Architecture reference
Reference architectures
Blueprints you can study and steal. These are the patterns 80% of companies actually run, made interactive so you can poke at them. Click any node for what it does, when to add it, what it costs, and what to use instead.
New here? Read the companion essay The 80% architecture for the named components, costs, and the symptoms that say you have outgrown the default.
The Starter SaaS
One VM, one database, a cron job. The MVP that boring engineering loves. Fewer moving parts than a hammer.
About 50 EUR/month all-in. Good for 0 to 10k users. Do not outgrow it for fun. Most pre-product-market-fit startups should live here longer than they do.
The 80% Standard SaaS
Stateless app fleet, Postgres, Redis, a queue, workers, and the usual external SaaS. The default that 80% of companies should be running.
About 500 to 2000 EUR/month for a healthy production deployment. Comfortable from 10k to 1M users. The default that 80% of companies should be running and most are not.
Modern AI-augmented SaaS
The 80% standard SaaS plus a 2026 LLM stack: gateway, vector DB, MCP, agents. Same bones as before, with a brain bolted on.
Add 200 to 1000 EUR/month on top of the standard SaaS for inference and a vector DB. Most of the gain comes from prompt caching and not from picking the cheapest model.
Excel-to-BI: the data team architecture
Sources on the left, dbt and warehouse in the middle, dashboards on the right, all stitched together by Airflow. Every reporting team has a version of this.
The shape barely changes from one company to another. Sources -> ingestion -> raw lake -> dbt -> warehouse -> BI. The interesting variation is which tools you pick at each step, not the topology.
Medallion lakehouse
Raw to bronze to silver to gold on object storage and Iceberg. Each layer is a separate dbt model with its own tests, owners, and access.
Bronze keeps source-of-truth raw. Silver is the cleaned, conformed model. Gold is what business users query. Each step is a separate dbt model with separate testing and access control.
Event-driven microservices
Services do not call each other. They publish events to Kafka and consumers replay the log to derive their own state. Sagas and CQRS sit on top of this shape.
Services do not call each other directly. They emit events. Consumers replay the log to derive their own state. Saga and choreography sit on top of this shape.
Multi-region active-active
Three regions serving traffic. Reads local, writes to a primary region. The hard part is not the compute, it is conflict resolution.
Active-active means every region serves traffic. Reads are local, writes go to the primary region. The hard problem is conflict resolution; most teams accept eventual consistency for the gain.
Data team toolchain
The whole stack: sources, ingestion, lake, warehouse, dbt, quality, BI, alerts, CI. The shape every mid-size data team converges on.
Most data teams above 5 engineers run a version of this stack. Names rotate, the shape does not.
Want more?
More patterns are coming: the IoT-to-warehouse pipeline, the multi-tenant B2B SaaS, the data-mesh layout, and the on-prem-to-cloud migration shape. If there is one you would like to see next, send a note via the contact link in the footer.