Back to course overview
Module 6Capstone 15 min

Capstone: Validate & document

Prove the foundation with the ten-questions test, then write the two documents that make it maintainable by someone who isn't you.

A warehouse is validated by one standard: can it answer real business questions correctly, with evidence? The ten-questions test: answer each through the proper interface (certified metrics via the layer where one exists; mart SQL where not), record the number and the definition it relied on.

  1. Net revenue for May 2026, and net of returns. (Two numbers; the gap is the returns program's cost.)
  2. Return rate overall, and by product category. (Which category should merchandising investigate?)
  3. True customer count, and the double-counting error the raw systems would have reported. (Your Module 4 victory, quantified.)
  4. Repeat-purchase rate — and the same number computed without resolution, on customer_ref. (The delta is the business case for entity resolution; put it in a sentence an executive would feel.)
  5. Top 5 customers by lifetime completed revenue, golden records only.
  6. Web vs. POS: revenue share and average order value.
  7. Weekday vs. weekend daily average revenue. (Bonus: does this justify the date dimension you correctly didn't build?)
  8. Null-rate trend for POS email over the 30 days — is intake improving?
  9. Which day had the volume anomaly, and what did the gate say about it at the time? (Your check_runs log is the evidence trail.)
  10. For net_revenue_after_returns: who owns it, is it certified, and what edge-case ruling does its description record? (If you can't answer from metrics.yaml alone, the definition isn't finished.)

The two documents

  • LINEAGE.md — for each mart table: grain sentence, sources, transform path (raw file → staging view → load pattern), and downstream consumers. One page. This is the map the next engineer — or the AI agent with a documentation tool — reads first.
  • QUALITY.md — the contract: every check with family, threshold, and the profiling evidence behind the threshold; the freshness/volume SLAs; what happens on failure (block? quarantine? page whom?). SLAs you can't currently meet get written down as gaps, not omitted — an honest quality doc beats an aspirational one.
Documentation is for machines now too

Everything in LINEAGE.md and QUALITY.md is text an LLM can consume. Well-documented warehouses aren't just kinder to colleagues — they measurably improve every AI agent pointed at them. You're writing the system prompt for tools that don't exist yet.