Infrastructure
Managed Services Always
Engineers should focus on the product, not setting up services that are already solved. Managed services have more ROI than custom infrastructure for most companies.
- Vercel, serverless, managed databases, managed observability.
- Custom infra only when managed services cannot meet specific requirements (compliance, scale, cost).
Phase-Dependent Decisions
Infrastructure choices evolve with the product phase:
- Prototyping: maximize speed. Vercel, managed everything. No production environment needed yet.
- Pre-release: production readiness requires more control. May move to cloud providers (GCP, AWS) with Terraform and CI/CD.
- Scale: performance bottlenecks drive specific decisions. Elasticsearch for search, read replicas, caching layers.
Each phase has different constraints. The infrastructure that was right for prototyping may be wrong for production.
Security and Performance
Pragmatic and phase-dependent. Customer personal information is always critical. Beyond that, risk tolerance depends on company phase.
- External security audits before production launch.
- For performance: define what “bad” looks like, move fast until monitoring shows problems.
- Monitoring is the key — you cannot optimize what you cannot see.
Decision Criteria
When evaluating infrastructure: managed service by default. Only move to custom when the managed option fails on compliance, scale, or cost. The burden of proof is on custom infra — it must justify the operational overhead.
When choosing between speed and control: match to phase. Prototyping needs speed. Production needs reliability. Scale needs data-driven decisions. Do not over-engineer for a future that may not arrive.
Anti-patterns
- Custom infra for solved problems — building your own logging, monitoring, or deployment pipelines when managed services exist.
- Same infrastructure for every phase — Vercel is right for prototyping, may be wrong for production at scale. Recognize when constraints shift.
- Performance optimization without monitoring — guessing at bottlenecks. Invest in visibility first.
Experience Notes
Previously: Vercel for everything. Now: Vercel for prototyping and static sites, cloud providers (GCP, AWS) for production workloads that need more control. The transition happened when production requirements outgrew what managed PaaS could handle at acceptable cost.