CI [scipy-dev] now uses python 3.7 #11409
Conversation
|
Maybe it could be better to add python 3.7 as another build to the Travis build matrix? Since scipy-dev is a cron, it's less systematically checked. I don't think build time would increase that much with another build as Travis appears to use multiple parallel workers... |
|
Ah, I'd not actually realised it left RC already.
Travis doesn't yet have Python 3.7 natively available, as far as I
understand from the related SciPy issues.
|
Yes, but we could very well install Python 3.7 with conda then numpy & scipy with pip as you doing there. Essentially it could be just copying the |
|
As long as we can use conda packages for all dependencies, then fine...
otherwise the run will be a lot slower than the others.
|
|
Why? The scipy/numpy manylinux wheels with OpenBlas would be comparable in speed, aren't they? In my experience in other projects I was originally biased toward using conda, but it turns out that due to slow dependency resolution in conda + downloading/unpacking mkl which is huge, it's often faster to just use wheels from PyPi, if test runtime is comparable. |
|
Well I'm not entirely certain why, but (here)[https://travis-ci.org/scikit-learn/scikit-learn/builds/399462496?utm_source=github_status&utm_medium=notification] it takes 10 more minutes to build for 3.7. |
|
Running
|
build_tools/travis/install.sh
Outdated
| @@ -82,8 +87,8 @@ elif [[ "$DISTRIB" == "scipy-dev" ]]; then | |||
| # Set up our own virtualenv environment to avoid travis' numpy. | |||
There was a problem hiding this comment.
If we end up using conda to have Python 3.7 I think we can get rid of this comment ("Set up our own virtualenv ...") and the one below:
# This venv points to the python interpreter of the travis build
# matrix.
|
Actually, the test runtime is an issue but it should improve quickly as soon as the binary packages get uploaded. |
They are already uploaded for stable numpy and scipy . The reason this build takes so long ,I think, is because it uses scipy-dev (i.e. numpy dev wheels aren't available for Py3.7 yet from what I gather). So having one separate build for scipy-dev (cron) and one for Py3.7 (non cron) using stable numpy and scipy would be a way around it.. |
| # matrix. | ||
| virtualenv --python=python ~/testvenv | ||
| source ~/testvenv/bin/activate | ||
| make_conda python=3.7 |
There was a problem hiding this comment.
I now wonder if we should just use python here without pinning the version.
There was a problem hiding this comment.
Anticipating Python 4 ? )
I think it doesn't hurt to indicate explicitly for the reader that it tests 3.7 ..
|
(merged master to resolve merge conflict) |
|
CI passed except for the PyPy3 build which is now temporary disabled. Merging, thanks @jnothman ! |
Fixing #11320...?