Introduction

The Nebari MLflow Pack deploys MLflow on a Nebari cluster: experiment tracking and a model registry behind Keycloak SSO, with a PostgreSQL backend store and TLS provisioned for you.

It wraps the community MLflow chart and adds three things that chart cannot know about: a NebariApp for routing and auth, a MLFLOW_SERVER_ALLOWED_HOSTS secret derived from that hostname, and defaults chosen so the combination actually works.

  browser ──► Envoy Gateway ──► MLflow ──► PostgreSQL
                   │          svc :80         (bundled)
              OIDC filter    pod :5000
                   │
               Keycloak

  JupyterHub notebooks ──────► MLflow  (in-cluster, no auth hop)
    MLFLOW_TRACKING_URI       svc :80

Two ways in, deliberately. Humans arrive through the gateway and authenticate against Keycloak; notebooks talk to the ClusterIP service directly, bypassing the OIDC filter because the MLflow Python client has no way through it. See Connecting JupyterHub.

What ships

• MLflow 3.7.0, served by uvicorn so the 3.x security-middleware flags are available.
• PostgreSQL 17.5 as the backend store, with an 8Gi PVC and schema migration on startup. On by default — the alternative is SQLite that loses every experiment on pod restart.
• A NebariApp producing an HTTPRoute, a cert-manager certificate, a Keycloak client, and an Envoy SecurityPolicy enforcing OIDC at the gateway.
• An allowed-hosts Secret computed from the NebariApp hostname plus the in-cluster service DNS name, injected via envFrom.

Know this before you deploy

Two defaults surprise people, and both are covered in full below.

• Artifacts are ephemeral out of the box. The backend store is durable; the artifact store defaults to a path inside the pod’s filesystem, so logged models and files vanish on restart. See Artifact storage.
• The PostgreSQL secret must exist first, and its name must be exactly <release-name>-postgresql. See Getting started.

In this guide

• Getting started — the credentials secret, install, first check
• Deploying on Nebari — the Argo CD Application, DNS, and certificates
• Connecting JupyterHub — tracking URI, the NetworkPolicy port trap, and verifying from a notebook
• Standalone deployment — running without Nebari, and locally

Guides

• PostgreSQL backend — the secret contract, sizing, and the SQLite fallback
• Artifact storage — why the default is ephemeral, and how to point it at a bucket
• Allowed hosts — MLflow’s security middleware and the Host header
• Troubleshooting — symptoms, causes, and the commands that distinguish them

Reference

• Configuration — every value this chart owns
• Authentication flow — the OIDC handshake, cookie format, and JWT claims
• NebariApp CRD — field-by-field reference