DecisionARCH-009

Should Northwind move to Fabric, and how?

The Premium capacity is being retired at renewal. Choose a target architecture for the company's models (Import, DirectQuery, Direct Lake, Lakehouse, Warehouse, mirroring) and record the decision as ADR-009.

Reported by: Elena Torres, BI Lead (your manager)
Opened: Mon 6 Jul 2026, 09:00
Needed by: Architecture board, Thu 23 Jul
Priority: High
Time box: 5 h

Labels: roadmapbudgetplatform

Previously at Northwind

The certified model from Sprint 05 now serves 140 people. Refresh takes 74 minutes and is creeping up; the operations dashboard still uses DirectQuery against the warehouse management system, and Sprint 08 showed how fragile the on-premises chain is.

Why the business cares This decision fixes the platform, cost and skills the BI team will live with for three to five years. Getting it wrong means either an expensive migration that changes nothing for users, or a cheap one that hits a wall in a year.

Brief

Elena TorresBI Lead (your manager)email Mon 6 Jul, 09:00

ADR-009: target architecture

Our P1 capacity won't be renewed in December: Microsoft is retiring Premium per-capacity SKUs, and the equivalent is an F64 Fabric capacity. That forces a decision, so let's make a good one.

Please write ADR-009 for the architecture board on the 23rd: options, constraints, trade-offs, your recommendation, consequences, risks and what would make us revisit it. Everything you need is in the folder.

I don't want the most modern option. I want the one this team can run.

Kenji MoriData Engineerchat Mon 6 Jul, 10:15

If we go Fabric I'd love to stop running nightly ETL into NorthwindDW and Power BI separately. Mirroring could replicate NorthwindDW into OneLake continuously. But nobody here knows Spark, and I'm not learning it by December.

Sarah ChenCFOemail Mon 6 Jul, 11:30

Cost

Whatever you propose: no increase in platform spend next year unless there's a business case I can defend. And Finance's access rules must not get weaker.

Jordan ReedVP Saleschat Mon 6 Jul, 12:02

Ops says they need everything real time. I need my sales numbers faster than 'tomorrow' too. Can Fabric do real time for everything?

What you are expected to do

Frame the decision: what is actually forced (and by when), and what is optional?
List the constraints and the real latency needs per workload. Challenge "real time for everything".
Compare at least three options with pros, cons, costs and risks, against the constraints.
Record your decision (Your work → Decision) and write ADR-009.
Explain the recommendation to the CFO in five sentences and to Kenji in technical terms.

Deliverables

ADR-009: context, constraints, options, recommendation, consequences, risks, revisit trigger
Latency and workload table: what each consumer really needs
Phased migration plan with a rollback at each phase
Five-sentence recommendation for the CFO

Evidence

current_estate.csv Today's main models: storage mode, size, refresh, consumers, latency needed.
constraints.csv
refresh_and_load.csv Refresh duration and source load over the last four weeks.
Storage options in Fabric (summary checked 1 Oct 2026)