Use pyface < 7

[cbrnr]

Fixes (works around) #8040 hopefully.

[cbrnr]

This seems to work, but it would be better if we used the most recent pyface (7.0.1). The problem is that for some reason, traits is pinned to 5.2.0, but pyface 7.0.1 requires the latest traits 6.1.0 (which is available in the default channel).

[larsoner]

Do we know which package actually requires pyface<7?

[cbrnr]

Do we know which package actually requires pyface<7?

No, I tried to let conda solve the conflict, but after working for 15 minutes it spat out a very long list of conflicts that didn't really help me (because every package seemed to conflict with several other packages).

[larsoner]

Which OS? I tried on macOS and Linux and in both cases conda update conda followed by conda create -n test -f environment.yml from master worked, in both cases installing pyface 7.0.1. So maybe this was just a transient issue?

[larsoner]

To check, maybe remove the numpydoc and the <7 pin from pyface and see if it succeeds here?

[cbrnr]

Nope, doesn't work (still downloads and installs 5.2.0). The error log mentions qt4 = traitsui.qt4:toolkit - do we require Qt4 (instead of Qt5) somewhere?

[larsoner]

If you change it to traits>=6 it should hopefully fail earlier. It's in win64:

https://anaconda.org/anaconda/traits

So some package must require it. I'll fire up the Windows VM for local debugging...

[cbrnr]

Also, how did conda create -n test -f environment.yml (as you've mentioned here) work? First, -f should be --file, but even then do I get an error because the file wants to install in an env called mne. That's why I have to use conda env update --file environment.yml.

[larsoner]

Sorry I meant conda env create not conda create

[cbrnr]

Sorry I meant conda env create not conda create

Thanks, this works (pretty inconsistent argument naming I would say). It does give the same results as conda env update --file environment.yml though, which is traits==5.2.0 on my macOS machine.

Why do you think this is related to Windows? The latest version is on conda for all platforms, and we get the issue on a Linux CI (and on my local macOS).

[larsoner]

I thought it was installing 6 on macOS and Linux but it wasn't. Changing it to have - traitsui>=6 and - traits>=6 fixes things for me on Linux (pytest mne/gui didn't work before but does now). No idea why it tries to pull traits==5.2.0 if you don't tell it to use 6+. For some reason it installs Python 3.6 though...

[larsoner]

Adding a python>=3.8 or even python>=3.7 does create conflicts. Argh...

[cbrnr]

This whole conda thing is a mess. We mix/overwrite with pip because the default channel lags behind recent versions. Should we consider moving to conda-forge as @hoechenberger has repeatedly suggested? It looks like such a situation can occur anytime with our env file with the defaults channel.

[cbrnr]

(Moving to conda-forge also has the advantage that we can use mamba instead of conda, which is much faster and has more features).

[larsoner]

Should we consider moving to conda-forge as @hoechenberger has repeatedly suggested? It looks like such a situation can occur anytime with our env file with the defaults channel.

I've seen problems with conda-forge due to packages not being ready for prime time in other repos that use conda-forge, so I wouldn't think of this as something that will (EDIT: necessarily) solve our problems. It might fix more problems than it creates, but it might also be the reverse.

In any case, this seems to be some base problem. Even this causes a conflict:

name: mne
channels:
- defaults
dependencies:
- python>=3.8
- traits>=6
- pyface>=7
- traitsui>=6

So maybe we should manually pin pyface until they sort it out, and raise an upstream issue.

[cbrnr]

Agreed - can you raise the upstream issue? You already have a great minimal example. I can pin pyface to the old version in the meantime.

[cbrnr]

It would still be worth giving conda-forge a try, I'd assume that it will solve more problem than it creates (especially the whole pyqt/PyQt5 conda/pip business).

[hoechenberger]

It would still be worth giving conda-forge a try, I'd assume that it will solve more problem than it creates (especially the whole pyqt/PyQt5 conda/pip business).

Yes this could definitely be worth a shot. The current approach with conda + pip leads to inconsistent environments that behave in the strangest manner if you try to install or update any conda packages. Not a fan.

[larsoner]

Thanks @cbrnr

[larsoner]

conda/conda#10095

[drammock]

by the way, it would have been fine to specify spyder-kerneles>=1.9.2 as @larsoner suggested. It just wouldn't actually fix the problem with macOS Catalina. But no need to go back and do it; it's already WIP in #8020