search

opendatahub-io / ai-edge

ODH integration with AI at the Edge usecases
Apache License 2.0
8 stars 17 forks source link

readme

Open Data Hub / AI Edge

As the adoption of AI and ML continues to surge, there's an increasing demand to deploy these workloads in edge environments such as factories, farms, and remote locations where traditional computing infrastructure is limited, connectivity is unreliable or constrained, and reliability is paramount.

This repository contains tools and artifacts comprising a proof of concept to bring the power of Open Data Hub and OpenShift AI to such edge environments.

[!WARNING] This is a proof of concept repository and is not intended for production use. The edge features are experimental and the repository is under active development.

Overview

The term edge is a highly overloaded term in the industry. For the purpose of this repository, the following diagram illustrates the concept of core and near edge environments.

Terminology

Use Case

The main objective currently is to showcase that a user can:

  1. Package an inference service by integrating a trained model with all required dependencies into a container image, ensuring it's ready for utilization at the near edge location(s) in a streamlined manner.
  2. Centrally control and manage the set of target near edge environments from the core OpenShift cluster, and configure the initialization of a GitOps workflow on each of those targets.
  3. Utilize GitOps tooling at each near edge environment to reconcile the manifests specified in the GitOps repository (managed centrally), and deploy the inference service.
  4. Monitor and observe the deployed AI/ML inference workloads in the near edge environments from the core OpenShift cluster.

[!NOTE] Developing and training a model is currently out of scope of this repository as this repository already contains example models in the examples/models/ directory. You can read more about developing and training a model in the Developing and training a model section.

Architecture: A Bird's Eye View

The diagram above illustrates the architecture of the proof of concept. The two main categories of clusters are the core and near edge clusters.

[!NOTE] The rest of this documentation assumes usage of ACM^2 for cluster and application management, and we highly recommend this approach. If you choose to use a different approach, you'll find some information on how to deploy the inference container application using standalone Argo CD instances on the edge clusters, in the dedicated Bring Your Own Cluster Management document.

Core Cluster

Near Edge Cluster(s)

Preparing the Infrastructure

Before you can start using the proof of concept, you need to prepare the infrastructure. This includes setting up the OpenShift clusters and installing the required components.

Follow the steps in Preparing the Infrastructure to set up the infrastructure.

Proof of Concept Walkthrough

The proof of concept is built to provide two main workflows as shown in the diagram above

Workflow 1: Building an Inference Application Container Image

In manifests/README.md we show how to take the trained models, stored in an S3 bucket or Git repository, build container image(s) with the model and serving runtime using OpenShift Pipelines, push the container image(s) to an image registry accessible by the near edge cluster(s), and update a clone of this repository with a pull request, configuring acm/odh-edge/apps/*/kustomization.yaml with location and digest (SHA-256) of the built images.

You can skip this step if you do not wish to rebuild the container images. If you use the default configuration as shown in this git repository, you will use already built container images from the https://quay.io/organization/rhoai-edge repositories.

Workflow 2: Deploying and Updating an Inference Application on Near Edge Clusters

With the remote near edge clusters now available in the cluster set and the Argo CD, it is now possible to deploy the applications to those clusters using the Argo CD Pull model integration in ACM GitOps.

We will create the objects in the respective namespaces from the acm/registration/ directory structure.

However, it is important to note that the actual application configuration, stored in acm/odh-edge/, does not come from our local git repository checkout. Instead, acm/registration/near-edge/overlays/tensorflow-housing-app/kustomization.yaml and acm/registration/near-edge/overlays/bike-rental-app/kustomization.yaml contain source URLs of the git repositories that will become configurations of the ACM Channels, and the applications in those remote git repositories will be deployed. By default, github.com/opendatahub-io/ai-edge is configured; edit the URLs to match your repositories.

Then, as a user with permissions to create/update the following resource types in the openshift-gitops namespace on the ACM hub cluster:

ApplicationSets
GitOpsClusters
ManagedClusterSetBindings
PlacementBindings
Placements
Policies

run the following command:

oc apply -k acm/registration/
Credentials for private repositories

If the GitOps repository that Argo CD on the edge clusters will deploy resources from requires some form of credentials, then these credentials will need to be provided to each cluster via a Secret. This can be done by specifying the Secret inside an ACM Policy resource on the ACM hub cluster in such a way that ACM will instruct the edge clusters to create it. The following example illustrates how this may be done:

cat << EOF | oc apply -f -
apiVersion: policy.open-cluster-management.io/v1
kind: Policy
metadata:
  name: namespace-policy
  namespace: openshift-gitops
spec:
  disabled: false
  policy-templates:
    - objectDefinition:
        apiVersion: policy.open-cluster-management.io/v1
        kind: ConfigurationPolicy
        metadata:
          name: ensure-namespace-exists
        spec:
          object-templates:
            - complianceType: musthave
              objectDefinition:
                apiVersion: v1
                kind: Secret
                metadata:
                  name: first-repo
                  namespace: openshift-gitops
                  labels:
                    argocd.argoproj.io/secret-type: repository
                stringData:
                  type: git
                  url: https://github.com/argoproj/private-repo
          pruneObjectBehavior: DeleteIfCreated
          severity: low
          remediationAction: enforce
EOF

View the application deployments

ACM

In the OpenShift Console of the ACM hub cluster in All Clusters > Applications, search for the application name. You should see a few results results: <app name>-appset, <cluster name>-<app name>, and <app name>-1. Clicking on the <cluster name>-<app name> entry, and then navigating to the Topology tab, you can see what objects were created on the remote cluster.

Everything should be shown green. If it is not, click the icon of the faulty object and check the displayed information for debugging clues.

Argo CD

The Argo CD provided by Openshift GitOps has a console on each near edge cluster, showing detailed information on each application that it manages on that cluster. On the particular edge cluster, open the console by navigating to the domain specified by the openshift-gitops-cluster Route in the openshift-gitops namespace on that cluster. Once logged-in to the Argo console, search for the application name, and then select the entry named in the format <cluster name>-<application name> to see more information.

Observability setup

Core Cluster

Edge cluster(s)

[!NOTE] If you are not deploying our example models then you must edit the static list of targets to contain the service pointing to your custom model.

Developing and training a model

This step is out of scope of this Proof of Concept repository, as this repository already contains trained models in the examples/models/ directory.

If you wish to develop and train different models, Jupyter notebooks provided by Open Data Hub (ODH) or Red Hat OpenShift AI can be used. To install ODH or OpenShift AI operators, admin privileges in the OpenShift cluster are needed.

Working and deploying your own models might require bigger changes to the definition and configuration of the pipelines and ACM setup below, so you might want to start with the pre-built models first.

Using local models in pipelines

In examples/model-upload/ you can upload a local model file to be used in our pipelines. This is done by uploading a model to a PVC and copying that model to our pipeline's workspace for use while it is running.

Upload model to PVC:

make MODEL_PATH="PATH_TO_A_FILE" NAME=my-model create

You should get a final output showing details of the upload

PVC name: model-upload-pvc
Size: 1G
Model path in pod: /workspace/model-upload-pvc/model_dir/model.model

You can set the SIZE and PVC values aswell

make MODEL_PATH="PATH_TO_A_FILE" NAME=my-model SIZE=1G PVC=my-new-PVC create

You can then use the copy-model-from-pvc task to copy the model from the model-workspace, which can be set to the PVC created in the last step, to the build-workspace-pv workspace, which is then used by the buildah task in the pipeline. copy-model-from-pvc task is not included in the pipeline by default, you have to add it yourself.

Contributing

See CONTRIBUTING.md.

Development

See DEVELOPMENT.md.