Joel Baxter & Kartik Mathur & Don Wake

Building Dynamic Machine Learning Pipelines with KubeDirector

August 14, 2020

Editor’s Note – HPE Ezmeral Container Platform is now HPE Ezmeral Runtime Enterprise. For more information on why the name was changed, please click here.

Imagine that you’re a data scientist who’s been asked to create an application or service that can predict travel time for a proposed taxi ride. The application needs to have the ability to update the service with new data as it becomes available, so its predictions take recent patterns into account. In this tutorial, we’ll show you how to set up a Machine Learning (ML) pipeline using KubeDirector to train, register, and query your model.

KubeDirector was introduced to the open source community to address stateful application deployment in standard Kubernetes clusters. In the latest release (version 0.5), KubeDirector now allows multiple clusters to share data very easily using a new feature called Connections. The new feature helps users create large-scale dynamic, stateful containerized applications such as are found in Machine Learning (ML) pipelines and allow them to constantly evolve as models are improved and the data changes.

A basic ML pipeline consists of three stages:

  • Training, where the input datasets are processed to create ML models
  • Model registration, where the models to be used are identified and characterized
  • Inferencing, where selected models are made available for answering queries

After a brief description of the KubeDirector architecture used for this solution, we’ll show you how to train, register, and finally query your model for answers. The sample dataset used, and available here under trip data, contains taxi ride data that will be used to infer total time per ride for given taxi pickup and drop-off locations. And yes, this is pre-2020 data. Things look much different now due to the pandemic. That’s why it’s important that your ML model is flexible and can take real-time data into account.

Our solution uses three KubeDirector Applications (kdapps): A training deployment kdapp named training-engine, a Jupyter Notebook kdapp named jupyter-notebook, and an inferencing deployment kdapp named deployment-engine.

A kdapp is a custom resource (CR) that can be created in any Kubernetes cluster. The kdapp instructs KubeDirector on how a particular kind of virtual application cluster should be deployed and managed. The three kdapps used in this solution are examples that can be found online in the KubeDirector github example catalog.

Register Your KubeDirector Applications with Kubernetes

Assuming that KubeDirector is already deployed and running in the Kubernetes cluster, these kdapp CRs can be created to register apps with KubeDirector, e.g. by using kubectl:

kubectl create –f cr-app-training-engine.json
kubectl create –f cr-app-jupyter-notebook.json
kubectl create –f cr-app-deployment-engine.json

This is a one-time step that builds a catalog of applications that data scientists can now instantiate as needed.

Once a kdapp is created, an instance of that app can be deployed by creating a KubeDirector virtual cluster (kdcluster) CR. A kdcluster identifies the desired kdapp and specifies runtime configuration parameters, such as the size and resource requirements of the virtual cluster.

KDApp Connections

For the purposes of this discussion, one of the most interesting parts of the kdcluster spec is the Connections stanza, which identifies other resources of interest to that kdcluster. This is detailed in the GitHub readme link here.

The KubeDirector GitHub repo includes examples of kdcluster CRs that instantiate these kdapps. This tutorial will show you how to create kdclusters much like those examples, except that yours will add information about the Connections that each kdcluster will use.

The input data and trained ML models for this example pipeline will live in a project repository of shared persistent storage. The kdapps used will expect to be able to access that repository through directories mounted within the app containers. There are multiple ways to implement this arrangement, but here you’ll use the FS Mounts feature of the HPE Ezmeral Container Platform to access our pre-integrated HPE Ezmeral Data Fabric (formerly known as MapR Data Platform) for persistent container storage. The final piece of the puzzle will be the use of ConfigMap resources to describe and locate the ML models used for drawing inferences.

These components come together to build the pipeline like so:

mlops figure 1

A project repository is used to store the key data components needed in any ML pipeline, such as source data, the model itself and the scoring script.

In this example, the project repository is maintained within the HPE Ezmeral Data Fabric, since KubeDirector was itself deployed using the HPE Ezmeral Container Platform (which includes our data fabric by default); KubeDirector is a key component of the HPE Ezmeral Container Platform. However, KubeDirector is an open-source project and it can be used with any Kubernetes cluster and corresponding project repository.


