Skip to content

Integrate Run:AI with MLflow

MLflow is an open-source platform to manage the ML lifecycle, including experimentation, reproducibility, deployment, and a central model registry. The purpose of this document is to explain how to run Jobs with MLflow using the Run:AI scheduler.


MLflow concepts and alternative architectures are discussed here. MLflow can run on various platforms. To work with Run:AI we would use the MLflow Kubernetes integration.

The MLflow documentation describes the Kubernetes integration as such:


When you run an MLflow Project on Kubernetes, MLflow constructs a new Docker image containing the Project’s contents; this image inherits from the Project’s Docker environment. MLflow then pushes the new Project image to your specified Docker registry and starts a Kubernetes Job on your specified Kubernetes cluster. This Kubernetes Job downloads the Project image and starts a corresponding Docker container. Finally, the container invokes your Project’s entry point, logging parameters, tags, metrics, and artifacts to your MLflow tracking server.

To run an MLflow job via Kubernetes, you specify an MLflow Kubernetes configuration file that contains a template. Here is an example from the MLflow documentation:

  "kube-context": ...,
  "repository-uri": ...,
  "kube-job-template-path": "/username/path/to/kubernetes_job_template.yaml"

The essence of the Run:AI integration is the modification of the kubernetes_job_template.yaml file. Specifically adding the Run:AI scheduler name and the Run:AI Project (Kubernetes namespace).

Step by Step Instructions


  • Install MLflow.
  • Make sure you have push access to a Docker repository from your local machine.
  • Make sure you are connected to Run:AI via the Run:AI Command-line interface.

The sample MLflow Project

The relevant sample files are here. These contain:

  • A Dockerfile. This file builds a base docker image containing python3 and the required mlflow dependencies. The Docker file is already compiled and available at
  • An MLflow project file MLproject. The project file contains the base image above as well as the python command-line to run.
  • The training python code
  • MLflow Kubernetes configuration files as in the MLflow documentation.
    • Kubernetes configuration file kubernetes_config.json
    • An MLflow Kubernetes Job template kubernetes_job_template.yaml


  • Edit kubernetes_config.json.
    • Set kube-context to the name of the Kubernetes context. You can find the context name by running runai list clusters or kubectl config get-contexts.
    • Set repository-uri to a repository and name of a docker image that will be used by MLflow (this is a different image than the base docker image described above). Your local machine needs permissions to be able to push this image to the docker registry.
  • Edit kubernetes_job_template.yaml.
    • Set the value of namespace to runai-<name of Run:AI project>.
    • Note the last line which adds the Run:AI scheduler to the configuration.
    • Do not change the lines marked by {replaced with....
    • Set the requested resources including GPUs. You can use the --dry-run flag of the runai submit command to gain insight on additional configurations


  • Perform docker login if required.
  • Run:
mlflow run mlproject -P alpha=5.0  -P l1-ratio=0.1  \
    --backend kubernetes --backend-config kubernetes_config.json

MLflow Tracking

The sample training code above does not contain references to an MLflow tracking server. This has been done to simplify the required setup. With MLflow-Kubernetes you will need a remote server architecture. Once you have such an architecture set up, you can use MLflow tracking in your code.

Using Interactive Workloads

With Run:AI you can also run interactive workloads. To run the Job as interactive, add the following to kubernetes_job_template.yaml:

    priorityClassName: "build"

See Also

Last update: December 9, 2021