How would you roll out data contracts across teams whose producers see no benefit?

## The real problem is incentives, not tooling The technical side of data contracts is unremarkable: a spec file, a diff, a gate. Every failed rollout fails socially. The producing team experiences a contract as new review burden protecting somebody else's dashboard, and they were not consulted about the dashboard. Any plan that does not address this is a plan to write specs nobody enforces. ## Start with pain, not coverage Begin with exactly one dataset, chosen for evidence rather than importance: there was an incident, it recurred, and a consuming team can state what it cost — a wrong number that reached a customer, an executive report restated, days of engineering time. That story is the mandate. A rollout that opens with "we should have contracts for all sources" has no story and dies in prioritisation. ## Make day one free Do not ask producers to author a contract from scratch. Generate the first version by introspecting what they already emit — the current schema, observed nullability, observed volume and lag turned into initial SLA clauses — and hand it to them as a pull request. Their job is to review, correct the semantics only they know (units, event versus ingest time, what one row means), and merge. Adoption cost approaches zero and the check is green from the first commit, which matters because a mechanism that fails on introduction gets disabled. ## Put the gate in their house The check runs in the producing repo, in their CI, with their branch protection. This is a deliberate political choice: engineers accept a failing test in their own pipeline and resent a platform team's after-the-fact escalation. Keep the failure message specific — name the field, the classification, and the two legitimate ways forward (deprecate instead of remove, or bump major with sign-off). Run in warn mode until the noise is gone, then make it blocking. ## Give producers something back This is the step most rollouts skip. Contracts must return value to the producing side or they are pure tax: - **Fewer interrupts.** A published contract with semantics answers the analyst questions that currently arrive in their channel. - **A defensible boundary.** "We committed to these fields and this freshness" also means nobody may demand arbitrary new fields at short notice. - **Fewer wrong-team pages.** A documented owner and SLA routes downstream incidents to whoever actually owns them. - **Compile-time safety**, if you generate typed stubs — the producer's own service catches the mistake. ## Measure the thing you claimed to fix Before expanding, instrument the argument. Count incidents caused by unannounced upstream change, time to detect them, and time to restore, per quarter, before and after. Also track process health: how many breaking changes were caught in a PR rather than in production, how long the median deprecation actually took, how many gates were bypassed. The bypass count is the honest indicator — a rising number means the gate has become theatre and the escape hatch needs repricing, not more datasets. ## Expand along dependency weight, not evenly The second and third datasets should be the ones with the most downstream consumers and the most cross-team exposure. Most organisations end with contracts on a modest set of core interfaces and none on the long tail — and that is the correct end state, not a failure. Universal coverage means universal review burden for datasets nobody outside one team reads. ## Where you cannot have a contract at all Be explicit about the limit, because a principal-level answer is judged on knowing it. A contract requires a counterparty who can be gated. For a vendor SaaS, a purchased data feed, or an internal system whose team will not participate, there is no pull request to block. What you have then is a **monitor** at your ingestion boundary: assert the shape and the volume on arrival, alert on deviation, and design downstream models to degrade rather than silently absorb the change. Call it what it is. Presenting a boundary monitor as a contract sets an expectation of prevention that you cannot meet. ## Escalation, only after evidence Organisation-wide policy — contracts required for any dataset crossing a team boundary — is the last step, not the first, and it only survives once several teams can point to a break the gate caught. Leading with policy produces compliance artifacts: specs written to satisfy an audit, never enforced, and quietly false within a quarter.