ADR-001: Fabric-first development approach¶
Date: 2026-05-27 Status: Accepted Deciders: Sandeep Jayaprakash
Context¶
Microsoft Fabric trial is active with a finite window. Fabric provides the full enterprise data platform story (Delta Lake, Spark, Auto Loader, CDC, Power BI) in a single environment. The alternative was building locally first using Polars/DuckDB and migrating to Fabric later, but this risks trial expiring before Fabric screenshots are captured.
Decision¶
Build the medallion pipeline in Fabric first (Bronze + Silver + Gold notebooks). Implement the local Polars lite tier in parallel as bandwidth allows, but treat it as non-blocking. Capture Fabric screenshots throughout — evidence of a running system on the enterprise platform is the priority during the trial window.
Alternatives considered¶
| Option | Pros | Cons | Why rejected |
|---|---|---|---|
| Local Polars first | No trial pressure, faster iteration | No Fabric screenshots if trial expires | Trial window is finite |
| Both simultaneously | Full coverage | Context switching overhead | Trial makes Fabric the forcing function |
| Fabric only | Maximum focus | Platform lock-in for reviewer demo | Local lite needed for post-trial demo |
Consequences¶
Positive: - Fabric screenshots and evidence captured before trial expires - Enterprise platform patterns demonstrated end-to-end - Auto Loader streaming simulation runs on real Spark infrastructure
Negative: - Local lite tier is secondary priority — built after Fabric phase - Any Fabric-specific bugs discovered late in trial window
Neutral: - Platform abstraction layer (ADR-002) makes this reversible
Implementation notes¶
- LAKEHOUSE_PLATFORM=fabric for all Fabric notebook work
- Screenshots checklist: docs/roadmap/scribe-iq-lakehouse-spec.md section 15.2
- Local fallback: LAKEHOUSE_PLATFORM=local_lite after trial expires