MLRun Community Edition

The Workbench2 MLRun module ships with an opt-in Helm install of the official MLRun Community Edition (CE) chart. CE provides the MLRun project store, Web UI, and MinIO-backed artifact bucket; the custom trainer sidecar continues to handle compute. This is the coexistence model adopted in ADR 0006 §Phase 5 .

Why coexistence (not replacement)?

Capability	Trainer sidecar (this module)	MLRun CE
sklearn / xgboost / lightgbm / pytorch tabular	✅ first-class	❌ requires custom job spec
Reproducible Python generator	✅	❌
MongoDB-native feature pipelines	✅	indirect
`ecosystem-runtime` adapter logging	✅	indirect
Project store / Web UI	❌	✅
MinIO artifact store	❌	✅
K8s scheduling for ad-hoc jobs	indirect	✅

When CE is enabled, both systems run side-by-side: the workbench backend prefers the CE API URL for project lookups (Settings.select_mlrun_api_url), while the trainer sidecar continues to receive every POST /train and POST /invocations call.

Installation


cd backend
./scripts/setup-mlrun-ce.sh

The script:

Verifies kubectl and helm are present.
Switches the Kube context to docker-desktop.
Creates the mlrun namespace.
Resolves MLRUN_CE_DBPATH (default LOCAL_DATA_PATH/mlrun-ce) and writes a values overlay merged with backend/k8s/mlrun-ce-values.yaml.
Adds the mlrun-ce Helm repo (idempotent).
Runs helm upgrade --install mlrun-ce mlrun/mlrun-ce.
Waits for the MLRun API pod to become Ready.

Service URLs (NodePort-based):

Service	URL	Purpose
MLRun API	`http://localhost:30070`	Programmatic project / run access
MLRun UI	`http://localhost:30060`	Web UI
MinIO	`http://localhost:30090`	Artifact bucket

Settings

Workbench2 picks up CE through the following backend/.env keys:


MLRUN_CE_ENABLED=true
MLRUN_CE_NAMESPACE=mlrun
MLRUN_CE_RELEASE=mlrun-ce
MLRUN_CE_API_URL=http://localhost:30070
MLRUN_CE_UI_URL=http://localhost:30060
MLRUN_CE_MINIO_URL=http://localhost:30090
# Optional persistence overrides (default: $LOCAL_DATA_PATH/mlrun-ce)
# MLRUN_CE_DBPATH=/Users/me/data/mlrun-ce
# MLRUN_CE_ARTIFACT_PATH=/Users/me/data/mlrun-ce/artifacts

Settings.select_mlrun_api_url() returns the CE URL when CE is enabled and non-blank, otherwise falls back to MLRUN_SIDECAR_URL, then to MLRUN_TRAINER_URL.

Teardown


cd backend
./scripts/teardown-mlrun-ce.sh

The script uninstalls the Helm release and removes the mlrun namespace. Persistent host paths are preserved so the same DBPATH can be reused across reinstalls — delete them manually if you want a clean slate.

Operational notes

Nuclio / Jupyter / Pipelines disabled: the bundled values file disables Nuclio, Jupyter, and Kubeflow Pipelines to keep the local install lightweight. Re-enable them in backend/k8s/mlrun-ce-values.yaml if you need them.
arm64 only: the local install is opinionated for Apple Silicon. On Intel hardware update the image tags before running the script.
CE storage size: MinIO is given a 10 GiB hostPath PVC; raise it for non-trivial artifact stores.

A live screenshot of the MLRun CE Web UI will be added once the Helm install has been run via ./scripts/setup-mlrun-ce.sh. The default deployment is opt-in (MLRUN_CE_ENABLED=false) so this screenshot is captured separately from the seed lifecycle.