Cloud Architecture & Infrastructure

Managed Postgres (RDS, Cloud SQL, Supabase), object storage (S3, R2), a CDN (Cloudflare, Fastly), and a container runtime (ECS Fargate, Cloud Run, Fly.io) cover 90% of real workloads. Resist the urge to adopt every new managed service — each one adds a vendor lock-in surface, a billing line, and an on-call surface area.

Serverless functions are great for sporadic workloads, terrible for sustained throughput. A long-running container is dramatically cheaper than the equivalent Lambda usage above roughly 100k requests per day. Run the math at your actual traffic, not at the brochure traffic.

AWS, GCP, and Azure all charge meaningfully for data leaving their network. A SaaS product that serves images, videos, or large API payloads to end users will see egress fees dominate its cloud bill within 12 months of product-market fit. Cloudflare R2 and Bunny.net price egress at zero or near-zero and are now the default object storage choice for any workload that serves to a browser.

Place compute close to data. Cross-region database queries are slow, expensive, and the most common cause of mysterious latency spikes in distributed deployments.

Most teams do not need true multi-cloud deployment. They do need to avoid architectural decisions that make multi-cloud impossible if the strategic situation changes. That means: use Postgres-flavoured databases, S3-compatible object stores, OpenTelemetry-based observability, and container-based compute. Stay clear of proprietary lock-in features (DynamoDB, BigQuery as the primary store, Cosmos DB) unless their specific capabilities are load-bearing for the product.

Backup and restore should be tested quarterly with a real restoration to a separate account. We have seen disaster recovery plans that worked perfectly on paper and failed in practice because nobody had restored from cold backup in 18 months.