Once you have registered the ML training kdapp and the Jupyter Notebook kdapp for Model Building, you’ll need to launch instances of those applications as KubeDirector clusters (kdclusters) to put the ML pipeline to work. First, launch an instance of the ML training kdapp:

kubectl create –f cr-cluster-training-engine.yaml

KubeDirector Connection creation: Jupyter Notebook kdcluster -> Training engine kdcluster

Next you’ll need to launch an instance of your Jupyter Notebook kdcluster. Before you do that, you need to modify the example kdcluster yaml file and include a new Connection stanza into the top level “spec” section. This modification would look similar to the following code:

  app: “jupyter-notebook”
  appCatalog: “local”
      -  “training-engine-instance”

Now, use the modified yaml file to launch an instance of your Jupyter Notebook kdcluster:

kubectl create –f cr-cluster-jupyter-notebook.yaml

The runtime configuration of the connected training-engine kdcluster, including its service endpoints, is therefore injected by KubeDirector into the jupyter-notebook app. This info is used by the notebook app when KubeDirector triggers configuration scripts to run within the app containers. >Note that these Connections can also be edited in the kdcluster spec later, for example if the notebook should be redirected to work with a different training engine.

Training the model

Now that the notebook kdcluster is running, a user accesses the Jupyter Notebook web UI via the network service port. Then, the usual process of directing the connected training deployment to generate models from the input data can be performed.

Using a little training magic

Those familiar with Jupyter Notebooks know that they include support for predefined magic functions that can be called within a notebook using a command line style syntax. These magic functions expand the capabilities of the notebook adding support for things like: %history, %edit, %rerun, %recall, %macro, %save, and %pastebin.

Jupyter Notebook also includes hooks that allow defining custom magics. The kdapps in our solution utilize these custom commands (magics) to seamlessly integrate with KubeDirector Connections. This includes custom magics to handle remotely submitting training code and retrieving results and logs. These magic functions make REST API calls to the API server that runs as part of a training environment.

For example the following magic functions are included:

  • %attachments: Returns a list of connected training environments
  • %logs --url: URL of the training server load balancer
  • %%<<trainingclustername>>: Submits training code to the training environment

The below screenshot from the GitHub site shows you an example of how a Jupyter Notebook can use these magic commands:

mlops figure 2

Note at the top of this screenshot of the Jupyter Notebook that the %attachments magic command is used to retrieve the Training Cluster name - traningengineinstance. Then, that name is used in the following magic command - %%trainingengineinstance. This single command then submits training code to the training environment – truly magic!

As the training engine generates models, it will store the model data into the project repository. From the training engine’s point of view, it is simply writing to a designated subdirectory of its filesystem.

Below is a snapshot from the HPE Ezmeral Container Platform Web User Interface. This illustrates the use of the integrated HPE Ezmeral Data Fabric. Specifically, the FsMounts feature is used here for organizing the data.

mlops figure 3

Model registration

Next, you’ll need to create a ConfigMap resource to store metadata about the model to be used in deployments. Here's an example ConfigMap for the model data generated above:

apiVersion: v1
kind: ConfigMap
  name: tensorflowmodel
  name: tensorflowmodel
  description: "example model"
  model-version: "1"
  path: /bd-fs-mnt/TenantShare/models/10yrdatasetchecknames/0_tf
  scoring-path: /bd-fs-mnt/TenantShare/code/

The two crucial parameters that will be used by the deployment-engine kdcluster are the path and the scoring-path, each of which identifies a location within the project repository as seen within an app container. The path is the location of the serialized model data. This path can be retrieved from the job log when the model was trained. The scoring-path locates a script that will be used by the deployment engine to deserialize and process the model.

Create inference deployment

This is what you’ve been waiting for! How long will my taxi ride take? You’ve trained a model from a data set, and now it’s time to extract value from your model. To do this, you need to create a deployment-engine.

A deployment-engine kdcluster is used to stand up services that will allow clients to draw ride-time inferences from the models you have created and registered.

A deployment engine will serve inferences for one or more models. To specify models to a deployment-engine kdcluster, the Connections feature once again comes in handy. Instead of naming other kdclusters of interest, this time it will be used to name ConfigMap resources, one for each model that the deployment should use. In this case, just use the one example model that was given.

