I can't build the last TAG

I'm trying to build the last tag from bob.blitz so I can use it with the new release of bob.extension, but....

https://gitlab.idiap.ch/bob/bob.blitz/-/jobs/141543

ping @andre.anjos @amohammadi

It's some API compatibility problem with numpy, but everything (bob.blitz and its bob dependencies) was built with numpy 1.13.3.

Can I have some help, please?

Thanks

If you search for "numpy:" on that log, you'll see 2 versions "1.9.3" (build) and "1.13.3" (deployment/test). Apparently, they are not compatible with each other.

good catch.

I remember discussing with @amohammadi that this is a desired behaviour

The numpy setup was exactly the same when we tagged this package last time.

https://gitlab.idiap.ch/bob/bob.blitz/-/jobs/130141

1.9.3 to build and 1.13.3 to test

This is a conda bug. Numpy 1.14 is used to build here:

numpy-base: 1.14.5-py27hdbf6ddf_1 defaults

While we asked for numpy 1.9

Ok, so how can we fix this?

We have to open an issue on Anconda issues on Github.

Does this happen for beta packages too?

no

Master is broken too

but nightlies is not

Nightlies is cheating

Shall we pin the numpy version? We can't wait for conda

Isn't it already pinned?

Not explicitly

Could you please show how you want to change it? I do not understand.

requirements:
  build:
    - {{ compiler('c') }}
    - {{ compiler('cxx') }}
    - pkg-config {{ pkg_config }}
    - cmake {{ cmake }}
  host:
    - python {{ python }}
    - setuptools {{ setuptools }}
    - bob.extension
    - numpy 1.13.3
    - libblitz {{ libblitz }}
    - boost {{ boost }}
  run:
    - python
    - setuptools
    - numpy 1.13.3
    - boost

This is going to conflict during build and test as soon as we update bob-devel.

Also, it will limit users to numpy 1.13.3 only which is very limiting since numpy guarantees backward compatibility.

Also, it will limit users to numpy 1.13.3 only which is very limiting since numpy guarantees backward compatibility.

I know, but today we have no way to release new stuff unless we try to solve it with conda, which I don't know how long it will take.

This is something that is not under our control and for me this is a big issue.

@amohammadi said that we should link bob.blitz with numpy-devel now

Looks like the numpy package is split into numpy-base, numpy-devel, and numpy packages: https://github.com/AnacondaRecipes/numpy-feedstock/blob/master/recipe/meta.yaml For packages that link against numpy at C level they should use numpy-devel. However, the numpy-devel had broken dependencies and it's fixed in numpy-devel 1.14 and the numpy metapackage had broken dependencies too which were fixed in https://github.com/AnacondaRecipes/numpy-feedstock/commit/211f8914f85df1fe337a3aa81cff94ec7a3e3b62#diff-e178b687b10a71a3348107ae3154e44c so we have to do two things:

1. Put numpy 1.14.4 (or even 1.14.5 while we are it) in bob.admin and bob-devel

2. Change the conda recipe of all C++ bob packages which link to numpy, from:

host: - numpy {{ numpy }} run: - {{ pin_compatible('numpy') }}

to:
```yaml
host:
 - numpy-devel {{ numpy }}