Microsoft Fabric Solution Architecture in London

The whole platform, not one workload. That means OneLake as the data foundation, the medallion lakehouse (Bronze→Silver→Gold) on Delta Lake, the ingestion method per source (batch, CDC or streaming), Lakehouse versus Warehouse serving, Direct Lake semantic models, the governance plane (OneLake catalog and Microsoft Purview), security from Entra groups down to row- and object-level security, CI/CD across DEV→QA→UAT→PROD, and the AI/agent consumers on top. The job is to defend every one of those choices with a reason and a named trade-off — that is what separates an architecture from a component list.

By workload, not preference. A Lakehouse suits data engineering, Spark, unstructured data and data science; a Warehouse suits SQL-centric analytical serving and dimensional BI with a strong T-SQL surface. A common shape is engineering through the Lakehouse to a curated Gold layer, then a Warehouse or Direct Lake semantic model for serving. We choose per workload, users, performance and governance need, and name the trade-off rather than defaulting to one.

As separated workspaces with their own connections, secrets and configuration, promoted through Git integration and Fabric deployment pipelines. A change flows feature branch → pull request → DEV → QA → UAT → PROD, so it is validated in lower environments and reversible. Building everything in one workspace — where a fix to a report breaks the pipeline feeding it, with no rollback — is the most common reason a Fabric estate cannot be trusted for operations.

In layers, designed in from the start: Microsoft Entra ID groups → workspace roles (Admin, Member, Contributor, Viewer) → OneLake security → row-level and object-level security, with Purview sensitivity labels and least privilege throughout. The mistake is handing out Contributor or Admin because it is quicker and treating RLS/OLS as an afterthought — retrofitting least privilege onto a live estate is far harder than designing it, and until it is done the platform cannot pass an audit or safely feed an agent.

We build Microsoft-first by default — Microsoft Fabric, OneLake, Power BI and Purview — because for most mid-market industrial estates already on Microsoft, it is the lowest-friction, best-governed path. Where a client’s estate already runs on Databricks, we deliver the same reference architecture on Unity Catalog and Delta Lake. The platform follows the estate; the discipline — requirements → design → trade-offs → why — does not change.

By governing what they reason over. A Fabric Data Agent or Copilot is only as good as its grounding, so the architecture provides a governed Gold layer, a semantic model with one definition of each measure, and — where you are heading toward agents — a Fabric IQ ontology of your business entities and relationships. We validate agents against known-correct answers before rollout and enforce that they honour the same security as your reports, so wider natural-language access never means wider data exposure.

It starts with a diagnostic on your requirements, current estate and constraints, then a reference architecture and a phased plan. First value lands in about six weeks — a governed slice end-to-end — rather than an 18-month programme. The full enterprise shape (all domains, governance, CI/CD, AI consumers) is sequenced after that, so you fund a working foundation first and expand from proof, not from a 50-slide roadmap.