Transpose operator inevitably triggers lazy dtypes metadata computation.

[dchigarev]

Doing df.T triggers dtypes property access and so its computation, even if it has been delayed by setting modin_frame._dtypes = None:

modin/modin/core/dataframe/pandas/dataframe/dataframe.py

Lines 2269 to 2272 in a80cb3b

	new_dtypes = pandas.Series(
	np.full(len(self.index), find_common_type(self.dtypes.values)),
	index=self.index,
	)

Data types computation may take up to several seconds for wide frames, we should avoid it when possible.

BTW df.T is a part of the reduction operations flow (columnarization at Series constructor), so it's generally slowing down every reduction method.

[devin-petersohn]

@dchigarev this should be a couple of lines of code to fix, and should improve performance significantly.