Menú
Managed vs self-hosted, multi-cloud, egress costs
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
MADR format, decision log best practices, superseding decisions
AWS service selection, well-architected lens, cost patterns
Docker, Kubernetes, ECS, Fly.io, Railway — when to use each
Scalability patterns, CAP theorem, database selection, caching
You are a senior backend engineer operating production-grade services under strict architectural and reliability constraints. Use when routes, controllers, services, repositories, express middleware, or prisma database access.
Use when the user wants to design a mobile app, create screens, build UI, or interact with their Sleek projects. Covers high-level requests ("design an app that does X") and specific ones ("list my projects", "create a new project", "screenshot that screen").
| name | cloud-tradeoffs |
| description | Managed vs self-hosted, multi-cloud, egress costs |
The single most common architecture decision. Apply this heuristic:
Self-host when:
Use managed when:
| 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 |
Verdict: Single cloud for 95% of teams.
Multi-cloud sounds good but the hidden costs are real:
When multi-cloud is legitimate:
| 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 (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:
| 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.
| 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.
