Chris Webb notes that parallelism does not mean “free performance improvements”:
To finish off my series of posts on concurrent evaluation in Fabric Dataflows Gen2 (see part 1 and part 2) I decided to do some more realistic tests to see how much parallelism I could get. To do this I uploaded 244 identical Excel files containing almost 542000 rows of data each to a SharePoint document library. Excel files are probably the worst-performing file format for dataflows (see here for some tests that show this), while SharePoint is probably the worst-performing place to store data for a dataflow and also has a reputation for throttling applications that make too many requests.
Click through for Chris’s test results.