MLOps

Simple open source MLOps stack with docker compose.

Design

Quickstart

This setup requires docker and docker compose. The easiest way to get started is to install docker desktop.

Create a .env file with the following keys. Values below are only placeholders and can be freely changed.

POSTGRES_DATABASE=postgres
POSTGRES_HOST=postgres
POSTGRES_PASSWORD=password
POSTGRES_PORT=5432
POSTGRES_USER=postgres

MINIO_ROOT_USER=minio
MINIO_ROOT_PASSWORD=password

Now run the following command:

docker compose up -d

Once the applications have started and are ready, you can access them on the following urls:

• MLflow: http://localhost:5001
• MinIO: http://localhost:9001
• Qdrant: http://localhost:6333/dashboard
• Argilla: http://localhost:6900

Applications

• PostgreSQL (SQL)
• MinIO (object storage)
• Qdrant (vector search)
• Argilla (data annotation)
• MLflow (experiment tracking & model registry)

Updating images

The following commands will stop the containers, update images if new versions have been published and redeploy containers.

docker compose stop
docker compose pull
docker compose up -d

Clunky aspects

• MLflow: no official image containing psycopg2. Workaround of running pip install psycopg2 prior to launching server. (Not enough benefits to create custom image).
• postgreSQL: Setting up databases per application. Mounting init-db.sql to docker-entrypoint-initdb.d folder which is automatically run only if postgres volume is empty.