Fix environment solving in tests

[philippjfr]

Unpin a bunch of stuff to see if that a) fixes the solve and b) still lets the tests pass.

[maximlt]

Ok this is still failing unfortunately. As for the environment, we had conda-forge in the mix which made things trickier probably. Removing it leads to a state where the env creation is failing on Mac only. It fails with this error:

Preparing transaction: ...working... done
Verifying transaction: ...working... done
Executing transaction: ...working... done
TaskFailed - taskid:develop_install
Command failed: 'pip install --no-deps --no-build-isolation -e .' returned -9

I've been able to reproduce it locally but I can't get my head around why this fails. I then tried to run the apparently failing command (pip install ...) directly and it succeeded. I added this piece in dodo.py:

import pyctdev._conda
# Mocking python_develop originally set to pip install ...
pyctdev._conda.python_develop = 'echo "TEST"'

And well:

Preparing transaction: ...working... done
Verifying transaction: ...working... done
Executing transaction: ...working... done
TaskFailed - taskid:develop_install
Command failed: 'echo "TEST"' returned -9

I don't understand what is going on there!

As for the tests, the CI is failing on Python 3.6 because it's pulling a dev release of panel that has an issue with 3.6. I've seen at least another test failing, due to holoviews itself this time.

[jbednar]

Bokeh 2.4 requires Python 3.7+, and with Panel tied tightly to Bokeh, does that mean we have to disable Python 3.6 tests?

[philippjfr]

Shouldn't have to since Holoviews does not have to pin tightly to any specific Panel version.

[Hoxbro]

A small comment about removing conda-forge is that I think it will only postpone the problems and not fix them, as the packages in default channels are not as up-to-date with the latest releases. This may give new users (and developers) problems when installing the holoviz stack (from pip or with conda-forge), with a newer version of numpy, pandas, or others.

Another thing to be noted is that many scientific packages follow NEP29, which will make it harder to solve the environment for python 3.6 in the future.

[maximlt]

This test MPLRendererTest.test_render_holomap_not_embedded is failing now on Ubuntu py38 and Mac py37. It seemed like a transient to @jlstevens but it would be worth having a look into it.

I'll submit a small commit soon to change a logging step, let's see whether this test still fails and if so how.

[jlstevens]

To get the hotfix release out for bokeh 2.4 support I will go ahead and merge. The one failure looks like an unusual transient and I now think there are fewer transients to worry about.

There are a number of TODO items to follow up on in subsequent PRs:

• Fix the issue with Sankey (test skipped right now)
• Improve/fix the render_holomap_not_embedded tests
• Improve the testing infrastructure and remove the hacks if possible.

Thanks @maximlt for figuring out the workarounds!