Dependency Hell

I started my Python package management journey years ago using pip, then more recently I embraced Anaconda and conda more fully (particularly with the “conda-forge” repository) to resolve complex dependencies along with system/binary dependencies. Recently, when attempting to update our team’s standard Python docker image with the latest versions of the packages we use, and include some new ones, it appears that relying on conda and conda-forge is untenable: I have been unable to resolve the appropriate set of versions for the scientific Python packages our team require for our work. I have moved back to pip for packages which are not provided in the default Anaconda repository. pip has and continues to make a number of improvements, and had no problem providing our extra dependencies.

Reproducing Conda Environments

This short summary is based on the Anaconda blog post here https://www.anaconda.com/moving-conda-environments/. The original blog post is a great high-level summary for the various methods in conda for reproducing environments.

• OS and platform specific (pulls from repos)

  # On source environment:
  conda list --explicit > spec-list.txt
  

  # New conda environment:
  conda create --name new_env_name --file spec-list.txt
  

• Different platforms and OS (pulls from repos, also includes pip installed packages)

  # On source environment:
  conda env export > env.yml
  

  # New conda environment:
  conda env create -f env.yml
  

• Platform and OS specific, no internet on target

Installing NetCDF Python Packages

I was trying to remember how I have installed netCDF4 and related libraries for Python, and what I need to do differently for Windows systems vs. the Linux systems I usually use.

On Linux, sometimes I use the system netCDF C libaries, but often I compile and install specific versions of HDF5 and netCDF4 from scratch. Here is how I have built netCDF for various Docker container images.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

# Build HDF5
cd hdf5-x.y.x
./configure --prefix=/usr/local --enable-shared --enable-hl
make
make install
cd ..

# Built NetCDF4
cd netcdf-x.y.z
LDFLAGS=-L/usr/local/lib
CPPFLAGS=-I/usr/local/include
./configure --enable-netcdf-4 --enable-dap --enable-shared --prefix=/usr/local --disable-doxygen
make
make install
cd ..

# NetCDF4 Fortran
cd netcdf-fortran-x.y.z
./configure --enable-shared --prefix=/usr/local
make
make install
cd ..

netcdf4-python #

I typically try to use pip to install Python libraries if I can. Use pip if you need to, but I am using conda and conda-forge as much as possible now, in fact by using conda, the above compilation steps are usually not necessary as far as I know. See below.