Temporarily pin fastparquet in CI

[jrbourbeau]

No description provided.

[jrbourbeau]

There are some CI failures which appear to be related to the latest fastparquet release using features from pandas which our minimum pandas version (0.25) doesn't support:

 ___________ ERROR collecting dask/dataframe/io/tests/test_parquet.py ___________
dask/dataframe/io/tests/test_parquet.py:25: in <module>
    import fastparquet
/usr/share/miniconda3/envs/test-environment/lib/python3.7/site-packages/fastparquet/__init__.py:5: in <module>
    from .core import read_thrift
/usr/share/miniconda3/envs/test-environment/lib/python3.7/site-packages/fastparquet/core.py:13: in <module>
    from .converted_types import convert, simple, converts_inplace
/usr/share/miniconda3/envs/test-environment/lib/python3.7/site-packages/fastparquet/converted_types.py:69: in <module>
    np.dtype('bool'): pd.BooleanDtype()
/usr/share/miniconda3/envs/test-environment/lib/python3.7/site-packages/pandas/__init__.py:214: in __getattr__
    raise AttributeError("module 'pandas' has no attribute '{}'".format(name))
E   AttributeError: module 'pandas' has no attribute 'BooleanDtype'

cc @martindurant

[martindurant]

Booleans came in in v1.0.0, the one after the 0.25 line - which is 2 years old (today!). We can either skip this up update our min version, if people want. Yes, I realise that the requirement in fastparquet ought to have been updated :|

[jsignell]

If we are following the loose 18 month policy that was discussed in #7378, then we can update to pandas to >=1.0.0 on July 30th. I'd say to fix this test, fastparquet should fix its requirements.txt.

[martindurant]

I'd say to fix this test, fastparquet should fix its requirements.txt.

That won't affect the released version, of course.

[jsignell]

That won't affect the released version, of course.

You could do a new build on conda-forge.

[martindurant]

Good idea.

[martindurant]

Can try kicking it again whenever we reckon conda-forge posts the new build

[jrbourbeau]

Thanks for the conda-forge update @martindurant. Bumping CI now...

[martindurant]

I don't know if the new build hasn't appeared yet in the channel metadata, or if the resolver will now insist on build 0, since it allows for the desired pandas version.

[jrbourbeau]

I might have bumped too early. I just did conda create -n test fastparquet pandas locally which pulled in the new fast parquet build (conda-forge/osx-64::fastparquet-0.7.0-py39hc89836e_1). Bumping CI again now

[jrbourbeau]

The linux Python 3.8 build is picking up build 1 of fastparquet, but the Python 3.7 build is installing build 0 still. Perhaps we should open a PR over in https://github.com/conda-forge/admin-requests (like, for example, this PR conda-forge/admin-requests#274) which marks build 0 as broken? cc'ing @jakirkham in case there's some other route we should go

[martindurant]

Ww should be able to remove build 0; but it might just get the earlier version. That would be OK too , I suppose.

[jrbourbeau]

Temporarily pinning fastparquet for Python 3.7 to try and unblock CI

[jrbourbeau]

Planning to merge this after CI finishes. I also opened up a PR in conda-forge-repodata-patches-feedstock which should handle patching the pandas dependency for fastparquet=0.7.0 (xref conda-forge/conda-forge-repodata-patches-feedstock#152)