Live connection = a thin report connects to an external semantic model (Power BI dataset or Analysis Services). No local data model, no refresh schedules in the PBIX — queries run against the source model in real time. Data refresh (Import) = data is ingested into a Power BI semantic model and refreshed on a schedule (plus Incremental Refresh/Hybrid Tables for scale). Use Live for single source of truth & central governance; use Import + Incremental for blazing interactive performance; use DirectQuery/Direct Lake/Composite when freshness demands or lakehouse patterns require it.

Live Connection

A report connects live to a published semantic model (Power BI dataset or Azure Analysis Services). The PBIX holds only visuals; measures and relationships live in the central model. No dataset refresh at the report level — the connected model handles processing/freshness.

Import + Data Refresh

Data is imported into your model; queries hit in-memory VertiPaq. Schedule refreshes (full or Incremental Refresh) to keep current. Best for snappy UX and complex DAX with heavy aggregations.

DirectQuery

No import; queries are pushed to the source (SQL, etc.) at interaction time. Freshness is maximal; performance depends on source and network. Consider Aggregations to accelerate.

Direct Lake (Fabric)

Reads parquet data directly from the lakehouse with near in-memory performance and minimal refresh. Great for very large data with freshness needs.

Composite Models

Blend Import + DirectQuery + Direct Lake in a single model, plus Hybrid Tables (hot cache for recent data, historical in import). Best of many worlds when tuned.

A — Aggregations
Accelerate DirectQuery with import aggregation tables; route big queries to cached summaries, fine-grain to source.

B — BI Semantic Layer
Live connection shines when a central, curated model (enterprise measures, RLS) must be reused across many reports.

C — Caching & Query Limits
Import = VertiPaq cache. DirectQuery respects source timeouts; cache visuals judiciously via query reduction settings.

D — Direct Lake
When you live in Fabric, Direct Lake reduces refresh ops and unlocks lake-scale models with interactive performance.

E — Encryption & Security
Use workspace-level permissions, RLS in the model (for Import/Direct), and rely on central model RLS for Live connections.

F — Freshness SLAs
If stakeholders demand sub-minute data, favor Live/DirectQuery/Direct Lake; for hourly/daily, Import + Incremental is ideal.

G — Gateway
On-prem sources need an Enterprise Data Gateway for both refresh and DirectQuery. Plan redundancy and version hygiene.

H — Hybrid Tables
Keep “hot” recent data in DirectQuery and “cold” history in Import within one table; perfect for high-volume fact tables.

I — Incremental Refresh
Partition by date; refresh only the latest slices; combine with Change Detection for near-real-time deltas.

J — Jitter & Latency
Live/DirectQuery depend on network + source health. Co-locate capacities close to data to crush round-trip delays.

K — Key Features
Some AI/quick insights require Import. Live leverages whatever the parent model exposes. Verify feature parity early.

L — Limits
Watch dataset sizes, model memory, DQ timeouts, and concurrency. Capacity matters; plan for peak, not average, load.

M — Modeling Discipline
Import/composite models live in your PBIX; Live pushes you to centralize measures & relationships in the shared model (good!).

N — Near Real-Time
Use DirectQuery/Direct Lake for “now,” or combine Incremental + change detection for “near-now.”

O — Observability
Monitor refresh history, partition ops, DQ query logs, and capacity metrics. Alert on failures and long-running queries.

P — Performance Engineering
Star schema, numeric surrogate keys, summarized facts, measure-driven formatting. In DQ, push filters early; avoid row-by-row UDFs at source.

Q — Query Folding
In refresh paths, ensure M steps fold to the source for speed. In DQ, push down predicates and aggregations where possible.

R — Row-Level Security
Define RLS in the central model for Live; for Import/DQ, maintain roles in the dataset. Test with role view before go-live.

S — Service & Capacity
Pick the right capacity tier. Import loves memory; DQ loves CPU + fast source. Direct Lake loves lake throughput and cache warmup.

T — Time Intelligence
Import makes complex DAX time calcs smooth. DQ can do it, but might translate to heavy SQL; consider pre-aggregations.

U — User Experience
Import = buttery interactivity. DQ = occasional spinner unless optimized. Live mirrors whatever the source model delivers.

V — Versioning
Live centralizes change control (great for governance). Import decentralizes; use deployment pipelines/ALM to stay sane.

W — Writeback & What-If
What-If parameters love Import. For true writeback, integrate external stores/APIs; keep models read-optimized.

X — XMLA & External Tools
Use XMLA endpoints/Tabular Editor for enterprise modeling, partitions, and scripted deployments — critical for both paths.

Y — Your SLA & Costs
Import costs are refresh + memory; DQ costs are source load + capacity. Live costs tie to the parent model’s processing window.

Z — Zero-Downtime
Staged deployments, blue-green refresh, partition swaps, and feature flags keep analytics online during updates.

Enterprise Thin Reports (Live)

1. Publish a certified semantic model with curated measures & RLS.
2. Build reports via Get data → Power BI datasets (no local model).
3. Manage processing centrally (partitions, incremental). Deploy via pipelines.

Use when: many reports must reuse one truth with tight governance.

Import + Incremental + Hybrid

1. Model star schema; define RangeStart/RangeEnd parameters.
2. Enable Incremental Refresh and, if needed, Hybrid Tables.
3. Warm cache post-refresh; validate partition health.

Use when: ultra-fast UX with hourly/daily freshness is OK.

Composite with Aggregations

1. Keep granular fact in DirectQuery.
2. Add Import aggregation tables (by Date/Region/Product).
3. Define aggregation mappings; verify hits in Performance Analyzer.

Use when: need near real-time on detail plus speed on summary.

Direct Lake (Fabric) Lakehouse

1. Land curated parquet in lakehouse delta tables.
2. Build a Direct Lake model; manage dataflows/medallion layers.
3. Cache warm-up for peak reports; monitor throughput.

Use when: huge data + low-latency without heavy refresh cycles.

Can I add local measures in a Live connection?

Traditional Live reports don’t host local models. Prefer thin reports and put measures in the parent dataset, or use composite over Power BI dataset when local calc is a must.

What about time intelligence on DirectQuery?

Works, but can push heavy SQL. Pre-aggregate by date or use aggregations to avoid slowdowns.

When do I choose Direct Lake over Import?

When data volume explodes and refresh windows become painful — Direct Lake gives lake-scale with interactive speed.

How do I keep costs predictable?

Import: manage refresh windows and memory. DQ: protect the source with caching/aggregations and query limits. Live: schedule processing for the central model in off-peak hours.