Kubernetes Authentication

MLflow includes built-in request auth providers for Kubernetes environments. These providers automatically add authorization headers to outgoing MLflow client requests using Kubernetes credentials.

:::important Applicability

This provider is designed for MLflow deployments behind a proxy with Kubernetes-based authentication, such as kube-rbac-proxy, or a custom auth MLflow plugin.

If your MLflow deployment does not use Kubernetes-based authentication, this provider is not applicable.

:::

Overview

Two auth providers are available, selected via MLFLOW_TRACKING_AUTH:

The kubernetes provider adds only a bearer token. The kubernetes-namespaced provider also adds an X-MLFLOW-WORKSPACE header derived from the Kubernetes namespace.

Prerequisites

Install the kubernetes Python package:

pip install mlflow[kubernetes]

Configuration

For token-only authentication:

export MLFLOW_TRACKING_AUTH=kubernetes

To also attach workspace headers (namespace-based workspace routing):

export MLFLOW_TRACKING_AUTH=kubernetes-namespaced

No additional configuration is needed. The provider automatically discovers credentials from the environment.

Credential Sources

The provider tries two credential sources in order for each piece of information, using the first one that succeeds. The namespace and token are resolved independently and may come from different sources.

1. Service Account (in-cluster)

When running inside a Kubernetes pod, the provider reads the mounted service account files:

• Namespace: /var/run/secrets/kubernetes.io/serviceaccount/namespace
• Token: /var/run/secrets/kubernetes.io/serviceaccount/token

This is the default path for pods with a service account mounted (which is the standard Kubernetes behavior).

2. Kubeconfig

When service account files are not available, the provider falls back to kubeconfig:

• Namespace: Extracted from the active context
• Token: Resolved via the Kubernetes Python client's ApiClient, which handles exec-based auth flows (EKS, GKE, AKS, OpenShift, OIDC)

Usage

Once configured, no code changes are needed. All MLflow client calls will automatically include the authorization headers with values sourced from Kubernetes:

import mlflow

mlflow.set_tracking_uri("http://mlflow-server:5000")

with mlflow.start_run():
    mlflow.log_param("learning_rate", 0.01)
    mlflow.log_metric("accuracy", 0.95)

If a workspace is explicitly set (e.g., via mlflow.set_workspace()), it takes priority over the Kubernetes namespace. The namespace from Kubernetes credentials is used as a default when no workspace is specified.