iamtxena/docker-library-test

Docker tests

Testing how to properly set up a Dockerfile to use python libraries installed from source.

Docker Cheat Sheet

Build image

Build an image from a Dockerfile.

You need to specify the path . and you can also set the image name with -t

docker build . -t test

Run image

Run an image, setting the container name, running a bash interactively and removing the container after the execution.

You can set the option -d (detach) to run it on the background.

docker run --name test --rm -it test bash

Start a container

Starts a generated container and run it interactively.

docker start -i test

Stop a container

Stops a running container.

docker stop test

Dockerfile setup

Base setup using qibo from the official Pypi repository

• Qibo library requires scipy, numpy and matplotlib. Those libraries are not possible to be installed using pip on an ARM Apple chip, and requires to be installed using conda from miniconda.

• There are a few pre-built docker images containing miniconda, the best one seems to be continuumio/miniconda3. Although on their github exists the alpine version of it, the one you pull from it, it is a debian-based image. The thing with this one is that you cannot install any other system library (but you can install any python library indeed), as the user is non-root. In my case, I need to install some other system packages, so it is not a suitable option for me.

• Another option is bulding the same image on your own. I tried to do that with the alpine version, but did not succeed. Conda uses glibc, and you need then to either install it on your own, or use a pre-built image with glibc installed. My code requires python 3.9 because I am using typings, so I need to take this into account when installing my own conda version to properly download one compatible with python 3.9.

• So, when following the continuumio docker steps to build the glibc, it raises an error because it cannot find the generated libraries. I tried a few times but I did not find the proper solution. Then, I decided to use another prebuilt image with glibc installed from frolvlad/alpine-glibc, and using the continuumio instructions to install conda in it, using the CONDA_VERSION=py_39_4.10.3 and its sha256sum.

• Just make sure to set the PATH variable with conda and bin directories, and initializing conda:

ENV PATH /bin:/sbin:/usr/bin:$PATH
ENV PATH /opt/conda/bin:$PATH
RUN conda init

• You can then install the required packages: scipy, numpy, and matplotlib using conda.

RUN conda install scipy numpy matplotlib

• And finally you can install qibo

RUN pip install qibo

Using qibo from a local source code installation