Bump minimum numpy version to 1.20

[headtr1ck]

• Closes xarray 2022.6.0 doesn't work well with numpy 1.20 #6818
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

Alternative to #6821

[dcherian]

Hmm... doesn't seem like we can do this yet:

Package           Required             Policy               Status
----------------- -------------------- -------------------- ------
python            3.8     (2019-10-14) 3.8     (2019-10-14) =
boto3             1.13    (2020-04-30) 1.18    (2021-07-15) <
bottleneck        1.3     (2019-11-13) 1.3     (2019-11-13) =
cartopy           0.19    (2021-04-23) 0.19    (2021-04-23) =
cdms2             3.1     (2018-11-02) 3.1     (2018-11-02) =
cfgrib            0.9     (2019-02-25) 0.9     (2019-02-25) =
cftime            1.4     (2021-02-01) 1.5     (2021-05-20) <
dask-core         2021.4  (2021-04-02) 2021.7  (2021-07-09) <
distributed       2021.4  (2021-04-02) 2021.7  (2021-07-10) <
flox              0.5     (2022-05-03) 0.5     (2022-05-03) =
h5netcdf          0.11    (2021-04-20) 0.11    (2021-04-20) =
h5py              3.1     (2020-11-07) 3.3     (2021-06-22) <
hdf5              1.10    (2017-09-18) 1.12    (2020-08-03) <
iris              2.4     (2020-02-21) 3.0     (2021-01-25) <
lxml              4.6     (2020-10-17) 4.6     (2020-10-17) =
matplotlib-base   3.4     (2021-03-31) 3.4     (2021-03-31) =
nc-time-axis      1.2     (2019-01-25) 1.3     (2021-06-11) <
netcdf4           1.5.3   (2019-03-24) 1.5     (2019-03-24) = (w)
numba             0.54    (2021-08-29) 0.53    (2021-03-12) > (!)
numpy             1.20    (2021-01-31) 1.19    (2020-07-16) > (!)
packaging         20.0    (2020-01-06) 21.0    (2021-07-03) <
pandas            1.2     (2020-12-26) 1.3     (2021-07-02) <
pint              0.17    (2021-03-22) 0.17    (2021-03-22) =
pseudonetcdf      3.1     (2019-10-18) 3.1     (2019-10-18) =
pydap             3.2     (2018-10-16) 3.2     (2018-10-16) =
rasterio          1.2     (2021-01-25) 1.2     (2021-01-25) =
scipy             1.6     (2020-12-31) 1.7     (2021-06-27) <
seaborn           0.11    (2020-09-08) 0.11    (2020-09-08) =
sparse            0.12    (2021-03-19) 0.12    (2021-03-19) =
toolz             0.11    (2020-09-24) 0.11    (2020-09-24) =
typing_extensions 3.7     (2019-02-25) 3.10    (2021-06-03) <
zarr              2.8     (2021-04-25) 2.8     (2021-04-25) =

[headtr1ck]

Ok seems @Illviljan and myself were a few month off xD #6821 (comment).

Nvmd then :)

[Illviljan]

Oh no, I read the wrong table, sorry about that. :(

It's this table https://numpy.org/neps/nep-0029-deprecation_policy.html#drop-schedule :
On Jan 31, 2023 drop support for NumPy 1.20 (initially released on Jan 31, 2021)

[dcherian]

Looks like we can bump a bunch of others if you're up for it :)

[headtr1ck]

It's this table https://numpy.org/neps/nep-0029-deprecation_policy.html#drop-schedule :
On Jan 31, 2023 drop support for NumPy 1.20 (initially released on Jan 31, 2021)

We are dropping 1.19 here, where this table says:
"On Jun 21, 2022 drop support for NumPy 1.19"
What is now correct?

[keewis]

you can probably use the minimum version policy CI to check (we have a policy that slightly deviates from NEP29).

I think we can drop 1.18, but 1.19 should probably still be supported for a while longer.

[headtr1ck]

Ok I think I found my misunderstanding.

Instead of dropping versions older than 18 month we want the oldest supported version to be at least 18 month old.

[headtr1ck]

Anyone know why min all deps fails?
As far as I can tell the only requirement of netCDF is python 3.8?

[headtr1ck]

I got rid of the complete npcompat module.
Not sure if you actually want to keep it around in case new incompatibilities arise?
But I guess we can always add it back.

[headtr1ck]

@dcherian I see that you were working on this the last time.
Do you have any idea why mamba fails to resolve the environment (for min-all-deps)?
With netcdf4=1.5.3 it fails for some python dependencies and with 1.5.4 it fails to find the correct cmds2?

We should really merge this, since all CI runs are currently broken due to new flox requiring numpy>=1.20

[tomvothecoder]

With netcdf4 1.5.4

Encountered problems while solving:
    - package cdms2-3.1.0-py27h6091dcd_0 requires python >=2.7,<2.8.0a0, but none of the providers can be installed

@tomvothecoder, I see you're committing to cdms. Is this something you've run in to?

since all CI runs are currently broken due to new flox requiring numpy>=1.20

sorry! trying to fix this in conda-forge/flox-feedstock#22

I haven't ran into this issue since I've only committed documentation and README updates for the repo.

@jasonb5 maintains the repo dependencies, so let's see if he can help. Jason, can you try taking a look at this issue? It looks like cdms2 is breaking the xarray builds.

As a side-note, cdms2 (and CDAT as a whole) is now in maintenance-only mode and support will cease around the end of CY 2023. Just an FYI to plan for possibly dropping cdms2 from xarray in the future. CDAT's successor is xCDAT (Xarray Climate Data Analysis Tools) .

[headtr1ck]

Is there a reason why the minimum versions are not in the additional requirements in setup.cfg?

[dcherian]

Is there a reason why the minimum versions are not in the additional requirements in setup.cfg?

See #5578

[jasonb5]

@dcherian @tomvothecoder cdms2-3.1.0 is old, the latest, final version is 3.1.5. Updating the version should probably resolve any dependency issues.

[headtr1ck]

We usually only fix the minor version, in this case cdms2=3.1. But it seems to work now...

LGTM!