## The Problem With Centralised Data Lakes for ML Teams
The centralised data lake model — a single platform team owning ingestion, transformation and serving of data for every consumer across the organisation — scales poorly once an enterprise has more than a handful of ML teams with genuinely different domains. A UK bank with separate ML teams for fraud detection, credit risk, marketing personalisation and regulatory reporting inevitably finds the central platform team becomes a bottleneck: every new feature request, schema change or data quality fix routes through a single overloaded team with no deep domain knowledge of fraud patterns or credit risk modelling, producing slow turnaround and data that technically exists but is poorly documented, inconsistently defined, and mistrusted by the ML teams who actually need to use it. Data mesh is an organisational and architectural response to exactly this failure mode.
## Core Principles: Domain Ownership and Data as a Product
Data mesh rests on four principles. Domain-oriented ownership means the fraud team, not a central platform team, owns and is accountable for the fraud-relevant data products they produce — they understand the domain, so they're best placed to define, document and maintain data quality for it. Data as a product means each domain publishes its data with the same rigour a software team applies to a public API: documented schema, defined SLAs for freshness and quality, a named owner, and discoverability through a shared catalogue, rather than a loosely-documented table that only the original author fully understands. Self-serve data infrastructure means a shared platform team provides the underlying tooling (storage, compute, cataloguing, access control) as a reusable platform, so domain teams can publish and consume data products without needing to build infrastructure from scratch themselves. Federated computational governance means cross-cutting policies — UK GDPR compliance, data classification, access control standards — are enforced consistently across all domains through automated tooling rather than a central team manually reviewing every dataset.
## Data Mesh vs Centralised Lake: A Direct Comparison for ML Teams
The practical difference for ML teams is where accountability and domain expertise sit. In a centralised lake, an ML engineer building a credit risk model requests data from the platform team, waits for it to be prepared, and often discovers quality issues only after model training reveals unexpected results — because nobody on the platform team deeply understood what 'correct' looked like for credit risk data. In a data mesh, the credit risk domain team publishes its own data product with documented quality guarantees, and the ML engineer building on it can trust the SLA because the people who understand the domain are the ones accountable for it. This doesn't mean data mesh is unconditionally better — for smaller UK organisations with only one or two ML teams, the organisational overhead of standing up domain ownership, product thinking and federated governance tooling often exceeds the benefit, and a well-run centralised lake remains the pragmatic choice until domain count and complexity genuinely justify the shift.
## UK Financial Services Examples
A UK retail bank restructuring its data platform around mesh principles assigned its fraud detection, credit decisioning, and customer analytics teams each ownership of their respective data products, with a central platform team providing shared cataloguing (via a tool like DataHub or Collibra) and automated UK GDPR classification tagging applied uniformly across all domains. The bank reported that time-to-availability for new ML features dropped from an average of 6-8 weeks (waiting on the central platform team's backlog) to 1-2 weeks once domain teams could publish their own governed data products directly. For UK insurers, a similar pattern applied to claims, underwriting and pricing domains allows each team's actuarial and data science staff — who understand their domain's regulatory nuances under Solvency UK and FCA conduct rules far better than a generalist platform team would — to own data quality directly, while a shared governance layer still enforces consistent data retention and access-control policy across all three domains.
## Getting Started With Data Mesh Without Overengineering It
Codeworth's recommended entry point for UK enterprises is not a big-bang migration but a phased pilot: pick the two domains experiencing the most acute platform-team bottleneck, define data product contracts for their highest-value datasets, and stand up shared cataloguing and access-control tooling incrementally rather than building a fully federated governance platform up front. Codeworth's MLOps engagements applying data mesh principles for UK enterprise clients typically run this pilot over 8-12 weeks, deliberately limited in scope, before recommending — or explicitly recommending against — a wider rollout, because the organisational change management required for domain teams to genuinely own data-as-a-product is usually the harder problem to solve than any of the underlying technology choices.
Related Codeworth Services
Need Help with Implementation?
The Codeworth team is ready to implement these ideas in your project. Free consultation.
Discuss the ProjectCodeworth team specialist.