GitLab CI and conda

I setup GitLab to host several projects at work and I have been quite pleased with it. I read that setting GitLab CI for test and deployment was easy so I decided to try it to automatically run the test suite and the sphinx documentation.

I found the official documentation to be quite good to setup a runner so I won't go into details here. I chose the Docker executor.

Here is my first .gitlab-ci.yml test:

image: python:3.4

before_script:
  - pip install -r requirements.txt

tests:
  stage: test
  script:
    - python -m unittest discover -v

Success, it works! Nice. But... 8 minutes 33 seconds build time for a test suite that runs in less than 1 second... that's a bit long.

Let's try using some caching to avoid having to download all the pip requirements every time. After googling, I found this post explaining that the cache path must be inside the build directory:

image: python:3.4

before_script:
  - export PIP_CACHE_DIR="pip-cache"
  - pip install -r requirements.txt

cache:
  paths:
    - pip-cache

tests:
  stage: test
  script:
    - python -m unittest discover -v

With the pip cache, the build time went down to about 6 minutes. A bit better, but far from acceptable.

Of course I knew the problem was not the download, but the installation of the pip requirements. I use pandas which explains why it takes a while to compile.

So how do you install pandas easily? With conda of course! There are even some nice docker images created by Continuum Analytics ready to be used.

So let's try again:

image: continuumio/miniconda3:latest

before_script:
  - conda env create -f environment.yml
  - source activate koopa

tests:
  stage: test
  script:
    - python -m unittest discover -v

Build time: 2 minutes 55 seconds. Nice but we need some cache to avoid downloading all the packages everytime. The first problem is that the cache path has to be in the build directory. Conda packages are saved in /opt/conda/pkgs by default. A solution is to replace that directory with a link to a local directory. It works but the problem is that Gitlab makes a compressed archive to save and restore the cache which takes quite some time in this case...

How to get a fast cache? Let's use a docker volume! I modified my /etc/gitlab-runner/config.toml to add two volumes:

[runners.docker]
  tls_verify = false
  image = "continuumio/miniconda3:latest"
  privileged = false
  disable_cache = false
  volumes = ["/cache", "/opt/cache/conda/pkgs:/opt/conda/pkgs:rw", "/opt/cache/pip:/opt/cache/pip:rw"]

One volume for conda packages and one for pip. My new .gitlab-ci.yml:

image: continuumio/miniconda3:latest

before_script:
  - export PIP_CACHE_DIR="/opt/cache/pip"
  - conda env create -f environment.yml
  - source activate koopa

tests:
  stage: test
  script:
    - python -m unittest discover -v

The build time is about 10 seconds!

Just a few days after my tests, GitLab announced GitLab Container Registry. I already thought about building my own docker image and this new feature would make it even easier than before. But I would have to remember to update my image if I change my requirements. Which I don't have to think about with the current solution.