The pipeline is green, the dashboard is available, and finance still maintains a separate spreadsheet. That is often a rational response. A successful load says little about whether returns were included, a store is missing, or yesterday’s figures were restated.
A trusted data product makes its limitations and responsibilities explicit. Zhamak Dehghani’s writing on data mesh places product thinking and domain ownership at the center of this problem. You can apply those disciplines within a shared warehouse; they do not require a wholesale reorganization.
Who can make a decision about this data?
Name a business owner who can settle competing definitions and a technical owner who can change the implementation. Both need time and funding for the work. An owner listed in a catalog is of limited use if every correction has to wait for the next transformation program.
What does the number include?
Document the grain, business meaning, exclusions, and time basis. Revenue by order date differs from revenue by settlement date; both may be valid. Make those distinctions visible where people consume the data, and preserve the ability to trace a reported figure to its source.
How fresh and complete does it need to be?
Set expectations around the decision being supported. An overnight refresh may be adequate for planning and unacceptable for dispatch. Publish the last successful update, expected arrival time, and known gaps. If a source is late, show that condition rather than presenting yesterday’s value as current.
Who should be allowed to use it?
Build classification, access approval, and retention into the normal delivery path. Sensitive fields should not travel into every downstream copy by default. Test access rules with realistic roles, including users who should be denied, and make revocation part of the same process as granting access.
What happens when a producer changes?
Agree a contract for schema and meaning, with compatibility checks and notice for breaking changes. A field can keep the same name and type while its meaning changes. Consumer tests should therefore check business expectations as well as structure. High-value reconciliations belong in automated checks, with tolerances agreed by the owner.
Who responds when the data is wrong?
Provide a support route, an incident owner, and a way to notify affected consumers. The response should include whether previous outputs need correcting. Fixing tonight’s pipeline does not repair a decision already made from an incorrect report.
Start with one product that supports a consequential decision and review it against these six agreements. Track reconciliation effort, repeat incidents, and the time needed to answer a new question. Those measures show whether the platform is becoming dependable in everyday use.
THE PRACTICAL TAKEAWAYTreat a data product as a service with consumers, commitments, and a correction process. The storage technology is only part of that service.