← CDO Compass
Technology & Innovation

The contract, the worker, and the librarian

How dbt, OpenMetadata, and the ODCS standard divide the work, and why a data contract is only worth as much as the layer that enforces it.

A data contract is a promise written in a file. This table will always have these columns. This identifier will never be empty. This data will refresh every day. The promise is clear, it is versioned, it is reviewed by the team. And yet, on its own, it protects nothing.

A promise on paper stops no one from breaking it tomorrow. You need a mechanism that checks the promise and blocks whatever fails it. Practitioners call this enforcement. Before we get there, it helps to understand who does what in a modern platform, because the most common confusion is the belief that two very popular tools, dbt and OpenMetadata, play the same role. They do not play the same role at all.

Part I · The Tools

dbt and OpenMetadata Do Different Jobs

The simplest way to tell them apart comes down to two images. dbt is the worker. OpenMetadata is the librarian.

It acts

dbt

the worker in the kitchen

It takes the raw data landing in the warehouse and turns it into clean, aggregated tables that are ready to use. dbt actually runs SQL and creates tables. But it only knows its own kitchen, its own project.

It observes

OpenMetadata

the librarian of the house

It builds nothing. It watches everything that exists, the tables from dbt but also the dashboards, pipelines, and models, and keeps a catalog: who owns what, how things connect, who uses them, whether quality holds.

The relationship between the two is concrete. When dbt runs, it leaves behind automatic records, the files manifest.json, catalog.json, and run_results.json, generated on every execution. OpenMetadata picks up these records and folds them into its catalog. From them it derives three things: lineage, meaning how tables depend on one another, documentation kept in sync with the code, and test results.

Neither one, on its own, is a contract. dbt can build a conformant table, but it only knows its own perimeter. OpenMetadata can describe and audit, but it describes what already exists. The contract is the formal agreement between the party producing the data and those consuming it. And it now, at last, has a standard.

Part II · The Contract

The Contract Now Has a Standard: ODCS

For a long time, every team wrote contracts its own way. That is no longer the case. The standards fight is over. Version 3.1.0 of the Open Data Contract Standard (ODCS) shipped in December 2025, under the Linux Foundation's Bitol project, with stricter validation and an official media type. In the same window, the competing specification was deprecated. In 2026, if you write a contract, you write it in ODCS.

Part III · Enforcement

Three Layers, Three Moments to Catch the Problem

Here is the heart of it. Since a contract is only worth as much as its verification, the real question is not whether to verify, but when. There are three possible moments. Think of an airport and its checkpoints: you can screen before boarding, at the gate, or in an overnight audit after everyone has landed.

The Three Enforcement MomentsEarlier is cheaper and more preventive. Later is easier to install but catches problems after the fact.
CODE USERS 1 · BEFORE at the CI gate cheapest, preventive 2 · DURING in the stream real time, complex 3 · AFTER in the catalog scheduled audit
1
Before · in CI
When a developer proposes a change, an automatic check tests whether it breaks the contract. If it does, it is blocked before it reaches production. The problem never touches the real data. Its limit: it only covers what goes through code.
datacontract-cli
2
During · streaming
Useful when data moves in real time, message by message, at the instant it is produced. You catch the bad record right in the pipe. More powerful, but reserved for streaming architectures and harder to set up.
broker-side enforcement
3
After · the catalog
The data already sits in the warehouse. On a schedule, OpenMetadata re-runs the checks and flags what failed. You learn a few hours later, not before, but for batch that is often acceptable, and you gain history and SLA reporting.
OpenMetadata 1.10+

The mental shift to make: these three layers are not competitors, they are three moments. The earlier you catch, the more preventive and cheaper it is, but coverage is limited to what goes through code. The later you catch, the easier it is to install, but you find out after the fact. OpenMetadata shipped data contracts in version 1.9 in August 2025, then full ODCS 3.1 import and export in version 1.10 in October 2025, which is what makes the third layer practical today.

A contract is only enforced to the extent of the layer that checks it. When someone says the contracts are live, the only good question is which of the three layers is actually enforcing them today. If the answer is that the catalog holds the file, the work is not done.

Bernard Millet, DataEdgePro.net

The Recommended Setup for a Warehouse Team

For most teams in a warehouse and batch environment, the solid and realistic stack combines two layers and leaves the third aside. Write the contracts in ODCS. Enforce them at the continuous integration gate, layer one, with an export to your existing quality engine: ODCS can delegate its rules to tools like Soda, Great Expectations, or dbt tests. Then wire in OpenMetadata for scheduled validation and SLA reporting, layer three. Layer two, real time, is only justified if you genuinely have streaming to protect.

The full chain then reads end to end. A contract declared in ODCS, versioned and reviewed in a pull request. A check in CI that blocks any violation. dbt building and testing the models, leaving its automatic records behind. OpenMetadata ingesting those records, connecting the lineage, applying the policies, and monitoring the contracts on a schedule. Each link does exactly one job, and none pretends to do its neighbor's.

Part IV · Adoption

No One Has Five Years of This

One point to keep in mind, especially if you are building a team or reading a job description around this stack. The enforcement side is genuinely young. ODCS 3.1.0 landed in December 2025. OpenMetadata shipped contract enforcement in version 1.9 in August 2025, and full ODCS support only in version 1.10 in October 2025. The capability is measured in months, not years.

So calibrate expectations accordingly. Nobody has more than a year of production experience enforcing ODCS contracts in OpenMetadata, because the capability only arrived in 2025. Screening for that tenure filters out everyone and no one at the same time. What you can look for instead is the durable reasoning underneath the tools. A practitioner who has spent years on schema design, data protection, and the politics of ownership between teams does not meet these tools as strangers. They recognize familiar problems in a cleaner form, which is exactly what turns weeks of ramp-up into days.

What Actually Transfers

The specific YAML dialect and catalog screens are learnable in weeks. These are the durable skills that make someone effective with the tools, and they predate the standard by years.

Schema evolution
Knowing which changes are safe, which break consumers, and how to version a breaking change without a production incident.
Quality gates
Deciding what to test, where to test it, and which failures should block versus warn. The rules matter more than the syntax.
Ownership
Negotiating the boundary between producers and consumers, and holding named owners to it. This is organisational work, not tooling.

A candidate who can explain why the catalog holding a file is not the same as the contract being enforced understands more than one who lists a tool they could not have used for long. Weigh judgment over years served, because on this stack the years simply are not there yet for anyone.

Conclusion

dbt builds and transforms. OpenMetadata catalogs, connects, and monitors. ODCS is the contract that binds the producer to the consumer, and it now has a settled standard behind it. But the standard is not the safeguard. The safeguard is the layer that checks the promise, and there are three places to put it: before, in CI, at the gate, in the stream, or after, in the catalog on a schedule. For a warehouse team, the first and the third together are usually enough.

That is the whole difference between a contract that reassures and a contract that protects. The first lives in a file. The second lives in a layer that, every day, verifies that the promise holds. Choose the layer deliberately, and treat the tooling as what it is: powerful, worth adopting, and young enough that judgment still matters more than tenure.

Sources & references

  1. Bitol Project, Open Data Contract Standard (ODCS) v3.1.0, released December 2025, hosted under the Linux Foundation. Origin: PayPal Data Contract Template. bitol.io
  2. OpenMetadata, Release notes: data contracts introduced in v1.9 (August 2025); SLA objects and full ODCS 3.1 import and export in v1.10 (October 2025). open-metadata.org
  3. dbt Labs, Model contracts, available since dbt Core v1.5. docs.getdbt.com
  4. Data Contract CLI, open-source validation and enforcement of ODCS contracts in CI. datacontract.com