Cached Miniconda in CircleCI

circle.yml

machine:
  environment:
    MINICONDA: "$HOME/miniconda"
    MINICONDA_PATH: "$MINICONDA/bin"
    PATH: "$MINICONDA_PATH:$PATH"
    CONDA: "$MINICONDA_PATH/conda"
    ANACONDA: "$MINICONDA_PATH/anaconda"
    # This ensures we don't accidentally pull in ~/.condarc
    # if it happens to be there for some strange reason
    CONDARC: "$MINICONDA/condarc.none"
    
checkout:
  post:
    - git fetch --depth 1000000
    - git fetch --tags

dependencies:
  override:
    - rm -rf virtualenvs venv .pyenv
    - |
      if [[ ! -f $MINICONDA/.finished ]]; then
          rm -rf $MINICONDA
          wget $ANACONDA_REPO_URL/static/extras/Miniconda3.sh
          bash Miniconda3.sh -b -p $MINICONDA
          # Use --system so the settings are placed in $MINICONDA
          # Add any other conda settings you like here
          $CONDA config --system --add channels conda-forge
          # If you'd like to make sure conda is up to date
          $CONDA install conda conda-build anaconda-client --yes
          touch $MINICONDA/.finished
      fi
    # Install some packages in our _cache environment that we know we're going to
    # need during the build process---say, your conda package's dependencies. Because
    # CircleCI saves this cache _before_ the build process proper, there is no way
    # to just let that build process populate the cache. So we install them into
    # this environment---and then promptly remove them. I have some rather large
    # packages in my particular use case so this helps me.
    - $CONDA create -n cache ... --yes # add packages here
    - $CONDA remove -n cache --all --yes
    # Ideally, what I'd do is blow away the tarballs and leave the extracted packages.
    # That would save the extraction cost during the build process. Unfortunately, the
    # absence of a tarball is interpreted as a missing package, because there's no MD5
    # signature available to verify. I understand why this is done (indeed I might 
    # have implemented it) but this is certainly a downside. So what I've chosen to
    # do is clean out the extracted packages here and suffer the extraction cost,
    # trading smaller cache tarball size for package extraction time. Experiments are
    # needed to determine if it is worthwhile.
    - $CONDA clean --packages --yes
    # Likely unnecessary, but just in case.
    - $CONDA build purge-all
  cache_directories:
    - ~/miniconda

test:
  override:
    - $CONDA build conda-recipe

deployment:
  production:
    branch: master
    commands:
      # The anaconda-client configuration is not cached so we have to set the URL here
      - $ANACONDA config --set url $ANACONDA_REPO_URL/api
      - $ANACONDA -t $ANACONDA_REPO_TOKEN upload --user creditflow --label dev --summary "auto-uploaded by circleci" $MINICONDA/conda-bld/*/*.tar.bz2 --force
 

An ongoing experiment in building a circle.yml that caches miniconda.

Thanks @mcg1969, this came in handy for me today.