KubeDirector Connection creation: Inference server kdcluster -> Model ConfigMap

For this deployment, the example cr-cluster-endpoint-wrapper.yaml file can be used. Similar to how the Jupyter Notebook kdcluster yaml file was modified, this kdcluster yaml file will be edited to include the Connection stanza. A new property in the top-level spec section is added, similar to the following:

  app: deployment-engine
  appCatalog: “local”
    -  “tensorflowmodel”

When these ConfigMap Connections are named in a kdcluster spec, KubeDirector will inject the contents of those ConfigMaps into a JSON file within the kdcluster's app containers where they can be used by the deployment app. If the contents of the ConfigMaps change, KubeDirector will immediately update this file. The deployment app can therefore reference this file to find the models it should use, which it can then load from the FS Mount directories exposed within its containers.

This key use of KubeDirector Applications, Clusters and Connections is what makes an ML pipeline created with KubeDirector very dynamic – it’s able to easily accept constantly updating and changing data sets and models of that data. The taxi ride data for New York City in 2020 will likely look very different than the sample pre-pandemic dataset. ML pipelines need to be constantly monitored and often retrained to accommodate this model drift, as the ML model lifecycle is a highly iterative and dynamic process.

A similar process could be implemented using ConfigMaps mounted into application pods using native Kubernetes mechanisms. An important benefit of the Connections feature, however, is that the connected set of ConfigMaps can grow, shrink, and change while the pod remains running. In terms of the use case here, that means you could add another model to be handled by the deployment engine without interrupting any current requests that the engine is processing.

Serving Queries

The “haproxy” service port on the inference deployment can now be used to service REST API queries, using the model created earlier to make inferences about how long a proposed taxi ride will take. An example script for making queries to this service can be found here.

Here’s that script in action, sending a query to the inferencing deployment previously created:

mlops figure 4

From this output, our example script queries the model via REST API calls, providing a few parameters, and the model returns a prediction.


While we’ve presented the steps from input dataset to query result sequentially here, in “steady state” operation all of these operations can be happening in different order and/or at the same time.

Building our pipeline using Kubernetes resources and the open source KubeDirector project lets us deal flexibly with a wide range of configurations and changes, with minimal disruption. The possibilities include:

  • Multiple notebooks per training deployment, with each notebook’s unique password given to a group or a single user for finer-grained access control.
  • Multiple training deployments to implement different training methods or access different datasets.
  • Changing existing notebooks to connect to different training deployments, without losing any of the working environment of the notebook.
  • Multiple inferencing deployments, for access control or load-balancing.
  • Multiple models served per inferencing deployment, at different service ports or different URLs.
  • Changing the set of models served by an inferencing deployment, without interrupting its operation.

Find out more about KubeDirector applications and KubeDirector clusters at the GitHub site: GitHub. From there you can download open-source KubeDirector and use it in your own Kubernetes clusters!

And to see how we use KubeDirector in our HPE Ezmeral Container Platform, check out the interactive demo environment here.

Look for more KubeDirector posts coming soon on the HPE DEV blog site.

This article was co-authored by:

  • Joel Baxter, Distinguished Technologist
  • Kartik Mathur, Master Technologist
  • Don Wake, Technical Marketing Engineer


Ted Dunning & Ellen Friedman

3 ways a data fabric enables a data-first approach

Mar 15, 2022
Nicolas Perez

A Functional Approach to Logging in Apache Spark

Feb 5, 2021
Cenz Wong

Getting Started with DataTaps in Kubernetes Pods

Jul 6, 2021
Kiran Kumar Mavatoor

Accessing HPE Ezmeral Data Fabric Object Storage from Spring Boot S3 Micro Service deployed in K3s cluster

Sep 13, 2021
Carol McDonald

An Inside Look at the Components of a Recommendation Engine

Jan 22, 2021
Carol McDonald

Analyzing Flight Delays with Apache Spark GraphFrames and MapR Database

Dec 16, 2020
Nicolas Perez

Apache Spark as a Distributed SQL Engine

Jan 7, 2021
Nicolas Perez

Apache Spark Packages, from XML to JSON

Dec 11, 2020

HPE Developer Newsletter

Stay in the loop.

Sign up for the HPE Developer Newsletter or visit the Newsletter Archive to see past content.

By clicking on “Subscribe Now”, I agree to HPE sending me personalized email communication about HPE and select HPE-Partner products, services, offers and events. I understand that my email address will be used in accordance with HPE Privacy Statement. You may unsubscribe from receiving HPE and HPE-Partner news and offers at any time by clicking on the Unsubscribe button at the bottom of the newsletter.

For more information on how HPE manages, uses, and protects your personal data please refer to HPE Privacy Statement.