Skip to content

Pattern — Power BI & Fabric Roadmap

TL;DR: Move new semantic models to Fabric Direct Lake as the default. Keep existing Import models on Power BI Premium until next major schema change. Avoid DirectQuery for performance-sensitive workloads. Build a TMDL-first development workflow so semantic models are version-controlled like code.

Problem

Power BI is the dominant BI consumer in Microsoft-aligned organizations, but it's evolving quickly:

  • Direct Lake (Fabric) — new storage mode with import-tier performance and DirectQuery-tier freshness
  • TMDL / fabric-cli — semantic models as code, finally
  • Semantic-link / SemPy — Python notebook access to semantic models
  • Copilot in Power BI — natural-language exploration over models

Most organizations have a sprawling Power BI estate built over 5-10 years that doesn't take advantage of any of this. The roadmap to modernize is real and worth investment.

Storage modes — pick the right one

Mode When Pros Cons
Import <100 GB model, batch refresh OK, complex DAX Fastest queries; works without Fabric Refresh latency; memory limits; long refresh windows
DirectQuery Real-time freshness required Always current; no refresh Per-query backend hits; aggregations needed for perf
Direct Lake (Fabric only) Modern lakehouse data, BI-first workloads Import-tier performance + DQ-tier freshness; no refresh Fabric-only; some complex DAX patterns don't work yet
Composite (DQ + Import) Mix of dimension (Import) + fact (DQ) Best of both for some workloads Complex; user-defined aggregations needed

Default by 2026

  • Net-new model: Direct Lake if data is in OneLake / shortcut-able to OneLake
  • Existing Import: stay on Import until next major schema change, then evaluate Direct Lake
  • DirectQuery: avoid except for true real-time requirements

Pattern: TMDL-first development

Old workflow:

  • Open Power BI Desktop, edit semantic model in UI, publish PBIX
  • Source of truth is binary PBIX; no diffs, no PR review

New workflow:

  • Semantic model defined in TMDL (.tmdl files) under git
  • Edited in Power BI Desktop OR Visual Studio Code with TMDL extension
  • Reviewed in PRs like any code
  • Deployed via Fabric REST API or fabric-cli
my-semantic-model/
├── .gitignore
├── definition.pbism                    # bindings, mode
├── model.tmdl                          # model-level config
├── tables/
│   ├── dim_customer.tmdl
│   ├── dim_product.tmdl
│   ├── fact_sales.tmdl
│   └── _measures.tmdl                  # measures separate file
├── relationships.tmdl
├── perspectives/
│   └── exec_view.tmdl
└── translations/
    ├── en-US.tmdl
    └── es-ES.tmdl

Diff a measure change in PR review. Finally.

Pattern: medallion alignment

Lakehouse layer Power BI artifact
Gold tables (star schema) Direct Lake semantic model — 1:1 mapping
Gold aggregates Pre-built aggregation tables in semantic model
Silver Reference for ad-hoc Power BI Datamarts (rare in production)
Bronze Never queried by Power BI directly

Each gold table maps to one fact or dimension. Star schema in gold = fast Direct Lake performance.

Pattern: workspace + capacity governance

Don't put all semantic models in one workspace. Pattern:

Workspaces:
├── Finance / Production (F64 capacity)
├── Finance / Development (F8 capacity)
├── Sales / Production (F32 capacity)
├── Sales / Development (F4 capacity)
├── Shared / Production (F128 capacity, hosts org-wide models)
└── Sandbox / All users (PPU licenses, no capacity)
  • One workspace per LOB per environment
  • Capacity sized for workspace's workload
  • Sandbox for self-service exploration; never source-of-truth
  • Promotion via Fabric Deployment Pipelines (Dev → Test → Prod)

Pattern: Copilot in Power BI

Available for any model on F64+ capacity. Treats your semantic model as a metadata corpus:

  • Users ask questions in natural language → Copilot generates DAX
  • Quality is proportional to model quality — bad column names, missing descriptions, ambiguous measures = bad Copilot
  • Investment in model curation (descriptions, synonyms, hidden columns, sample questions) is now ROI-positive

Roadmap step: pick 5-10 highest-value models, curate them properly, enable Copilot, train users.

Pattern: external tools / DAX optimization

Even with Direct Lake, DAX matters. Tools:

  • DAX Studio — query trace, server timings, query plan
  • Tabular Editor — power editing for Import models, calc groups, perspectives
  • Best Practice Analyzer — automated rule checking on semantic models

Use as part of CI: lint TMDL with BPA before merging PRs.

Migration roadmap (typical enterprise)

gantt
    title Power BI estate modernization (typical enterprise)
    dateFormat YYYY-MM
    section Foundation
    Fabric capacity provisioned     :2026-01, 1M
    OneLake shortcut to ADLS gold   :2026-02, 1M
    TMDL workflow established       :2026-02, 2M

    section Wave 1 - new models on Direct Lake
    Pilot Finance dashboards        :2026-04, 4M
    Pilot Sales dashboards          :2026-06, 4M

    section Wave 2 - Migrate Import models
    Tier-1 (highest-traffic) models :2026-09, 6M
    Tier-2 models                   :2027-03, 6M

    section Wave 3 - Curation + Copilot
    Curate top 10 models for Copilot :2027-01, 6M
    Enable Copilot in Finance        :milestone, 2027-04, 0d
    Enable Copilot org-wide          :milestone, 2027-09, 0d

    section Continuous
    DirectQuery models reviewed      :2027-01, 12M
    Sandbox cleanup quarterly        :2026-04, 24M

Anti-patterns

Anti-pattern What to do
Edit PBIX, publish, no source control TMDL in git
One huge workspace for everything LOB workspaces, dev/test/prod separation
Direct Lake for non-star data Build a star in gold first
DirectQuery to operational DB Build a Direct Lake semantic model on a gold mirror
Copilot enabled on poorly-curated models Curate first; Copilot only as good as the model
Aggregations as an afterthought Plan aggregations during model design
Manual deployment portal-clicking Fabric Deployment Pipelines + git PR reviews
Trying to migrate every Excel report ever Inventory; archive ones not used in 90 days

Trade-offs

✅ Why modernize

  • Direct Lake is genuinely better than Import for large models
  • TMDL + git is finally a real dev experience
  • Copilot value scales with curation investment
  • One platform (Fabric) for storage + semantic + BI = simpler ops

⚠️ Why be patient

  • Curation takes time; rushing produces bad Copilot experiences
  • Some DAX patterns don't yet work in Direct Lake
  • Capacity sizing for Fabric is a learning curve
  • Existing Import models work — don't migrate just for fashion