MLflow Tracking Stack

Docker Compose stack for a local MLflow tracking server: MLflow, MinIO, Ofelia, and a choice of PostgreSQL, MySQL, MSSQL, or SQLite as the backend store. Requires Docker with Compose.

Services

The stack combines a base file with a database overlay.

Base (compose.yml)

Service	Role
minio	S3-compatible artifact store.
minio_client	One-shot bootstrap: creates the mlflow bucket in MinIO.
server	mlflow server connected to MinIO and the chosen database.
ofelia	Scheduler that runs mlflow gc and backup jobs via Docker labels.

Backend overlays

Overlay	Services	Role
compose.postgres.yml	backend, backend_backup	PostgreSQL 17 backend with pg_dump backups.
compose.mysql.yml	backend, backend_backup	MySQL 8.4 backend with mysqldump backups.
compose.mssql.yml	backend, backend_backup	SQL Server 2022 Dev backend with .bak backups.
compose.sqlite.yml	backend_backup	SQLite file-volume backend with file-copy backups.

Services start in dependency order; health checks gate each stage.

Environment variables

Copy .env.example to .env and adjust before first run. Change all secrets before any shared or production use.

Name	Purpose	Default
AWS_ACCESS_KEY_ID	MinIO root user / S3 access key for MLflow.	minio_key
AWS_SECRET_ACCESS_KEY	MinIO root password / S3 secret for MLflow.	minio_secret
MINIO_PORT	MinIO S3 API port.	9000
MINIO_CONSOLE_PORT	MinIO web console port.	9090
DB_PORT	Database port inside the stack.	5432
DB_NAME	MLflow metadata database name.	mlflow_database
DB_USER	MLflow database user.	mlflow_user
DB_PASSWORD	Password for DB_USER.	mlflow
MLFLOW_BACKEND_STORE_URI	SQLAlchemy URI for mlflow server --backend-store-uri.	Depends on the database
DEFAULT_ARTIFACT_ROOT	Artifact root; must match the bucket created by minio_client.	s3://mlflow/
MLFLOW_SERVER_PORT	Host port for the MLflow UI/API.	5000

Run

Pick a database backend and pass both compose files:

# PostgreSQL
docker compose -f compose.yml -f compose.postgres.yml up -d

# MySQL
docker compose -f compose.yml -f compose.mysql.yml up -d

# MSSQL
docker compose -f compose.yml -f compose.mssql.yml up -d

# SQLite
docker compose -f compose.yml -f compose.sqlite.yml up -d

Open in the browser (ports from .env):

• MLflow UI: http://localhost:5000
• MinIO console: http://localhost:9090 — log in with AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY

Stop with the same -f flags:

docker compose -f compose.yml -f compose.postgres.yml down

Use MLflow from Python

import mlflow

mlflow.set_tracking_uri("http://localhost:5000")

Logging artifacts also requires direct MinIO access:

import os

os.environ["MLFLOW_TRACKING_URI"] = "http://localhost:5000"
os.environ["MLFLOW_S3_ENDPOINT_URL"] = "http://127.0.0.1:9000"
os.environ["AWS_ACCESS_KEY_ID"] = "minio_key"
os.environ["AWS_SECRET_ACCESS_KEY"] = "minio_secret"

If MinIO rejects virtual-hosted-style requests, also set AWS_S3_ADDRESSING_STYLE=path.

From another container on the same Compose network: use http://server:5000 and http://minio:9000.

Notes

• MinIO and database volumes persist across restarts.
• Ofelia runs mlflow gc every 24 hours to purge deleted runs.
• Database backups run every 12 hours. Use restore_backup.sh inside backend_backup to restore a dump.