What three things does a minimal data contract specify?

Storage tier, retention, encryption

A producer wants to drop the `amount_cents` column and add `amount_minor`. Under a healthy contract regime, what happens?

CI flags it as breaking; the producer ships *both* columns for the deprecation window, then drops the old one after every consumer has migrated

They ship and the consumers discover it at 9 a.m. Monday

CI flags it as breaking; the producer ships *both* columns for the deprecation window, then drops the old one after every consumer has migrated

They file a JIRA ticket and wait six months

They email a screenshot of the new schema

Data Contracts — Semantic Web Academy

Overview

Data Contracts

Producer-signed schema + SLA — moves the responsibility for breakage to the producing team.

Why it matters

A contract turns 'oops, we changed the column' from a P2 incident on the consumer team to a blocked CI check on the producer team.

Going deeper

A practical data contract YAML looks like:

dataset: prod.analytics.fct_orders
owner: team-checkout@company
version: 2.3.0
sla:
  freshness: 1h
  availability: 99.9%
schema:
  - { name: order_id,   type: string,    nullable: false, pk: true }
  - { name: amount_cents, type: int64,   nullable: false }
  - { name: placed_at,  type: timestamp, nullable: false }
checks:
  - amount_cents > 0
  - order_id is unique
breaking_changes_require: [migration_window, downstream_signoff]

The crucial line is the last one: breaking changes require explicit downstream sign-off. That single rule is what flips incident ownership from the consumer back to the producer, where it belongs.

Analogy

A data contract is the OpenAPI spec for a dataset.

Backend services stopped quietly renaming fields in REST responses years ago — because OpenAPI / Protobuf made the contract explicit, machine-checked, and breakable-builds. Data pipelines, by contrast, mostly still rename columns at 3 a.m. and break a dozen downstream models at 9 a.m.

A data contract closes that gap: the producer team publishes a versioned schema + freshness SLA + ownership block; CI fails their build when they violate it. The consumer is no longer playing defense.

Make it stick

Use the prompts below to anchor data contracts to something you actually own.

›Which of your team's datasets has the most downstream consumers? It is your highest-ROI candidate for a v1 contract.
›Last time a producer broke a consumer in your org, who paid the cost in hours — and who *should* have, under a contract regime?
›What's the minimum CI check you could ship next sprint to enforce 'no column drop without sign-off'?

Reading in progress · 0 of 2 activities done