condaenv

a working replacement of the conda env command

$ conda env -f environment.yaml doesnt follow your instructions. It will happily create an environment that does not respect the channels and versions of dependencies (channel-name::package=version) you specified. This is unacceptable. Maintaining a conda environment.yaml of any complexity feels like

Operation

• conda install -n root pyyaml # make sure yaml is installed
• condaenv -f environment.yaml -n environmentname

If conda is unable to install the packages as specified it will error instead of silently succeeding (which conda env would do, thus ruining the next 3 weekends because your environment doesnt not contain what you specified it to contain)

Why

TL/DR: conda <4.6 , when calling conda env install/force ignores the channel pin.

We, and many other people use conda environment yamls, such as below, to describe self contained environments using packages from different channels. I'm not sure if the conda developers are deprecating the conda env command, but it is common in conda version <4.6 that conda will simply ignore the channel pins channel::package and take the most recent build across all channels listed.

name: stitchup

channels:
  - loopbio
  - conda-forge
  - defaults

dependencies:
  - conda-forge::python=2.7
  - loopbio::ffmpeg=4.0.2=nonvidia_gpl*
  - loopbio::opencv=3.4.3

Depending on who (which channel) has the most recent opencv build of that version, a different package will be installed. When combined with an environment of any decent complexty, and with the fact that conda(forge) developers often rebuild minor or micro build increments with new or changed dependencies!!! will often lead to environments being installed that do not contain what you specified them to contain.

To make matters worse, conda env often silently succeeds, which means you will just get random crashes because you have mixed incompatible and/or poorly specified (by other packages) dependencies.

If you want to use conda envs environment.yaml then the only way to guarentee that you get the build from the channel you requested is to use hashed build strings. But if you are doing this, you might as well use explicit environment / lists of packages anyway.

A 'known' workaround has been to just call conda create not conda env with all the packages in the environment yaml. That's what this tool does

In principle this is fixed in conda 4.6.