A tool to generate Dockerfiles from templates and scripts for stacks. Be careful, there are rough edges!

Install poetry(https://python-poetry.org)

curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/get-poetry.py | python

To install stacker clone the repo and:

cd stacker
poetry install 

Now stacker has been installed as a commandline tool.

• Generate dockerfiles

stacker --generate

This command will generate all possible dockerfiles for all OSes and frameworks as defined in the config file.

• stacker/configs/dlrs.yaml

A loosely defined specification file

---
version: "0.6.0"
description: "Deep Learning Reference Stack"
license: |
  # SPDX-License-Identifier: MIT
  # Copyright (c) 2019 Intel Corporation

stack:
  dlrs:
    version: "0.6.0"
    tag: dlrs
    ubuntu:
      version: "18.10"
      tag: ubuntu
      os_pkgs:
        - openssh-server
        - openssh-client
        - wget
        - curl
        - python37
      python:

Similarly spec files for each stack can be defined, or even put into the one single file.

• stacker/slices/component.dockerfile

Each component is defined as dockerfiles templates

Eg. A tensorflow dockerfile cane be something like:

RUN pip install tensorflow=={{tf_version}}

Here {{tf_version}} is a variable that would be dynamically injected from the spec file value spec.stack.dlrs.ubuntu.tensorflow.version

Another template is for the OS

FROM {{os}}
RUN {{pkg_install}}  

Here, there are 2 placeholder variables {{os}} and {{pkg_install}} Again, OS and system packages would be injected dynamically based on the spec file

• parse spec file - DONE
• parse slices of dockerfiles per app - DONE
• update dockerfile based on the yaml file values - DONE
• enable elementary cli support - DONE

TODO

• give option to build dockerimages from generated dockerfiles
• convert dockerfiles to singularity recepies
• lint dockerfiles
• resize images

• Maintainable and upgradable solution
• Use yaml specification files to define configuaration values
• Use Dockerfiles as templates which can accept substitution from the spec files
• Limit use of shell scripts/custom templates use Dockerfiles where possible
• Use official Docker and Singularity clients where possible