Package enviornment pulling in old dependencies

[rob-p]

I had a user report this in the salmon repository, but it seems more related to bioconda and the current packaging. I don't know what might cause this, so any suggests are welcome! The original report is copied below (and due to @bounlu, who I am tagging here):

---

ok, this is not really a bug but I think it needs attention.

i am using anaconda environment for many tools and keep it updated as much as possible with the conda update --all command. but the bottleneck is salmon, which has very old dependencies, and it's a dilemma either to update salmon from 0.8.1 (pretty old buggy version) to the latest one (0.11.3) and downgrade a bunch of other important packages, or vice versa.

Describe the bug
a bug is a species of animal kingdom, a small insect (just kidding)

To Reproduce
Steps and data to reproduce the behavior:

$ conda update salmon
Solving environment: done

## Package Plan ##

  environment location: /home/software/anaconda2

  added / updated specs: 
    - salmon


The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    salmon-0.11.3              |       h86b0361_2         2.9 MB  bioconda
    blas-1.0                   |              mkl           6 KB
    numpy-1.14.3               |   py27h28100ab_1          41 KB
    ------------------------------------------------------------
                                           Total:         3.0 MB

The following packages will be UPDATED:

    jemalloc:     4.5.0-0                               bioconda    --> 5.1.0-hfc679d8_0      conda-forge
    libgcc-ng:    7.2.0-hdf63c60_3                      conda-forge --> 8.2.0-hdf63c60_1                 
    libstdcxx-ng: 7.2.0-hdf63c60_3                      conda-forge --> 8.2.0-hdf63c60_1                 
    salmon:       0.8.1-0                               bioconda    --> 0.11.3-h86b0361_2     bioconda   

The following packages will be DOWNGRADED:

    blas:         1.1-openblas                          conda-forge --> 1.0-mkl                          
    fastqc:       0.11.7-5                              bioconda    --> 0.11.6-2              bioconda   
    gsl:          2.4-blas_openblash47a8a8e_1           conda-forge [blas_openblas] --> 2.1-2                 conda-forge
    numpy:        1.15.1-py27_blas_openblashd3ea46f_1   conda-forge [blas_openblas] --> 1.14.3-py27h28100ab_1            
    openjdk:      8.0.144-zulu8.23.0.3_2                conda-forge --> 8.0.121-1                        
    scikit-learn: 0.19.2-py27_blas_openblasha84fab4_201 conda-forge [blas_openblas] --> 0.19.1-py27hedc7406_0            
    scipy:        1.1.0-py27_blas_openblash7943236_201  conda-forge [blas_openblas] --> 1.1.0-py27hd20e5f9_0

Expected behavior
salmon should be updated to the latest version without the requirement of downgrading of its dependencies.

Linux nscc04 2.6.32-696.30.1.el6.x86_64 #1 SMP Fri May 18 11:50:44 EDT 2018 x86_64 x86_64 x86_64 GNU/Linux

[bgruening]

This looks for me like the update is preferring defaults or conda-forge. Can you make sure that conda-forge has priority?

[PertuyF]

Just to mention, I saw a similar behaviour a few times. I my case it was always related to the command updating numpy, and preferring mkl feature over openblas. Despite my .condarc specifying:

channel_priority: True
channels:
  - bioconda
  - conda-forge
  - defaults

Most of the time I got it solved using something like conda install <package> blas=*==openblas instead of conda update <package> as a workaround. But of course finding where it comes from would be better.

[rob-p]

Tagging @bounlu to see if trying any of this helps in his case.

[PertuyF]

Looks like it could be related to features (see discussion here about similar issue with openjdk).
If you specify pinned_packages in a .condarc you will be able to avoid switching from bioconda/conda-forge to defaults.

For instance, referring to my last example, the following file allows to restrict blas to conda-forge channel, thus enforcing numpy to use the feature openblas from this channel.

pinned_packages:
  - conda-forge::blas

Could it be a solution with minor side-effects?