Create custom docker image for CI about alibi HOT 8 CLOSED

Currently, the .travis.yml uses three Python images (3.5, 3.6, and 3.7). Should there be three different custom images, one for each of these versions of Python?

from alibi.

I think one could use a single image and manage the different Python versions for CI using virtualenv/conda. That being said, the CI docker image would probably be distinct from a docker image containing alibi to be used immediately (it would only include a python interpreter, alibi and all its dependencies), in which case it would make sense to have different versions. Were you thinking of the CI or the user facing image?

from alibi.

I was thinking of the CI (speeding up the

  - pip install -r requirements/requirements_all.txt
  - python -m spacy download en_core_web_md

part)

from alibi.

Yes, that would be the main reason for using a custom image (and we could use it for new builds and releases too). Were you interested in contributing a Dockerfile for this?

from alibi.

I'm interested in attempting this - yes. I was thinking that we could have a Dockerfile of this kind:

FROM python:3.7

COPY requirements/requirements_all.txt ./

RUN pip install -r requirements_all.txt
RUN python -m spacy download en_core_web_md

We could then build a Docker image, push it to DockerHub (e.g. as alibi/alibi-ci:latest), and then using this in the travis yml file as follows:

dist: xenial
language: python
image: alibi/alibi-ci:latest
before_install:
  - sudo apt-get install pandoc
script:
  - flake8 alibi
  - mypy alibi
  - travis_wait pytest
  - make build_pypi
  - make build_docs
after_success:
  - codecov

Before I run any of this, I just wanted to check how we would go about such a pull request.

1. Is this what you had in mind? At present, I wouldn't know how to manage different Python versions with virtualenv/conda and just a single Docker image, but I can look into it.
2. The travis file shouldn't reference a Docker image that's in my name, which is what would happen if I pushed the image. So, should the process be:

• one pull request with new Dockerfile;
• you build the Docker image, and push it to Docker hub, and let me know its tag
• another pull request to update the travis file, referencing the new image
  ?

from alibi.

So, after having another look at Travis, if I understand correctly then custom build images are not supported out of the box... (travis-ci/travis-ci#7726). One could get around it by doing docker pull etc. in before_install and then making sure the tests run inside the image (and inside several virtualenvs separately), but I'm not sure if it's worth the hassle at this point.

from alibi.

I'm inclined to agree. Judging by the last successful travis build:

• sudo apt-get install pandoc takes only 13 seconds
• pip install -r requirements/requirements_all.txt takes about 2 minutes
• python -m spacy download en_core_web_md takes 7 seconds.

By contrast, pytest takes almost 6 minutes! So, using a custom Docker image to speed up a 2-minute part of a 9-minute process might not be worthwhile.

I'll have a go at caching pip files instead

from alibi.

Closing as seems like the benefits would be minimal from the current setup.

from alibi.