Our build based on conda and MKL updated the NumPy version from 1.20.3 to 1.21.2.
The test test_check_pandas_sparse_valid[longfloat-longdouble-floating] start to fail with the following:

E               ValueError: Input contains NaN, infinity or a value too large for dtype('float128').

xref: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=33572&view=logs&j=ee99577d-f9c7-545c-0f56-ef7efc8f5936&t=268f7a52-f266-5aea-5799-fda7f893c9ab&l=913