Data Mesh in Practice
Decentralising data ownership to the domains that produce it, with a self-serve platform and federated governance — an organisational change more than a technical one.
Definition
Four principles: domain ownership of data, data as a product, a self-serve data platform, and federated computational governance.
The problem it addresses
A central data team owns pipelines for data it does not understand, produced by teams who have no incentive to keep it usable. The central team becomes a bottleneck, the domain teams treat data as exhaust, and quality problems have no owner.
That is a genuine and common failure, and it is fundamentally organisational.
What each principle means in practice
Domain ownership. The team producing the data owns its published form: schema, quality, freshness and documentation. They know the semantics; nobody else does.
Data as a product. A published dataset has consumers, a documented contract, quality guarantees, freshness SLAs and a support path. This is the principle that does the most work — it converts data from a by-product into something with an owner and a standard.
Self-serve platform. Domains cannot each build ingestion, storage, cataloguing and access control. The platform provides that as self-service capability — otherwise decentralisation is just each team solving the same problem badly.
Federated governance. Global standards where interoperability requires them — identifiers, security classification, interchange formats — decided by representatives from the domains, with everything else left to the domains.
Where it fails
Adopted as a technology. There is no data mesh product. Buying a catalogue and declaring a mesh changes nothing.
Decentralisation without the platform. Each domain builds its own pipeline, badly, and the estate is worse than the centralised version it replaced. The platform is a precondition, not a later phase.
Domains without capability or incentive. A team with no data engineering skill and no reward for data quality will not produce a good data product.
Applied at the wrong scale. In a small organisation a central data team is more efficient. The mesh addresses a coordination problem that only exists above a certain size.
The honest assessment
The product thinking and domain ownership principles are valuable at almost any scale and can be adopted independently. The full model is an organisational programme, and adopting the label without the platform and the incentives produces a decentralised mess.
Interview question
"What must exist before decentralising data ownership makes things better rather than worse?"