed	Self-Hosted
PostgreSQL	RDS Aurora, Supabase, Neon	EC2 + pgBackup, Railway
Redis	ElastiCache, Upstash	EC2, Fly.io
Kafka	MSK, Confluent Cloud	EC2 cluster, Redpanda
Elasticsearch	OpenSearch Service	EC2, Hetzner
Kubernetes	EKS, GKE, AKS	k3s on EC2/Hetzner
Object Storage	S3, GCS	MinIO
Auth	Cognito, Auth0, Clerk	Keycloak, Ory Hydra

Multi-Cloud vs Single Cloud

Verdict: Single cloud for 95% of teams.

Multi-cloud sounds good but the hidden costs are real:

• Egress fees when data crosses clouds ($0.08–0.12/GB)
• Each cloud has different IAM, networking, monitoring — ops complexity doubles
• You lose cloud-native integrations (e.g., Lambda → SQS → DynamoDB is much tighter than cross-cloud)
• "Avoiding vendor lock-in" is often theoretical — you're already locked into your own architecture

When multi-cloud is legitimate:

• Regulatory requirement to have data in specific jurisdictions
• Disaster recovery across independent failure domains
• Specific services only available on one cloud (e.g., Azure OpenAI, GCP BigQuery)

Cloud Provider Selection

Strength	Provider
Broadest service catalog, market leader	AWS
AI/ML workloads, BigQuery, GKE	GCP
Enterprise, Microsoft ecosystem, OpenAI	Azure
Simplicity, predictable pricing, small teams	DigitalOcean / Fly.io / Railway
Cost-efficient raw compute (Europe-heavy)	Hetzner
Edge compute, CDN-first	Cloudflare Workers

For most startups: AWS or GCP. AWS if you want the most hiring pool and tooling. GCP if ML/data is core.

Egress Cost Trap

Egress (data leaving a cloud provider) is the hidden tax of cloud:

Provider	Egress cost (approx)
AWS	~$0.09/GB to internet
GCP	~$0.08/GB to internet
Azure	~$0.087/GB to internet
Cloudflare R2	$0/GB (zero egress)
Hetzner	~€0.01/GB (very cheap)

Decisions affected by egress:

• Video / media delivery → use CDN (CloudFront, Cloudflare) to avoid repeated origin egress
• Cross-region replication → quantify the egress cost before designing multi-region
• Object storage for large files → consider Cloudflare R2 if you serve assets directly

Serverless vs Always-On

Factor	Serverless (Lambda, Cloud Run)	Always-On (ECS, K8s)
Traffic pattern	Spiky or unpredictable	Steady / predictable
Cold start tolerance	Yes	No
Max execution time	15 min (Lambda)	Unlimited
Stateful workloads	No	Yes
Cost at low traffic	Very cheap (pay per request)	Fixed cost regardless
Cost at high traffic	Can exceed always-on	Cheaper per request
Ops burden	Very low	Higher

Breakeven rule of thumb: If your Lambda functions run more than ~50% of the time, EC2/Fargate is cheaper.

Infrastructure as Code: Tool Selection

Tool	When to use
Terraform / OpenTofu	Multi-cloud, team already knows HCL, most common
Pulumi	Prefer real languages (TypeScript, Python, Go) over HCL
AWS CDK	AWS-only, team is TypeScript/Python native
Ansible	Configuration management, not infra provisioning
Helm	Kubernetes application packaging
Crossplane	K8s-native IaC, control plane for cloud resources

Default: Terraform (OpenTofu if you want truly open-source). Switch to Pulumi if the team struggles with HCL or needs complex logic.

Cost Optimization Patterns

1. Right-size before you reserve. Run on on-demand for 1–2 months, analyze actual utilization, then buy Reserved Instances / Savings Plans.
2. Spot instances for fault-tolerant workloads. 60–90% cheaper. Use for batch jobs, workers, CI runners — not stateful services.
3. Auto-scaling over over-provisioning. CPU/memory-based scaling for ECS/EC2; DynamoDB on-demand for spiky DB traffic.
4. S3 Intelligent-Tiering for data you don't access frequently. Automatic cost reduction, no retrieval penalties for Frequent and Infrequent tiers.
5. Delete what you don't use. Unattached EBS volumes, unused Elastic IPs, old snapshots, and forgotten load balancers silently accumulate cost.