docker-compose.yaml

# This version hides the underlying postgres database on a backend network
# different from the frontend network on which nginx interfaces mlflow.
# And mlflow is hidden behind nginx, allowing user auth to be implemented.
# 
# The following environment vars can be set in the shell before running
# docker-compose (default values are these; at minimum change DB_PW).
# export DB_NAME=mlflow
# export DB_USER=postgres
# export DB_SERVER=db  # ie defaults to db container; can replace with ip address
# export PGPASS=~/.pgpass  # path of .pgpass file (contains pw)
# export DB_PORT=5432
# export FILESTORE=mlruns_vol  # ie defaults to docker volume; can replace with filesys dir
# export MLFLOW_PORT=5000
# Those defaults are set automatically so you only need specify the ones you
# want to change - eg a new DB_PW value should be set but the rest are optional.
#
# Note an AWS S3 bucket can be used instead of local drive for the artifacts
# store, via the commented-out environment lines below.

version: '3.3'

services:
    db:
        restart: always
        image: postgres:13
        container_name: mlflow_db
        expose:
            - ${DB_PORT:-5432}
        # networks:
        #     - backend
        environment:
            # - MUID=$UID
            #  - MGID=$GID
            - POSTGRES_DB=${DB_NAME:-mlflow}
            - POSTGRES_USER=${DB_USER:-postgres}
            - POSTGRES_PASSWORD_FILE=/run/secrets/pg_admin_pw
        secrets:
            - pg_admin_pw
        volumes:
            - datapg_vol:/var/lib/postgresql/data

    app:
        restart: always
        build: ./mlflow
        image: mlflow_server
        container_name: mlflow_server
        expose:
            - 5001
        # networks:
        #     - frontend
        #     - backend
        environment:
          - BACKEND=postgresql://${DB_USER:-postgres}@${DB_SERVER:-db}:${DB_PORT:-5432}/${DB_NAME:-mlflow}
          - ARTIFACTS=/storage/mlruns  # in-container path to filestore in filesys
          # For artifact store in AWS S3 (uses boto that was installed in container):
          # Commment out ARTIFACTS line above and instead use:
          #  - ARTIFACTS="s3://mlflow_bucket/my_mlflow_dir/"
          #  - AWS_ACCESS_KEY_ID=${AWS_ACCESS_KEY_ID}
          #  - AWS_SECRET_ACCESS_KEY=${AWS_SECRET_ACCESS_KEY}
          #  - AWS_DEFAULT_REGION=${AWS_DEFAULT_REGION}
        volumes:
            - ${FILESTORE:-/storage/mlruns}:/storage/mlruns  # can comment out this line if using S3
            - ${PGPASS:-~/.pgpass}:/root/.pgpass  # provides the pw for BACKEND database
            - condaenv_vol:/opt/conda  # provides continuity/speed when looping runs with same container
        command: 
            - sh    # (sh allows for var substitution of BACKEND and ARTIFACTS)
            - -c
            - mlflow server 
                --port 5001
                --host 0.0.0.0 
                --backend-store-uri $${BACKEND} 
                --default-artifact-root $${ARTIFACTS}
        # depends_on:
        #     - db

    nginx:
        restart: always
        build: ./nginx
        image: mlflow_nginx
        container_name: mlflow_nginx
        ports:
            - "${MLFLOW_PORT:-5000}:80"
        # networks:
        #     - frontend
        depends_on:
            - app

# networks:
#     frontend:
#         driver: bridge
#     backend:
#         driver: bridge

secrets:
    pg_admin_pw:
        file: ~/.pgadminpw

volumes:
    mlruns_vol:
    datapg_vol:
    condaenv_vol: