Every extract is a promise nobody wrote down
Extracts are agreements about shape and freshness that live in someone's head. When the upstream team changes a column, the breakage surfaces three layers downstream in a finance report, days later.
A contract moves that promise into the repository: schema, owner, freshness threshold, null policy and known exceptions, versioned next to the pipeline that consumes it.
What a useful contract contains
Four things: the schema with types and allowed nulls, the freshness expectation measured from source event time, the owner who is paged when it breaks, and the replay procedure. Anything less is documentation rather than a contract.
We keep contracts in the same repository as the transformation code so a change to either requires a review of both.
Quality gates before dashboards, not after
The cheapest place to catch a bad batch is the boundary. Volume checks, distribution checks and referential checks at ingestion cost minutes; the same defect found in a board pack costs a week of credibility.
Gates that only warn are decoration. A failing gate should stop the downstream run and page the owner, with a documented override for the cases where a partial load is genuinely better than none.
The metric dictionary is the deliverable
Contract-tested pipelines produce something more valuable than clean tables: a single agreed definition per metric, with an owner and a lineage trail.
Once definitions are explicit, the reconciliation meetings stop. Analysts get their time back, and the argument moves to where it belongs — whether the metric is the right one, not which number is true.
