Modeling, tests, and docs as code
Analytics engineering is an engineering discipline. Everything versioned, reviewed, tested.
Approach
We operate on the modern stack: dbt for transformations, an orchestrator (Airflow or Dagster), tests in CI, and a metric glossary treated as living documentation. Every pull request ships a preview and an impact plan.
Clear layers: staging isolates sources, intermediate builds reusable blocks, marts ship models ready for BI and reverse-ETL.
Mandatory code review. No model grows without owner and tests.
Docs generated from the code, not written aside: if the model changes, the doc changes.
Tests & docs
- T-01Schema tests (unique, not_null, relationships) on every critical table
- T-02Contract tests on marts: breaking changes fail CI
- T-03Freshness tests per table SLA
- T-04Docs generated via dbt-docs with full lineage
- T-05Exposure tracking: every dashboard linked to its source model
How a mart actually looks
A typical revenue mart with an enforced contract, a 2-day late-arrival window, and an incremental key. Every PR runs the contract test; a breaking change fails CI.
0
{{
config(
materialized = 1 ,
unique_key = 2 ,
on_schema_change = 3 ,
contract = { 4 : true }
)
}}
WITH paid AS (
SELECT
date_trunc( 5 , order_ts) AS period,
customer_id,
SUM(amount_cents) / 100.0 AS revenue_usd
FROM {{ ref( 6 ) }}
WHERE status = 7
{% if is_incremental() %}
AND order_ts >= (SELECT max(period) - interval 8 FROM {{ this }})
{% endif %}
GROUP BY 1, 2
)
SELECT
period,
COUNT(DISTINCT customer_id) AS customers,
SUM(revenue_usd) AS revenue_usd
FROM paid
GROUP BY period;
Metric glossary
Every project starts with an agreed glossary. We keep a public example so you can see the format.
Wire a real pipeline to your warehouse
30-min stack review. We come back with a scoped plan, an owner, and a freshness contract. No SDR, no slide deck.