Data products
A data product is a dataset treated as a product: it has an owner, a contract, and typed ports, so a consumer knows the shape and the guarantees, not just the location of some files. It is the unit that gets cataloged, versioned, shared, and (when you choose) listed on a marketplace.
What makes a dataset a product
| Facet | What it means |
|---|---|
| Typed ports | Input and output contracts declare the shape a product accepts and the shape it emits, validated against the open Data Product Descriptor Specification. |
| Certification | Quality, freshness, and compliance are recorded as a certified state, not a badge someone drew: "certified" has a record behind it. |
| Provenance | A provenance envelope travels on the product, listing provenance appears on any marketplace entry, and a provenance consumer on the read side lets a downstream user see what a product was derived from. |
| Subscriptions | Consumers subscribe with a lifecycle (and cancellation), gated by consumer access checks and federation access checks for consumption across an estate boundary. |
| Clean-room origin | A joint result computed under mutual distrust in a clean room can be promoted into a governed data product: the commercially interesting end of a clean-room engagement. |
The shape of a product
Figure: a curated model becomes a product carrying typed ports, a recorded certification state, and a provenance envelope; it is cataloged and versioned, then consumed by subscription or published as a signed cross-org listing. Ceiling: certification and provenance are recorded states; SLA definitions exist but SLA enforcement is not shown because it is not a claim (see below).
Table equivalent
| From | Relationship | To | State (meaning) |
|---|---|---|---|
| Curated model / table | is promoted into | Typed ports | verified (green) |
| Typed ports · Certification · Provenance | are recorded on | Cataloged + versioned product | verified → neutral |
| Cataloged product | consumed via | Consumer subscription (access + federation checks) | candidate (cyan) |
| Cataloged product | published as | Cross-org listing (Data Exchange) | candidate, dotted |
Cataloged, versioned, shared
- Cataloged. A product is a first-class catalog object with an owner and a domain. See the catalog and lineage.
- Versioned. Ports and definitions evolve through the catalog's promotion flow, so a change is a recorded event.
- Shared inside the estate via workspace shares and open protocols, or across organizations as a signed listing on the Data Exchange, where the data never crosses, only the signed attestation does.
A product carries SLA definitions and the platform records freshness, availability, and quality monitoring and scoring. Do not read that as SLA enforcement: the end-to-end enforcement path is not something to claim. Say the definitions and the recorded scores exist, not that a breach is automatically acted on. The same discipline applies to lineage: capture is real, but do not assume classification tags propagate automatically along it. See the catalog and lineage.
Across the deployment matrix
- Single-cell: a product, its descriptor, and its subscriptions live on the cell.
- Multi-cell: federation access checks govern consumption across cells; a product's catalog entry is visible through the replicated, federated catalog.
- Sovereign / air-gapped: defining, certifying, and consuming a product needs no external registry; provenance verifies from the envelope.
- Cloud-enabled: the same product model, with federated discovery as a convenience.
Related
- Data Modeling, where a product's underlying model is built.
- The catalog and lineage: how a product is recorded, versioned, and made discoverable.
- The marketplace journey, the producer and consumer journey around a product.
- Data Exchange, listing a product across an organization boundary.
- Open sharing, sharing a product inside an estate.