Replicate errors found in the 3.7 env in travis

[massich]

Try to replicate the errors we are facing in 3.7 travis runs in the following PRs:

• #5819
• #5718

[massich]

I do replicate locally. I'll try to see what is all that about

[larsoner]

FYI same failures are happening on maint/0.17, so we'll want to backport

[massich]

FYI same failures are happening on maint/0.17, so we'll want to backport

yes there has been a numpy realease and it's not compatible with scipy=1.2.0 (I think)

If this PR is green I suggest to merge it, and make a PR removing those 2 lines + fixing whatever is wrong.

[larsoner]

We probably don't need to fix anything, just wait for the SciPy 1.2.0 + NumPy 1.16.0 wheels to land. 1.16.0 came out yesterday IIRC so it might take a day or two ... ?

[massich]

numpy/numpy#12736

[massich]

@larsoner yes, we can just do the PR to keep track of it, re run the CI in a couple of days and if its green. Done!!

[massich]

maybe there was a nicer way to rollback to 1.15.4, but ..

[larsoner]

Actually I get it to go away locally if I add:

ignore:numpy.ufunc size changed:RuntimeWarning

to setup.cfg.

This makes sense if the sklearn wheels are old, which they almost certainly are.

Can you see if this fixes things for you? If so we should add it for a day or two rather than pin the NumPy version.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (master@e0df416). Click here to learn what that means.
The diff coverage is n/a.

@@            Coverage Diff            @@
##             master    #5834   +/-   ##
=========================================
  Coverage          ?   88.58%           
=========================================
  Files             ?      369           
  Lines             ?    69125           
  Branches          ?    11649           
=========================================
  Hits              ?    61232           
  Misses            ?     5038           
  Partials          ?     2855

[massich]

It does fix it on my end.
How did you come up with that? got it. It's just that I would not thought of that.

[massich]

I guess that if anything is crucial, some other test would actually break, some buffer would go crazy, or something.

[larsoner]

I looked at the topmost error in the log, it was our suite turning a warning into an error

[mattip]

This message is coming from scikit-learn apparently they are using cython<=0.29 where the numpy.pxd was updated to not emit that warning when upgrading NumPy.

See scikit-learn/scikit-learn#12829