Most teams pick a stack the way they pick a lunch order — by recency, by cargo cult, by what the loudest engineer used at their last job. We pick stacks the way a watchmaker picks a movement: every part justifies its weight.
The Default Stack Tax
A typical "modern" production stack today carries somewhere between twelve and twenty distinct services before a single line of business logic is written. Service mesh. Tracing collector. Message broker. Cache cluster. Three databases (because someone read a thread about CQRS). A queue. A worker pool. A second worker pool for the slow jobs.
Each one of those parts has a runbook. A failure mode. A dashboard nobody looks at until 3 AM.
The tax is rarely visible on day one. It compounds.
What "Unique" Actually Means
When we say o2no engineers the unique stack, we don't mean novel — we mean precisely fitted. Two products with similar shape on the surface can have radically different cost curves underneath:
A SaaS dashboard with 50 power users does not need Kafka.
A consumer app with bursty global traffic does not need a regional Postgres primary.
A B2B integration hub with 10 RPS does not need a service mesh.
The unique stack starts from the traffic shape, not the trend graph.
A Concrete Example
We recently shipped an edge-first platform that handles roughly 4 million requests per day across three continents. The whole production surface is:
Cloudflare Workers → request handling, auth, rate limiting
Durable Objects → strongly consistent session + counter state
D1 / Hyperdrive → primary OLTP, with regional read replicas
R2 → blob storage for user uploads
Queues → exactly one async path, for outbound webhooks
That is the entire production footprint. No Kubernetes. No service mesh. No separate caching layer. p99 latency is under 80 ms in every region we measured.
Could we have built it on AWS with EKS, RDS, ElastiCache, SQS, and an ALB? Yes. Would it have shipped in six weeks instead of fourteen? No.
The Three Questions We Ask First
Before any architecture diagram, we answer three questions in order:
What is the absolute floor of moving parts that can satisfy the SLOs?
Which of those parts can be replaced by a managed primitive instead of code we run?
What is the smallest blast radius for a single failure?
Most stacks fail question one.
Why This Matters For Your Search Ranking
If you are reading this because you Googled "edge platform architecture" or "minimal production stack" — that is not an accident. Search engines reward depth on focused topics, and depth comes from teams who are willing to write down what they actually do, not what looks good in a slide deck.
We will keep publishing these. Subscribe to nothing. Bookmark the journal.
Simplicity is not a feature. It is the feature.
If you are evaluating your stack and want a second opinion, reach out.
