A Minimal Guide to Deploying MLflow 2.6 on Kubernetes
Introduction
Deploying MLflow on Kubernetes can be a straightforward process if you know what you're doing. This blog post aims to provide a minimal guide to get you up and running with MLflow 2.6 on a Kubernetes cluster. We'll use the namespace my-space for this example.
Prerequisites
A running Kubernetes cluster
kubectl installed and configured to interact with your cluster
Step 1: Create the Deployment YAML
Create a file named mlflow-minimal-deployment.yaml and paste the following content:
apiVersion: v1 kind: Namespace metadata: name: my-space --- apiVersion: apps/v1 kind: Deployment metadata: name: mlflow-server namespace: my-space spec: replicas: 1 selector: matchLabels: app: mlflow-server template: metadata: labels: app: mlflow-server name: mlflow-server-pod spec: containers: - name: mlflow-server image: ghcr.io/mlflow/mlflow:v2.6.0 command: ["mlflow", "server"] args: ["--host", "0.0.0.0", "--port", "5000"] ports: - containerPort: 5000 ---
apiVersion: v1 kind: Service metadata: name: mlflow-service namespace: my-space spec: selector: app: mlflow-server ports: - protocol: TCP port: 5000 targetPort: 5000
Step 2: Apply the Deployment
Apply the YAML file to create the deployment and service:
kubectl apply -f mlflow-minimal-deployment.yaml
Step 3: Verify the Deployment
Check if the pod is running:
kubectl get pods -n my-space
Step 4: Port Forwarding
To access the MLflow server from your local machine, you can use Kubernetes port forwarding:
kubectl port-forward -n my-space mlflow-server-pod 5000:5000
After running this command, you should be able to access the MLflow server at http://localhost:5000 from your web browser.
Step 5: Access MLflow within the Cluster
The cluster-internal URL for the MLflow service would be:
http://mlflow-service.my-space.svc.cluster.local:5000
You can use this tracking URL in other services within the same Kubernetes cluster, such as Kubeflow, to log your runs.
Troubleshooting Tips
Pod not starting: Check the logs using kubectl logs -n my-space mlflow-server-pod.
Service not accessible: Make sure the service is running using kubectl get svc -n my-space.
Port issues: Ensure that the port 5000 is not being used by another service in the same namespace.
Conclusion
Deploying MLflow 2.6 on Kubernetes doesn't have to be complicated. This guide provides a minimal setup to get you started. Feel free to expand upon this for your specific use-cases.