Direct Lake vs Import: an engineering decision guide
Direct Lake lets a semantic model read Delta tables in OneLake directly, without copying them in a scheduled import. It can remove long refreshes — but it comes with per-SKU guardrails and different failure modes, so it is not an automatic upgrade from Import.
Side by side
| SKU | Rows per table | Parquet files per table | Max model size on OneLake | Max memory |
|---|---|---|---|---|
| F2 | 300 million | 1,000 | 10 GB | 3 GB |
| F16 | 300 million | 1,000 | 20 GB | 5 GB |
| F32 | 300 million | 1,000 | 40 GB | 10 GB |
| F64 | 1.5 billion | 5,000 | Unlimited | 25 GB |
| F128 | 3 billion | 5,000 | Unlimited | 50 GB |
| F256 | 6 billion | 5,000 | Unlimited | 100 GB |
| F512 | 12 billion | 10,000 | Unlimited | 200 GB |
How each mode works
Import copies data into the model's in-memory VertiPaq store during a refresh. Queries are fast and predictable, but the refresh takes time and memory — a full refresh needs roughly twice the model's size while the old copy stays online.
Direct Lake keeps the data in Delta tables in OneLake. A lightweight refresh (framing) points the model at the latest table versions, and columns are paged into memory when queries need them. There is no long import, but query performance depends on how well the Delta tables are laid out.
What happens at the limits
If a table exceeds the guardrails, behaviour depends on the storage mode. Direct Lake on SQL endpoints falls back to DirectQuery if fallback is enabled: queries still work, but more slowly. Direct Lake on OneLake behaves like Import: the refresh fails and the model cannot be queried until the Delta tables are optimised back within limits.
Many small Parquet files are a common hidden cause: the file and row-group guardrail is per table, so un-compacted tables can hit it long before the row limit. Regular OPTIMIZE / V-Order maintenance matters.
How to choose
Choose Import when data volumes fit comfortably within the per-model memory limit, refresh windows are acceptable, and you rely on modelling features that Direct Lake restricts.
Choose Direct Lake when data already lands in a Lakehouse or Warehouse as Delta, refresh times are the bottleneck, and your tables are maintained (compacted, sensibly partitioned) within the SKU's guardrails.
The short answer
- Existing Import models that refresh fine: leave them.
- Large fact tables already in OneLake with painful refreshes: Direct Lake is worth a proof of concept.
- Check rows and files per table against your SKU's guardrails before migrating.
Frequently asked questions
- Is Direct Lake always faster than Import?
- No. Import queries are typically as fast or faster once data is loaded. Direct Lake's advantage is avoiding long import refreshes, and its query speed depends on Delta table layout.
- What is Direct Lake fallback?
- For Direct Lake on SQL endpoints, when a table exceeds guardrails or uses unsupported features, queries can fall back to DirectQuery against the SQL endpoint if fallback is enabled — results still return, but more slowly.
- Can a model mix Direct Lake and Import tables?
- Microsoft supports composite designs in some configurations; check the current Direct Lake documentation for the storage modes your tables use before designing around it.
More comparisons
Sources: Direct Lake overview (Microsoft Learn). Figures verified 2026-10-10; prices are list prices and vary by region and agreement.