Why is not using the (optional) Dask Client in combination with pandas.diff not computing the right statistics?

Question

I\'m having some trouble understanding why when I am not using the Dask Client, the output of a certain computation changes.

I have a sizeable dataset (some 10s of mi