BridgeAI-gitops-infra

Prerequisities

A Linux or OSX Based computer with at least 16GB of RAM.

Dependencies

Ensure you have the following installed:

• Docker
• kind
• kubectl
• flux
• gnupg
• jq this one or the brew formula
• github cli

Ensure you have the following accounts:

• Github
• Dockerhub

You will need to setup a CLASSIC github personal access token that has repo and package:read permissions. It will need to be set as an ENVAR that the CLI tools can access, we recommend either setting it in the terminal or as a part of your shell profile.

export GITHUB_TOKEN=<token>

You will also need a DockerHub account with a public repository setup that we can push containers to via an access token.

OSX

Install using homebrew:

brew install kind kubectl fluxcd/tap/[email protected] gnupg jq gh

You should install docker desktop from the Docker website.

Linux

Install the listed dependencies using a package manager of your choice.

Useful Knowledge

Having an understanding of the following subjects will useful:

• Python
• Kubernetes
• Git
• Docker

Getting started

Clone the repository.

git clone https://github.com/digicatapult/bridgeAI-gitops-infra.git

Ensure that the Docker engine has at least 12GB of RAM resourced to it and several GB of swap.

Run the ./scripts/add-kind-cluster.sh script.

Select the GitOps branch you wish to track and run ./scripts/install-flux.sh -b <branch>

This will bring up a kind cluster with flux installed and will automatically install nginx and airflow

Verify that all the helm charts have successfully installed

flux get helmreleases -A

Local Workflow usage

See local workflow document for running the localised demonstrator.

Local Application Instances

Airflow

You can access the airflow UI on http://localhost:3080/airflow Username: admin Password: admin

MLFlow

You can access the MLFlow UI on http://localhost:3080/mlflow Username: admin Password: password You can set the MLFLOW_TRACKING_URI clientside as http://localhost:3080/ if you wish to use the mlflow binary clientside. Airflow has a variable set to locate the mlflow_tracking_uri

MLFlow MinIO

MinIO is used as the storage backend to store MLFlow's artifacts

To access MinIO UI perform port forwarding on the svc:

kubectl port-forward svc/mlflow-minio 9001 -n default

Access MinIO UI on: http://127.0.0.1:9001 Username: admin Password: password

DVC and Evidently

We use a separate instance of MinIO as the storage backends for DVC data and Evidently reports.

To access this instance of MinIO UI perform port forwarding on the svc:

kubectl port-forward svc/minio 9002:9001 -n default

Access MinIO UI on: http://127.0.0.1:9002 Username: admin Password: password

KServe

KServe is used as our model serving and inference framework. We are using it in rawDeployment mode which means a container is spawned for running the model application. This a Kubernetes CRD and can only be queried using kubectl commands. You can see which CRDs can be queried using the kubectl get crd |grep kserve command. We typically only make use of the inferenceservices CRD in this demonstrator.

Prediction Service

The Prediction service is a REST API for interfacing with the house-price prediction inference service. It is presented as a swagger UI and is available at http://localhost:3080/swagger

Cleanup

To delete a cluster

kind delete clusters bridgeai-gitops-infra

Verify if cluster is deleted

kubectl config get-contexts