RLS: pandas 1.5.1

[datapythonista]

It's been more than two weeks since the the release of pandas 1.5.0. There are several fixes to it since then, probably worth releasing a 1.5.1.

I see we've got some regressions for 1.5.1:
https://github.com/pandas-dev/pandas/issues?q=is%3Aopen+is%3Aissue+label%3ARegression+milestone%3A1.5.1

Is there any that should be fixed before a 1.5.1?

[phofl]

We should pin/fix the pydata sphinx theme issue and #48987 is ready

[MarcoGorelli]

shall we just revert #46174 to solve #48826 ?

[mroeschke]

We can discuss more during the dev call next week, but IMO the existing regression don't appear to be absolute blockers for a 1.5.1 release

[datapythonista]

Waiting for #49032 and #49137 and will release pandas 1.5.1. I moved all other PRs in the milestone to 1.5.2.

[MarcoGorelli]

Looks like #49032 isn't quite ready - shall we just revert the original PR, and then, get the (fixed) performance improvement in for 2.0.0?

Have opened #49140 if we just want to revert

[phofl]

+1 on reverting for now

[GYHHAHA]

Sounds good. Just revert that now and consider moving PERF to 2.0.0.

[mroeschke]

Thanks @MarcoGorelli for the revert in #48826; just merged it so we should be good to go once the backport PR is merged.

[datapythonista]

I'll start now the tests for the release and the release afterwards if everything looks good. I'll keep you updated here.

[datapythonista]

Our step to build the sdist is failing because of a problem in conda. Looks like the blas package downloaded from Anaconda is corrupt.

This is the package in repodata.json:

    "blas-1.0-openblas.tar.bz2": {
      "build": "openblas",
      "build_number": 7,
      "depends": [],
      "md5": "671f6e045d03a842f379145e7fc43c0d",
      "name": "blas",
      "sha256": "ed246aa0bf2809030f911cefd5b16a2c4de0d1b21f75f76e5178122d2213727e",
      "size": 49374,                          
      "subdir": "linux-64",           
      "timestamp": 1528224118045,             
      "track_features": "nomkl",
      "version": "1.0"                        
    },

But the downloaded file is significantly smaller, and sha256 checksum doesn't match: https://repo.anaconda.com/pkgs/main/linux-64/blas-1.0-openblas.tar.bz2

Seems like a global problem, and not in the instance of the CDN accessed by my network, since I tried with a vpn and the problem persists.

I'm able to create the sdist without problems without the docker image used in our release instructions. An environment with just numpy and cython from conda-forge, using the same blas=1.0=openblas package but from conda-forge seems to be working fine, and I guess the resulting sdist should be the same.

Does anyone know if there is a reason to use the pandas-release docker with all its dependencies from the Anaconda channel, instead of using my local environment with those 3 dependencies?

[phofl]

It also looks quite old, the newest version on conda-forge is 2.116

[datapythonista]

It also looks quite old, the newest version on conda-forge is 2.116

I don't think for building the sdist it even makes much of a difference, feels like only setup.py is executed for it, and numpy is only used to call numpy.get_include() which contains the C headers of numpy. So, as far as numpy can be imported and is updated, I don't think any of its dependencies really matter.

But surely no reason that I know to use that old blas and use the Anaconda channel. If there are no more comments I'll move forward building the dist with a fresh local environment, and not from the pandas-build docker image in the pandas-release repo.

[datapythonista]

Pushing tag and starting wheel builds.