A7 - environments

environment

defines the software configuration used for jobs and deployments

environments can be versioned
there are two useful categories:
- curated
- custom

curated environment

a prebuilt environment maintained/provided for common ML workloads

advantages:
- less setup
- common frameworks already configured
- faster to get started
- often optimized for Azure ML scenarios

custom environment

defined by the user by specifying base image, Conda/Python dependencies, and possibly Docker build configuration

Conda is a common environment specification

name: training-env

dependencies:
  - python=3.11
  - pip
  - pip:
      - numpy
      - pandas
      - scikit-learn
      - mlflow

Conda dependencies can be one layer within the overall runtime environment

CONTAINER IMAGE
├── Linux
├── system libraries
├── maybe CUDA
└── Conda/Python environment
    ├── Python
    ├── numpy
    ├── sklearn
    └── MLflow

one can also use their own Docker image, or the environment based on the container image
this gives maximum control, but also requires more maintenance and responsibility
environment categories have a trade-off between convenience and control
pinning versions for dependencies is also needed for stronger reproducibility
building environments are costly and time consuming, so avoid changing definitions every run
for GPU compute, GPU-enabled software is needed
curated environments reduce the burden of configuring each of this