Decision record: remediation approach for Fabric Lakehouse migration risks
Status: Recommended — roadmap delivered to the client; implementation owned by the client’s data platform team, not yet executed by Chris.
Context. An engineering organization migrating from legacy Synapse to a Medallion Architecture on Microsoft Fabric had real CI/CD discipline in place, but the physical Lakehouse deployment carried risks that would not surface until production scale: 100+ tables in a single shared schema, ingestion built on destructive full-table overwrites burning Capacity Units, and a Gold layer not structured for Power BI’s Direct Lake mode. An architecture health and risk assessment was conducted to determine how to close these gaps before they caused production incidents.
Options considered.
- Continue operating as-is and address Direct Lake readiness reactively only if/when performance or cost became a visible production problem — the do-nothing baseline a remediation roadmap is implicitly justified against.
- Attempt schema isolation, CDC ingestion redesign, and Direct Lake/V-Order optimization all at once rather than in sequence — risks compounding failure modes by building ingestion and consumption-layer changes on top of an still-ungoverned schema.
- A phased, sequenced roadmap — schema-level isolation, then metadata-driven CDC ingestion, then Star Schema/V-Order readiness — the option recommended.
Decision. Recommend the three-phase, sequenced roadmap, with each phase building on the one before it. Delivered as a prioritized plan; execution is owned by the client’s data platform team.
Why. Schema-level isolation addresses the highest-blast-radius risk first — 100+ tables ungoverned in one shared schema — before any ingestion or consumption-layer work is layered on top of it. Sequencing avoids compounding risk that a single concurrent overhaul would carry. Metadata-driven CDC directly targets the identified cost/performance failure mode rather than optimizing around it, and Star Schema/V-Order readiness is a functional prerequisite for Direct Lake, so gating it last matches the actual dependency.
Consequences. The client’s data platform team has a concrete, prioritized sequence to close the identified gaps rather than an open-ended list of concerns. Because this is a delivered roadmap rather than a completed build, the risk reductions below become real only once each phase is executed — not before: a governed schema foundation once Phase 1 lands, ingestion that stops burning Capacity Units once Phase 2 lands, and Direct Lake-ready consumption once Phase 3 lands.