Back to All Engineering Notes
Fabric & Modernization•8 min read•2026-10-10

Direct Lake Fallback to DirectQuery: How to Detect and Prevent It

Your Direct Lake model is quietly running as DirectQuery. Why it happens, how to see it, and how to stop it — including when you should not.

Core Engineering Takeaways

  • •Only Direct Lake on SQL endpoints can fall back to DirectQuery; Direct Lake on OneLake fails to refresh instead when guardrails are exceeded.
  • •Fallback keeps queries working but slower, which is why it often goes unnoticed for months.
  • •The model's Direct Lake behavior property (Automatic, DirectLakeOnly, DirectQueryOnly) controls whether fallback is allowed.
  • •Most fallback comes from exceeding per-SKU guardrails (rows, files, row groups) or from features the SQL endpoint forces into DirectQuery.

1. Two kinds of Direct Lake, two failure modes

Direct Lake lets a semantic model read Delta tables in OneLake directly instead of importing them. A lightweight refresh — framing — points the model at the latest version of each table, and columns are loaded into memory when queries need them.

There are two flavours, and they fail differently when something goes wrong:

  • Direct Lake on SQL endpoints reads tables through a Lakehouse or Warehouse SQL analytics endpoint. When it cannot serve a query in Direct Lake mode, it can fall back to DirectQuery against the SQL endpoint — if fallback is enabled. Queries still return results, just more slowly.
  • Direct Lake on OneLake reads Delta tables directly. It has no DirectQuery fallback: if guardrails are exceeded, it behaves like Import — the refresh fails and the model cannot be queried until the tables are back within limits.

So the question "why is my Direct Lake model slow?" usually has the answer it is not running in Direct Lake — and the question "why did my Direct Lake model stop working?" usually has the answer it exceeded a guardrail with nowhere to fall back to.

2. What causes fallback

The two most common causes:

Guardrails. Every Fabric SKU caps how big a Direct Lake table can be. For example:

  • F2 to F32: 300 million rows and 1,000 Parquet files (and row groups) per table.
  • F64: 1.5 billion rows and 5,000 files per table.

The files and row-groups limit catches many teams out. A table written by frequent small appends can accumulate thousands of small Parquet files long before it gets near the row limit. The Direct Lake vs Import guide lists the guardrails for every SKU.

Features the SQL endpoint cannot serve in Direct Lake. With Direct Lake on SQL endpoints, some designs force queries to DirectQuery — for example, tables based on SQL views rather than Delta tables, or security defined on the SQL endpoint that the semantic model has to respect. Check the current Microsoft documentation for the full list, because it changes as Direct Lake matures.

Scaling a capacity down is a sneaky third cause: a table comfortably within F64's guardrails can exceed F32's.

3. How to detect fallback

Fallback is invisible to report users — visuals just get slower — so you have to look for it deliberately.

  • Performance Analyzer in Power BI Desktop: when a visual's query falls back, you see a DirectQuery entry for it. A Direct Lake visual should not have one.
  • DAX Studio with Server Timings enabled: fallback shows up as SQL queries sent to the SQL endpoint instead of (or alongside) storage-engine scans.
  • Traces / Profiler: DirectQuery events against the model mean something is not being served from memory.

Test with realistic filters and the largest tables, and re-test after data volumes grow: a model that runs fully in Direct Lake today can start falling back when a table crosses a guardrail next quarter.

Add a quarterly check of row and file counts per table against your SKU's guardrails. Growth is predictable; the fallback it causes does not have to be a surprise.

4. Control it with the Direct Lake behavior setting

The semantic model property Direct Lake behavior decides what happens:

  • Automatic (the default): use Direct Lake, fall back to DirectQuery when needed.
  • DirectLakeOnly: never fall back. Queries that cannot run in Direct Lake fail instead.
  • DirectQueryOnly: always use DirectQuery — useful for testing.

You can set it with Tabular Editor, through TMSL/XMLA (directLakeBehavior on the model), or with Semantic Link in a Fabric notebook.

Setting DirectLakeOnly is a good choice for production models where predictable performance matters more than availability: a failure is loud and gets fixed, while silent fallback can degrade a report for months. It also avoids planning for both execution paths. The trade-off is real, though — if a table crosses a guardrail overnight, reports fail rather than slow down, so pair it with monitoring.

5. Prevent it

  1. Keep Delta tables compacted. Regular OPTIMIZE (and V-Order) keeps the number of Parquet files and row groups well under the per-table guardrail.
  2. Model on Delta tables, not views, for anything that should run in Direct Lake on SQL endpoints.
  3. Size the SKU from the largest table, not the total data volume. Rows per table is the limit that bites.
  4. Archive or aggregate history you do not need at row level, so large fact tables stay inside the guardrails as they grow.
  5. Test after every scale-down of a capacity.

My free Direct Lake checker checks row counts and common fallback triggers against each SKU's guardrails, which is a quick way to see how much headroom a model has.

6. When fallback is acceptable

Fallback is not always a bug. For a rarely used drill-through page over a huge table, an occasional slower DirectQuery answer may be a perfectly good trade for not holding that table in memory. The problem is unintended fallback on the pages people use every day.

Decide deliberately: run hot paths fully in Direct Lake, and if you allow fallback anywhere, know where and why. If your Direct Lake models are slow and you are not sure why, the performance tuning service starts with exactly this diagnosis.

Power BI Diagnostic Audit

How close is your model to a guardrail?

Simulate row counts and fallback triggers against every Fabric SKU, or get a fixed-price audit of the model.

Get the next Engineering Note

Practical DAX, semantic model and Fabric notes. 1–2 a month, unsubscribe any time.

1–2 useful notes a month. Unsubscribe at any time.

Piyush JainPower BI & Fabric Engineer
#Direct Lake#DirectQuery#Fabric#Performance