The eight sessions 🎛️ · one platform, architected across all of them
The running project is Daybreak, the coffee-subscription brand you already met in the SQL and warehouse courses. It is three years old, it has one Postgres, and its dashboards query production. You are the person who has to decide what it becomes next, and write down why.
Requirements to constraints
Nobody hands you an architecture problem. They hand you a complaint. Turning "the dashboard is slow" into four numbers you can design against, on paper, before any technology is named.
🧱 Session 2 · easyBoundaries and contracts
Where one team's problem stops being another team's problem. The three couplings that actually hurt, what belongs in a data contract, and why the producer's build should go red instead of your dashboard.
🗄️ Session 3 · mediumStorage and serving
Row store, column store, lake, stream - chosen by access pattern, not by fashion. The freshness-cost curve, PACELC, and why designing to the average latency is designing for nobody.
🕸️ Session 4 · mediumOrchestration topology
Cron knows what time it is. An orchestrator knows what finished. Idempotency as a design property, backfills that are boring, and SLAs on data rather than on services.
🎛️ Session 5 · hands-onDesigning for failure
The tradeoff board. Nine levers, four meters, two requirements - and the moment where switching the requirement without touching a lever flips the verdict. Plus: retries do not fix a contract break.
🛡️ Session 6 · hands-onThe governed shell
Access, lineage, PII boundaries, retention, deletion and cost attribution - designed at the boundary rather than bolted on afterwards, because deletion cannot be retrofitted into a lake.
🤖 Session 7 · hardestAI as a system component
A model is a latency budget, a cost line and a failure mode. Retrieval indexes go stale, agent loops have unbounded cost, and DORA 2025 found AI raises throughput and instability at the same time.
📝 Session 8 · hardestThe design review
ADRs, one-way and two-way doors, and the four questions a good reviewer asks. Capstone: take apart a deliberately broken Daybreak architecture and write the decision record that fixes it.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Eight stamps means you have argued your way through every meter on the board at least once. Stamps live in this browser only.
What this course is not, so you land in the right place 🧭
| If your question is | Go here instead |
|---|---|
| What layers does a data platform have, and what do they cost? | learn-ai-infra-with-phoebe - an inventory and an investment case |
| How do I build the ingestion-to-serving pipeline? | learn-data-engineering-with-phoebe - construction, on real DuckDB |
| How do I test, deploy and operate it? | learn-dataops-with-phoebe - CI/CD, MLOps, DBOps |
| What should the schema look like? | learn-data-modeling-with-phoebe - dimensions and facts |
| Which of these do I choose, and how do I defend it? | you are in the right place |
Choose your path 🗺️
Three doors into the same conviction: an architecture is not a diagram, it is a set of trades you can name and defend.
What is honest here 🧾
This course quotes published work where published work exists, and labels the rest as a teaching model.
| Claim | Status |
|---|---|
| AI adoption raises delivery throughput and instability at the same time | DORA State of DevOps 2025. Epics completed per developer up 66.2 percent, with a negative relationship to delivery stability |
| Elite change failure rate sits near 5 percent | DORA four keys. The 2025 report replaced the four tiers with seven team profiles |
| Give consumers at least 90 days notice of a breaking schema change | Published industry guidance, not a law. Your release cycle sets the real number |
| The model box is the small box in an ML system | Sculley et al., NeurIPS 2015, and it has not stopped being true |
| Every number on the tradeoff board | A teaching model. Shapes of trades are real, digits are calibrated for teaching and are not a benchmark of your platform. Every constant is written down on the session that uses it. |
The knowledge map 🧠
The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.