Data Contract

data-contract

Treats a dataset, table, or topic as a product with an owner and a guarantee — the agreement a producing team makes to everyone reading their data. Distinct from api-contract-draft (synchronous request/response + webhook interfaces): this governs data at rest and in motion — the warehouse table, the Kafka topic, the daily export — where the failure mode isn't a 500, it's a downstream model silently trained on a column whose meaning changed last Tuesday.

How to respond

1. Establish producer, consumers, and the asset. What dataset (table / topic / file feed), who produces it, and who consumes it — because the consumer list is what makes a change "breaking". Ask at most 2 questions, spent on the asset's update mechanism (batch / streaming / CDC) and whether any consumer is outside the producing team's control (another team, a partner, a trained model). A dataset with unknown consumers can't have a contract — surfacing that is the first value.

2. Specify schema with semantics, not just types. Types prevent crashes; semantics prevent silent wrongness. Every field gets: name, type, nullability, unit or allowed-values, and a one-line meaning. ✅ "amount — int64, not null, minor currency units (cents), the net charged after discounts" — ❌ "amount — integer" (cents or dollars? gross or net? the model that assumed dollars is now off by 100×). Mark every field's [required] vs [optional] and whether null is a valid value distinct from absent.

3. Set quality guarantees as measurable SLAs (catalog in reference.md) — each with a threshold, a check, and a breach action:

   • Freshness — max acceptable lag ("partition for day D lands by 06:00 D+1")
   • Completeness — expected volume / no gaps ("row count within ±10% of trailing-7-day median; no missing partitions")
   • Validity — % rows passing field constraints ("≥99.9% of email rows match the format")
   • Uniqueness — the key that's actually unique ("event_id unique; duplicates are a contract breach, not a dedup hint")
   • Distribution — drift bounds on the columns consumers depend on ("country cardinality and null-rate within historical band") A quality SLA without a breach action ("page the producer" / "quarantine the partition" / "alert consumers") is a wish; name the action.

4. Write the schema-evolution policy — the section that earns the contract. State precisely what's safe and what isn't, for data consumers specifically: adding an optional column is usually safe; adding a not null column without a default breaks every writer; changing a field's meaning while keeping its name and type is the silent killer and is always breaking; removing or renaming is breaking; narrowing an enum is breaking, widening may be. Every breaking change requires a version bump + a deprecation window + consumer notification — and the contract names the window (e.g. "30 days, or until all registered consumers ack").

5. Make ownership and classification first-class. Named owner (team/role, not a person), an on-call or contact, PII/sensitivity classification per field, retention period, and residency if constrained — the same discipline as nfr-spec, scoped to this asset. An unowned dataset is an incident waiting for a quarter-end.

6. Specify delivery. Where it lives, format (Parquet/Avro/JSON), partitioning scheme, update cadence, and how a consumer registers as a dependent (so the breaking-change notification has an address list). For streams: delivery semantics (at-least-once default — consumers dedupe on the unique key), ordering guarantees, and retention.

7. Emit with templates/data-contract.md in one message: header (asset, owner, consumers), schema table with semantics, quality SLA table with breach actions, evolution policy, delivery spec, classification. Where the asset already exists, note which guarantees are aspirational vs currently met — an honest contract beats a flattering one.

Useful references in this skill

• reference.md — the data-quality dimension catalog, the schema-evolution safe/breaking matrix, and field-semantics rules
• templates/data-contract.md — the contract skeleton with schema and SLA tables

Quality bar

• Every field has a meaning and a unit/allowed-values, not just a type. The cents-vs-dollars class of bug is designed out at contract time.
• Every quality SLA has a threshold, a check, and a breach action. "High quality" is not an SLA; "freshness < 6h, checked hourly, breach pages the owner" is.
• The evolution policy names the silent killer — meaning-change-under-stable-name — as breaking. Most contracts miss it and most data incidents are it.
• Consumers are enumerated, so "breaking" is defined and notifications have an address. Unknown consumers are flagged, not assumed away.
• The asset has a named owner and per-field classification. No owner, no contract.
• Existing-asset contracts mark aspirational vs met. Claiming a freshness SLA the pipeline doesn't hit is how the contract loses its authority on day one.

When to use this skill

• ✅ Publishing a table/topic/feed other teams or models will depend on
• ✅ "We need a data contract for X" / stopping downstream breakage from schema drift
• ✅ Formalizing the guarantee on a dataset that's already widely consumed
• ✅ A data-mesh / data-as-a-product initiative needing per-asset contracts

When NOT to use this skill

• ❌ A synchronous REST/RPC API or webhook — that's api-contract-draft
• ❌ The physical schema migration (expand/backfill/contract) — that's migration-plan
• ❌ System-wide NFRs for a service — that's nfr-spec; this is one data asset's guarantee
• ❌ A private dataset with exactly one consumer inside the same team — a comment in the code may suffice

Anti-patterns to avoid

• ❌ Types without semantics. status: string with no allowed-values list means every consumer invents their own set and one of them is wrong.
• ❌ Quality as an adjective. "Clean, reliable data" — unmeasurable, unenforceable, unfalsifiable. Dimensions with thresholds or nothing.
• ❌ Ignoring meaning-change. Renaming the concept behind revenue from gross to net while the column name and type hold — the most expensive data bug there is, and zero schema validators catch it. The evolution policy must.
• ❌ The orphan dataset. No owner, no consumer list — so no one can be paged and no one can be warned. That's not a product, it's a liability.
• ❌ Aspirational SLAs on a pipeline that misses them. A freshness promise the batch job blows nightly trains consumers to ignore the whole contract.
• ❌ Treating additive as automatically safe. A new not null column with no default breaks every existing writer; a new column can break a SELECT * consumer's schema check. Additive ≠ free in data.