Do they just remove things from the main channel? That seems pretty bad. Do you think we should open an issue on that?

They removed numpy=1.11.0, matplotlib=1.5.1, and scipy=0.17.0.

You may still be able to access the older packages (e.g. pre Sept. 2017) by explicitly adding the "free" channel. It used to be included by default, but they changed that fairly recently.

e.g. conda config --set restore_free_channel true
https://docs.conda.io/projects/conda/en/latest/user-guide/configuration/free-channel.html#using-the-free-channel

The rationale for this decision on Anaconda's end is explained here:
https://www.anaconda.com/why-we-removed-the-free-channel-in-conda-4-7/

I figured they wanted to prune the search space, but breaking CI is annoying ;) Thanks for the explanation!

Not 100% sure if that is what is going on here, but I had recently experienced similar issues and think that was the cause. I think otherwise packages only tend to get removed if they were "broken", but in those cases they would typically be replaced with a working version with a newer build number.

You can always pin conda

conda install conda=4.6

The packages will be available there.

@glemaitre yeah I was also considering that but I feel like we'd just be asking for more issues down the road? Not sure.

Oh, this would have saved time fixing the release branches...

Lots of packages were broken in the last conda version as well. For instance, the csv files of the scikit-learn package were corrupted. I think it was something linked to some compression applied to these files. Maybe conda is rebuilding the packages in the meanwhile.

We could also use conda-forge more in the CI. It's more transparent and easier to understand what is happening there.

Thanks! I think this PR could have used a link toward the comment where this issue was detected (#14507) as otherwise @lesteve ended up proposing a independent PR to solve it in #14515

I think this PR could have used a link toward the comment where this issue was detected

That would have helped indeed but to be fair I could have checked existing PRs too ...