MLflow¶

Open-source platform for managing the end-to-end machine learning lifecycle, including experiment tracking, model registry, and deployment. Provides a centralized tracking server for logging metrics, parameters, and artifacts.


Category	observability
Official Docs	MLflow Documentation
CLI Install	`./cli o11y mlflow install`
CLI Uninstall	`./cli o11y mlflow uninstall`
Namespace	`mlflow`

Overview¶

MLflow provides comprehensive ML lifecycle management: - Experiment Tracking: Log parameters, metrics, and artifacts - Model Registry: Version and stage ML models - Model Serving: Deploy models as REST endpoints - S3 Artifact Storage: Uses Terraform-provisioned S3 bucket - Web UI: Visual experiment comparison and analysis

Installation¶

Prerequisites¶

Configure in .env:

MLFLOW_USERNAME=admin
MLFLOW_PASSWORD=your-password

Install¶

./cli o11y mlflow install

The installer: 1. Provisions an S3 bucket via Terraform for artifact storage 2. Adds the community Helm chart repository 3. Renders Helm values with credentials and bucket configuration 4. Deploys MLflow to the mlflow namespace

Verification¶

# Check pods
kubectl get pods -n mlflow

# Check service
kubectl get svc -n mlflow

# Port-forward for UI access
kubectl port-forward svc/mlflow 5000:5000 -n mlflow --address 0.0.0.0 &

# Access MLflow UI
open http://localhost:5000

Log in with MLFLOW_USERNAME and MLFLOW_PASSWORD.

Configuration¶

Environment Variables¶

Variable	Description
`MLFLOW_USERNAME`	Admin username
`MLFLOW_PASSWORD`	Admin password
`DOMAIN`	Domain for ingress (optional)

Infrastructure¶

MLflow uses Terraform to provision: - S3 bucket for artifact storage (models, datasets, logs) - IAM roles via IRSA for S3 access

Usage¶

Python SDK¶

import mlflow

mlflow.set_tracking_uri("http://mlflow.mlflow:5000")
mlflow.set_experiment("my-experiment")

with mlflow.start_run():
    mlflow.log_param("model", "qwen3-30b-instruct-fp8")
    mlflow.log_metric("accuracy", 0.95)
    mlflow.log_artifact("output.json")

Model Registry¶

# Register a model
mlflow.register_model("runs:/<run-id>/model", "my-model")

# Load a registered model
model = mlflow.pyfunc.load_model("models:/my-model/Production")

Troubleshooting¶

UI not loading¶

# Check pod status
kubectl get pods -n mlflow

# Check logs
kubectl logs -n mlflow -l app=mlflow

Artifact upload fails¶

# Check S3 bucket access
kubectl exec -it -n mlflow <mlflow-pod> -- aws s3 ls s3://<bucket-name>/

# Check IRSA configuration
kubectl describe sa -n mlflow mlflow