conda 4.7.5

[minrk]

gets much faster installs

had to move WORKDIR $HOME up due to tempfile permission issues with conda 4.7 (fixes #891, fixes #878). Conda 4.7 creates some tempfiles in the CWD instead of /tmp (actually conda_package_handling).

[parente]

@minrk Thank you for helping out here while I'm heads down elsewhere.

[Bidek56]

This PR seems to break: "conda update", has anyone tried to do: "conda update --yes --all" on the latest image? Thx

[maresb]

I wonder if there's a better way than to stop removing qt. Removing qt reduces the image size from 4.01 GB to 3.68 GB.

If I understand correctly, pyqt is unfortunately an (extraneous?) dependency of matplotlib. Thus subsequent conda install commands will try to reinstall qt, which is annoying.

I think the ideal solution would be to remove qt and convince conda that it's okay without. Does anyone know if this is doable?

[dhirschfeld]

I think in a container where you don't need the gui front-ends you can install matplotlib-base directly rather than matplotlib :

> conda search --info matplotlib=3.1.1=py37_1

matplotlib 3.1.1 py37_1
-----------------------
file name   : matplotlib-3.1.1-py37_1.tar.bz2
name        : matplotlib
version     : 3.1.1
build       : py37_1
build number: 1
size        : 6 KB
license     : PSF-based
subdir      : win-64
url         : https://conda.anaconda.org/conda-forge/win-64/matplotlib-3.1.1-py37_1.tar.bz2
md5         : fe8dc51d1eff32684a89c08e77934201
timestamp   : 2019-07-24 17:46:41 UTC
dependencies:
  - matplotlib-base 3.1.1 py37h2852a4a_1
  - pyqt
  - python >=3.7,<3.8.0a0

i.e. matplotlib just repackages matplotlib-base plus the qt gui toolkit:

> conda search --info matplotlib-base=3.1.1=py37h2852a4a_1

matplotlib-base 3.1.1 py37h2852a4a_1
------------------------------------
file name   : matplotlib-base-3.1.1-py37h2852a4a_1.tar.bz2
name        : matplotlib-base
version     : 3.1.1
build       : py37h2852a4a_1
build number: 1
size        : 6.6 MB
license     : PSF-based
subdir      : win-64
url         : https://conda.anaconda.org/conda-forge/win-64/matplotlib-base-3.1.1-py37h2852a4a_1.tar.bz2
md5         : 4e22816c97c10fecee29fa4e05a7884d
timestamp   : 2019-07-24 17:46:01 UTC
dependencies:
  - cycler >=0.10
  - freetype >=2.9.1,<3.0a0
  - kiwisolver
  - libpng >=1.6.35,<1.7.0a0
  - numpy >=1.14.6,<2.0a0
  - pyparsing
  - python >=3.7,<3.8.0a0
  - python-dateutil
  - setuptools
  - tornado
  - vc >=14,<15.0a0
  - zlib >=1.2.11,<1.3.0a0

[parente]

Thanks for the ref @dhirschfeld. matplotlib-base didn't exist when we originally setup the Dockerfile and so we resorted to the uninstall hack.

I'm assuming @minrk recently removed the uninstall to avoid conda environment inconsistencies. I think we should switch to the base package unless my assumption is incorrect.

[maresb]

In the context of scipy-notebook, @dhirschfeld's solution works great, so I made the pull request above.

Frustratingly, other conda packages which I want to install later depend on matplotlib instead of matplotlib-base (e.g. geopandas and descartes). As soon as I installl one of those, I'm unfortunately back in the previous situation. (Note to self: I should submit PRs to switch the corresponding feedstocks to matplotlib-base.)

[dhirschfeld]

Note to self: I should submit PRs to switch the corresponding feedstocks to matplotlib-base

I'd strongly encourage you to do so - it'll improve the whole ecosystem 👍