Definition of the data contract
“A data contract defines the agreement between a data producer and consumers.” The sentence comes from the presentation of the Open Data Contract Standard (ODCS), an open standard of the Bitol project, which its website attaches to the LF AI & Data foundation of the Linux Foundation.
In practice, the data contract takes the form of a versioned file, read by people and verified by tools. The producer of a dataset declares in it what it provides: structure, meaning of fields, quality rules, service levels, contacts. The ODCS uses the YAML format for this.
What a data contract contains, and what it changes
The ODCS, at version 3.2.0 as of 12 November 2026, divides the contract into sections. Here are some of them.
- Fundamentals: identifier, version and status of the contract, domain.
- Schema.
- Data quality. A rule is expressed there in text, by a predefined metric such as the number of rows or null values, by an SQL query, or in the format of a third-party tool.
- Service-level agreement: for example latency, update frequency, retention period.
- Team, roles, support and its communication channels.
- Pricing, infrastructure and servers.
Another open format, the Data Contract Specification (datacontract-specification.com), declares itself deprecated since version 3.1.0 of the ODCS, to which its authors recommend migrating.
The contract makes a breaking change visible at the producer, before it reaches consumers. The two moments when it is verified are described in our comparison of data quality testing methods. It has its place between the domains of a data mesh and, for each source, at ingestion into the data platform.
Two subjects remain beyond its reach. The contract does not say who actually consumes the data: that inventory relies on data lineage. And a deviation on a point it says nothing about escapes it, which data observability can flag.
A concrete example: splitting a field under contract
Consider an online news publisher whose subscriptions team publishes the table of active subscriptions every morning, read by the advertising sales house and by the audience research team. The contract, stored with the pipeline code, declares the table key, the definition of the “end date” field, the time of availability and the address of the team to contact.
One day the team wants to split the “offer” field in two. The contract check, run in its integration pipeline, rejects the change. The team then publishes a second version of the contract, announces the withdrawal date of the first and maintains both tables while consumers adapt their processing.
Data contract, schema, service-level agreement, data product: not to be confused
| Concept | What it sets | Relationship to the data contract |
|---|---|---|
| Schema | The columns, their types and their constraints | A section of the contract, silent on the meaning of fields, freshness and owner |
| Service-level agreement (SLA) | Measurable commitments: lead time, frequency, availability | A section of the contract, applied to a dataset |
| Data product | A dataset published as a product, with its owner | The product is what is delivered; the contract, the verifiable description of what is promised |
| Contract in the legal sense | A mutual agreement between two parties | According to the Data Contract Specification, the data contract is a specification maintained by the data provider, and the word “contract” can mislead |
The quality rules a contract records are written with the business teams, threshold and owner included. This drafting work is described in our data quality offering.