circle ci is failing with:

Package pandas conflicts for:
seaborn -> pandas[version='>=0.14.0|>=0.22.0']
seaborn -> statsmodels[version='>=0.5.0'] -> pandas[version='>=0.14|>=0.21']
pandas==0.25.0

but I fail to understand why this version spec cannot be resolved... based on https://docs.conda.io/projects/conda-build/en/latest/resources/package-spec.html#package-match-specifications it seems that pandas[version='>=0.14.0|>=0.22.0'] should be equivalent to pandas[version='>=0.14.0'] and pandas[version='>=0.14|>=0.21'] to pandas[version='>=0.14'] and neither should conflict with pandas==0.25.0. What do I miss?

Also why did this test not fail in the #20143 PR itself?

Also why did this test not fail in the #20143 PR itself?

For the circleci build, sklearn/_min_dependencies.py was not updated so circleci did not use the new minimum pandas version.

For test_min_dependencies_readme.py, there was a small bug in the test.

I suspect the failure has to do with:

numpy==1.14.5
pandas==0.25.0 -> numpy[version='>=1.13.3,<2.0a0|>=1.14.6,<2.0a0']

where numpy>=1.14.6.

Also why did this test not fail in the #20143 PR itself?

For the circleci build, sklearn/_min_dependencies.py was not updated so circleci did not use the new minimum pandas version.

Just to clarify: the pandas version was set in sklearn/_min_dependencies.py to version 0.24.3 working fine.
For some reasons, unclear to me, the README.rst said that the minimal dependency was 0.25 : technically there is no reason to increase pandas min dependency to 0.25, then no reason to increase numpy minimal dependency, you can just fix the readme. This is why I asked in #20143.

pandas 0.25 added support for the sparse accessor, which we use to better support dataframes with sparse columns. I see two paths forward:

1. Use pandas 0.24.3 and only update the README.
2. Use pandas 0.25.0, but numpy would need to be bumped to 1.14.6.

As the difference between numpy 1.14.5 and 1.14.6 is one single bugfix release, I'd go for option 2: pandas 0.25.0 and numpy 1.14.6.

@lorentzenchr @thomasjpfan , Thanks for your answers.

I merged because I was under the impression that everyone agreed this was a good small fix.