try dask.array

[aaronspring]

• tries to fix make dask.array dependency not dask to speedup pip install #70, not successful
• speeds up CI

[codecov]

Codecov Report

Merging #71 (75f9184) into master (e01291b) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master      #71   +/-   ##
=======================================
  Coverage   97.18%   97.18%           
=======================================
  Files           2        2           
  Lines         249      249           
  Branches       71       71           
=======================================
  Hits          242      242           
  Misses          5        5           
  Partials        2        2           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e01291b...75f9184. Read the comment docs.

[jrbourbeau]

Thanks @aaronspring! FWIW I'm not sure this will make an large impact on what dependencies are installed. Dask's required dependencies are listed here

https://github.com/dask/dask/blob/c19a8300d514a8ee488e0da5e81c01155d55fdbe/setup.py#L32-L39

which are always installed, regardless if one pip installs dask, dask[array], dask[array, dataframe], dask[complete], etc.

Adding the [array] extra to the dask install will cause numpy to also be installed

https://github.com/dask/dask/blob/c19a8300d514a8ee488e0da5e81c01155d55fdbe/setup.py#L13

This is probably a good thing as it will require xhistogram to install a numpy version which is compatible with dask.array, but will, as of today, not result in a big difference since numpy is listed as a separate dependency already.

That's all to say that this change seems totally reasonable, but isn't the cause for libraries like bokeh being installed as referenced over in #70

[aaronspring]

thanks @jrbourbeau for the quick description

ok. so there is no way around install_requires and the other extras (bokeh in diagnostics) were not even installed before.

so this PR here would only ensure that a matching numpy version is used @rabernat but xarray already kind of does that by itself. should this PR still go forward?

[jrbourbeau]

Ah, I see what the issue might be now. dask[array] (and my comment here #71 (review)) is just a pip thing and isn't supported when installing dask with conda. conda install dask will pull in other packages like bokeh though. Specifically the packages listed here.

If you just want to conda install the minimum required dependencies for Dask (which would be equivalent to pip install dask), you'll want to install the dask-core package from conda-forge.

[aaronspring]

using dask-core reduces Setup Conda Env from 5:37 https://github.com/xgcm/xhistogram/runs/4258344230?check_suite_focus=true to 2:47 https://github.com/xgcm/xhistogram/runs/4258767295?check_suite_focus=true

[aaronspring]

thanks for the explanation. the originating issue was xarray-contrib/xskillscore#359, then we discovered that xhistogram relies on dask.

also xhistogram on conda-forge relies on dask https://github.com/conda-forge/xhistogram-feedstock/blob/master/recipe/meta.yaml - not sure whether there also specifying dask.array works, but you said dask[array] is a pip thing, so likely not, but we could use dask-core on the next release maybe

[rabernat]

If we are happy about this I will merge. 👍 if you think we are done here.

[aaronspring]

happy to merge and maybe we can ship the conda version with dask-core, see conda-forge/xhistogram-feedstock#9