-
Notifications
You must be signed in to change notification settings - Fork 17
Fix macos version matrix #29
Conversation
Group all macOS entries together. Remove redundant entry for Python 3.5. Pin scipy version for python 3.6 build.
Why are no Travis runs starting? |
Now some Linux builds are failing: they are trying to build wheels for numpy and scipy, and the scipy build is failing because numpy isn't there yet... Or at least that's my interpretation |
None of the jobs marked daily build are working here now |
Any chance @thomasjpfan can help with this?
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We have wheels built for all but Mac on Py3.6. There it is going through the same compilation errors as the linux builds due to no wheel for Scipy on that platform, it seems.
Let's just see if we can ship that wheel then look into longer-term fixes.
The build I wanted finished but didn't upload for some unknown reason... |
I'll twine the wheels I have for now. |
Indeed we need to find the oldest known good set of dependicies for numpy / scipy for each Python version to build the scikit-learn wheel and the the most recent known good set of dependicies for numpy / scipy to run the tests. |
It seems we are going to release on Friday again :) |
I don't know the English version but I think you can guess the meaning: https://www.estcequonmetenprodaujourdhui.info/ |
I think if we changed macpython to install numpy strictly before scipy, it could work... Anyway, it's green here and so, pending the switch to Azure, we should merge |
Hopefully Mac Py3.6 turns up on http://wheels.scipy.org/ soon. I won't be
able to twine it, unfortunately. Could you, please, @ogrisel?
|
Yes I will. Thanks for fixing the remaining builds. |
I think I understand the issue described in scikit-learn/scikit-learn/pull/13915#issuecomment-495319072:
The test dependencies are not updated to the latest stable automatically: so it can be the case that building on old numpy could run the tests on a too recent version of scipy that would cause problem at test time.
Let's see if pinning everything fixes the problem.