What a data product is

The phrase is used in two related ways. The older, broader one goes back to DJ Patil, who in 2012 described a data product as one that reaches its goal through the use of data.

On that reading a recommendation engine or a route planner is a data product. The narrower use, now common in data teams and in data sales, means the data itself packaged as something others consume: a table, file, feed or API with a scope, an owner and documentation.

The meaning used here

This page uses the narrower meaning, because that is the thing a seller lists and a buyer licenses. The way of working that produces such products inside a company is a separate idea, data as a product, covered below.

Zhamak Dehghani, writing on martinfowler.com in 2020, described a data product as bundling three components: the code that builds and serves it, the data together with its metadata, and the infrastructure it runs on. For a commercial seller the same idea translates into a practical checklist.

• Contents: what records, fields, time span and geography are in, and what is deliberately left out.
• Format and delivery: files, a database share or an API, and how often it refreshes.
• Documentation: a data dictionary, collection method and known limits.
• Quality measures: completeness, error rates and how quickly problems are fixed.
• Owner: a named person or team who answers questions and maintains it.
• Terms: the licence, permitted uses and price structure.

Data as a product is often confused with a data product, but it is a way of working, not a thing to buy. It is one of four data mesh principles Dehghani set out in 2019 and 2020: ownership by business domain, data as a product, a self-serve data platform and federated governance.

It says the team that creates a set of data, say the logistics team, should publish it for others with the care a product team gives software, and treat the people using it as customers. In her 2019 essay, shared data counts as a product when it is:

• Discoverable: listed in a catalogue with its owner and origin.
• Addressable: reachable at a stable, predictable location.
• Trustworthy: the owner commits to quality levels and tests the data as it is made.
• Self-describing: schemas, meaning and samples are written down.
• Interoperable: built to shared standards so it joins with other teams' data.
• Secure: access is controlled for each product.

Sell one, practise the other

You can sell a data product; you can only practise data as a product. The principle was written for data shared inside a company, and it says nothing about whether that data may legally be sold. But a seller already working this way is far quicker to diligence than one whose data lives in undocumented exports.

Illustrative only. An online marketplace for used construction equipment receives a broker's request for its listing prices from a lender that values machinery. The first deal is a single spreadsheet export. Then two more lenders ask for the same thing, each wanting slightly different columns.

1. 1

   Fix the scope

   One standard schema: equipment type, make, model year, hours, region, asking price, sold price where known, listing date.

2. 2

   Set the cadence

   A monthly file, delivered on a known date, covering the trailing month plus revisions.

3. 3

   Write it down

   A two-page dictionary and a note on how sold prices are captured and where they are missing.

4. 4

   Name an owner

   One analyst on the seller's side handles questions and fixes errors within an agreed time.

5. 5

   Standard terms

   A licence template with tiers for internal valuation use and for wider distribution.

Each new buyer now costs the seller little extra work, which is the commercial point of a data product: build once, license many times.

When a seller says it has a data product, test the claim. Ask who owns it day to day, when it last changed shape, and what happens if a monthly delivery is late.

Products without owners

A product without an owner tends to decay into a raw dataset within a year or two. Also check that every buyer's licence matches the same product definition; selling slightly different custom cuts to each buyer erodes the efficiency that made packaging worthwhile.