Base image with Conda python package manager

[atselousov]

Is it possible to use conda package manager in our base image?

What it can give:

• Using of conda can simplify the installation of packages which have non-python dependencies (e.g. see installation here; library which is required by torchvision is installed with one line).

• Another useful use-case is supporting envs with different versions of libraries (e.g. PyTorch, TensorFlow, etc.). Choose of different environments is supported by jupyter-notebook and user switch easily between existed envs after they started a job. Environments are organized in a similar way in AWS SageMaker.

• It can also give chance to users to create their own environments with cookiecutter template (e.g. during setup stage with the support of custom environment.yml files).

[atemate]

For now, it's not clear what tech risks we take once we decide to transition from pip to conda. For that, we need to understand clearly the following aspects:

1. the problem: why do we need to go away from pip?
2. the final result: what we will get at the end of transition from pip to conda
3. is there anything that can not be installed via conda but can via pip?
4. take a look at other widely used DOckerfiles, to research why they use pip or conda? It might happen, that conda behaves unstable on docker.

Deep refactoring of base image is a big work, so we need to understand the problem and the risks before we start. For now, as I see, there's no problem but only some bearable inconveniences.