Installation

The MLRun module has three independently switchable layers:

Layer	Default state	Required for
Trainer sidecar (`:8003`)	Enabled	Training, scoring, generated code
Docker Desktop Kubernetes	Off	Deployments tab, `deploy_to_k8s.py`
MLRun Community Edition	Off	Project store + UI for runs/artifacts

1. Prerequisites

Docker Desktop ≥ 4.30 with at least 6 CPU / 8 GB RAM allocated.
The Workbench2 backend (http://localhost:8001) and frontend (http://localhost:5270) running, with MongoDB reachable.
A Mongo master.bank_transactions collection populated for the spend-risk use-case, and master.bank_customer populated for the customer-personality use-case (both ship in the demo seeders).

2. Trainer sidecar

The trainer sidecar is a Python 3.11 FastAPI service that exposes:

POST /train — trains an sklearn / xgboost / lightgbm / pytorch model and returns metrics + model id.
POST /invocations — ecosystem-runtime compatible scoring endpoint backed by the freshly trained model.

Start it with the bundled compose file:


cd backend
./run_mlrun.sh up

The script wraps docker compose -f docker-compose.mlrun.yml up -d and publishes the trainer on http://localhost:8003.

3. Docker Desktop Kubernetes (optional)

Enable Kubernetes inside Docker Desktop, then bootstrap the ecosystem-workbench namespace and PV mounts:


cd backend
./scripts/setup-k8s.sh

In backend/.env, flip the K8s feature flag:


K8S_ENABLED=true
K8S_CONTEXT=docker-desktop
K8S_NAMESPACE=ecosystem-workbench
K8S_RUNTIME_URL=http://localhost:30091

Restart the backend. The MLRun console now shows a Deployments tab, and the Deploy to Kubernetes action becomes available on every training row.

4. MLRun Community Edition (optional)

When you want a full MLRun project store + UI on top of the trainer sidecar:


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

The script wraps helm upgrade --install mlrun-ce mlrun/mlrun-ce with the values in backend/k8s/mlrun-ce-values.yaml. NodePorts:

Service	URL
MLRun API	`http://localhost:30070`
MLRun UI	`http://localhost:30060`
MinIO	`http://localhost:30090`

In backend/.env, flip the CE flags:


MLRUN_CE_ENABLED=true
MLRUN_CE_NAMESPACE=mlrun
MLRUN_CE_RELEASE=mlrun-ce
MLRUN_CE_API_URL=http://localhost:30070

See MLRun Community Edition for the complete coexistence model.

5. Seed the two reference use-cases


cd backend
./venv/bin/python scripts/seed_mlrun_use_cases.py

Useful flags:

Flag	Effect
`--use-case spend_risk`	Seed only the spend-risk lifecycle.
`--use-case customer_personality`	Seed only the personality lifecycle.
`--reset`	Cascade-delete prior seed rows before seeding.
`--dry-run`	Validate preflight without writing to MongoDB or training.
`--skip-training`	Seed metadata only; skip the live trainer calls.
`--skip-k8s`	Skip the optional Kubernetes deploy step.

Successful execution writes a SEED_MLRUN_USE_CASE row to ecosystem_meta.activities per use-case with the framework list, succeeded count, and elapsed seconds.

6. Verify in the console

Open http://localhost:5270/mlrun-console. Both Customer Spend Risk and Customer Personality rows should be visible in the Configurations table, and the Training Runs tab should list four successful runs per use-case.

MLRun console — training runs