mlflow

MLflow

MLflow is an open-source AI and ML platform for experiment tracking, model registry, artifact storage, and model lifecycle workflows.

What this template deploys

This Akash SDL deploys:

• MLflow Tracking Server
• PostgreSQL backend store
• MinIO S3-compatible artifact store
• Caddy reverse proxy
• Persistent storage for PostgreSQL, MinIO, and MLflow working data

Public port:

• 80 — MLflow UI and Tracking API

Internal services:

• mlflow:5000
• db:5432
• minio:9000

Architecture

MLflow uses PostgreSQL as the backend store for experiments, runs, metrics, params, tags, traces, and model registry metadata.

MinIO is used as an S3-compatible artifact store for run artifacts, models, plots, and other files.

The MLflow server runs with:

mlflow server \
  --backend-store-uri postgresql+psycopg2://mlflow:<password>@db:5432/mlflow \
  --serve-artifacts \
  --artifacts-destination s3://mlflow-artifacts

With --serve-artifacts, MLflow clients can upload artifacts through the MLflow Tracking Server instead of connecting directly to MinIO.

Before deployment

Change these placeholder values in deploy.yaml before deploying:

CHANGE_ME_POSTGRES_PASSWORD
CHANGE_ME_MINIO_ACCESS_KEY
CHANGE_ME_MINIO_SECRET_KEY

Generate strong values, for example:

openssl rand -base64 32

Use the same MinIO values in both places:

AWS_ACCESS_KEY_ID=CHANGE_ME_MINIO_ACCESS_KEY
AWS_SECRET_ACCESS_KEY=CHANGE_ME_MINIO_SECRET_KEY
MINIO_ROOT_USER=CHANGE_ME_MINIO_ACCESS_KEY
MINIO_ROOT_PASSWORD=CHANGE_ME_MINIO_SECRET_KEY

Storage

Default persistent storage sizes:

PostgreSQL data: 20Gi
MinIO artifacts: 50Gi
MLflow working data: 10Gi

The MinIO setup container creates this private bucket:

mlflow-artifacts

Artifacts are stored at:

s3://mlflow-artifacts

Using MLflow

After deployment, open the public Akash URI in your browser.

To connect from Python:

import mlflow

mlflow.set_tracking_uri("http://<akash-host>")
mlflow.set_experiment("akash-test")

with mlflow.start_run():
    mlflow.log_param("provider", "akash")
    mlflow.log_metric("score", 0.95)

To log a file artifact:

import mlflow

mlflow.set_tracking_uri("http://<akash-host>")

with mlflow.start_run():
    with open("example.txt", "w") as f:
        f.write("hello from Akash")
    mlflow.log_artifact("example.txt")

Notes

The MLflow container installs psycopg2-binary and boto3 on startup so the official MLflow image can connect to PostgreSQL and MinIO.

The MinIO API and console are not exposed publicly in this template. MLflow is the public entrypoint.

For production deployments, consider adding authentication or placing MLflow behind a secured ingress/proxy.