Check Where a Power Query Transformation Stops Folding to Its Source
A short final table does not reveal how much source data a query retrieved to produce it. Before rearranging transformations for performance, identify which work is being requested from the database and which work remains with Power Query.
Use a read-only practice database connection in Power Query Desktop and a small, authorised sample. Record the connector and source table. CSV and Excel files do not provide the source compute engine used for query folding; this investigation needs a source that supports it.
Inspect evidence for a particular step
Where the connector supports it, right-click an Applied Step and use View Native Query or View data source query. Inspect the request after a simple filter and again after a later transformation. Microsoft cautions that a disabled option indicates stopped folding only for a source that normally exposes this capability; unsupported connectors require other evidence.
For a fictional source, choose ten labelled records, three with Region East. Apply the East filter and inspect both the returned identifiers and the generated request. Record whether that request contains the intended restriction rather than inferring its presence solely from the three-row preview.
Add the transformation you actually need, then inspect the new step. Preserve the earlier evidence and identify the change in the request or capability. Do not assume that every transformation of the same general kind behaves identically across connectors.
Keep the output question unchanged
Power Query Online has separate folding indicators; Microsoft explicitly limits that feature to Online. An indicator describes the query up to that point, and later optimisation can sometimes restore folding. Do not prescribe those icons as a required Desktop control or treat one intermediate state as the final query's permanent state.
If considering a different step order, compare the same expected record identifiers, values and ordering requirements. A faster query that keeps the first three records instead of the three East records answers a different question. Use an agreed source-side diagnostic if the connector cannot show the request directly.
Measure refresh behaviour only after the outputs agree, keeping the source and workload conditions with the observation. Folding may reduce transferred data or local work, but an inspected request is evidence about execution, not a universal speed promise. Keep privacy settings intact while investigating where processing occurs.
Sources: Microsoft: query folding examples; Microsoft: supporting procedure.