gem-hunter

This repo will show you how to deploy your own Zero-shot Object Detection API onto GKE using CI/CD ⭐⭐⭐

• Setup alert service
• Distributed tracing using Jaeger and Opentelemetry
• Logs management

1. Create GKE Cluster using Terraform
2. Deploy serving service manually
   1. Deploy Nginx Ingress Controller
   2. Deploy API
3. Deploy monitoring service
4. Continuous deployment to GKE using Jenkins pipeline
   1. Create Google Compute Engine
   2. Install Docker and Jenkins in GCE
   3. Connect to Jenkins UI in GCE
   4. Setup Jenkins
   5. Continuous deployment
5. API result

• Gcloud CLI can be installed following this document:
  • For mac: https://cloud.google.com/sdk/docs/install#mac
  • For ubuntu: https://cloud.google.com/sdk/docs/install#deb
• Initialize the Google Cloud CLI.

gcloud init
Y

• Pick you cloud project then type Enter.
• Check if the Google Cloud CLI is installed successfully.

gcloud -v

gcloud components install gke-gcloud-auth-plugin

• Create your service account, and select Kubernetes Engine Admin role therefore you will have full management of Kubernetes Cluster and their Kubernetes API object for your service account.
• Create new key as json type for your service account. Download this json file and save it in terraform directory. Update credentials in terraform/main.tf with your json directory.

• Go to IAM, click on GRANT ACCESS, then add new principals, this principal is your service account created in step 1.3. Finally, select Owner role.

• Change the default value of variable project_id in terraform/variables.tf with your project id on Google Cloud. Then run the following command to create GKE cluster:

gcloud auth application-default login

cd terraform
terraform init
terraform plan
terraform apply

• After you run these command lines, you will see the GKE cluster is deployed at asia-southeast1 with its node machine type is: e2-standard-2 (2 vCPU, 1 core, 8 GB RAM and costs $144.35/1month). You can change these settings in terraform/variables.tf to your desired setting.
• Remember not to set enable_autopilot=true in terraform/main.tf as Prometheus service cannot scrape node metrics from Autopilot cluster.

• After the cluster was created successfully, click on your cluster and select Connect button. Then copy and paste the Command-line access into you terminal.
• You can check the connection by using this command

alias k=kubectl
k get nodes

Using Helm Chart to deploy the application on GKE. See the installation here.

cd helm/nginx-ingress
k create ns nginx-ingress
kubens nginx-ingress
helm upgrade --install nginx-ingress-controller .

cd helm/app
k create ns model-serving
kubens model-serving
helm upgrade --install app .

• This will create 3 pods.
• Obtain the IP address of nginx-ingress.

k get ing

• Add the domain name zod.com (set up in helm/app/templates/nginx-ingress.yaml) of this IP to /etc/hosts

sudo nano /etc/hosts
[YOUR_INGRESS_IP_ADDRESS] zod.com

• Then you can access the API UI by zod.com/docs

Using Prometheus and Grafana for monitoring both node and containers (pods). Prometheus will scrape the metrics from both node and containers then display by using Grafana UI. Lastly, the system health alerts will be sent to Discord.

• First install Kube-prometheus-stack which will contain every component of monitoring stack. Then deploy on a new namespace called monitoring

helm repo add prometheus-community https://prometheus-community.github.io/helm-charts
k create ns monitoring
kubens monitoring
helm install kube-prometheus-stack --namespace monitoring prometheus-community/kube-prometheus-stack

• Log in to Prometheus UI

k port-forward -n monitoring svc/kube-prometheus-stack-prometheus 9090:9090

• Login to Grafana UI

k port-forward -n monitoring svc/kube-prometheus-stack-grafana 8080:80

username: admin
password: prom-operator

Jenkins is deployed on Google Compute Engine using Ansible.

NOTE: Make sure you have installed miniforge, you can see the installation here.

Then you can install ansible by running the below command:

conda create -n [your_desired_env_name] python=3.11
conda activate [your_desired_env_name]
pip install ansible

Check if ansible is successfully installed

ansible --version

• Create your service account, and select Compute Admin role (Full control of all Compute Engine resources) for your service account.
• Create new key as json type for your service account. Download this json file and save it in secrets directory. Update your project and service_account_file in ansible/deploy_jenkins/create_compute_instance.yaml.
• In the terminal run the following command lines to create Google Compute Engine:

cd ansible
ansible-playbook create_compute_instance.yaml

• Create ssh key, and select this directory [YOUR DIR]/.ssh/id_rsa

ssh-keygen -t rsa

Then run cat [YOUR DIR]/.ssh/id_rsa.pub and copy the content.

• In the Google Computing Engine settings, select Metadata and add your SSH key.
• Run cp example.inventory inventory and replace all value inside the double single quotes in the created inventory file.

cd ansible/deploy_jenkins
ansible-playbook -i ../inventory deploy_jenkins.yaml

• Access the instance by using this command

ssh -i ~/.ssh/id_rsa YOUR_USERNAME@YOUR_EXTERNAL_IP

• Check if jenkins container is already running ?

sudo docker ps

• Open web brower and type [YOUR_EXTERNAL_IP]:8081 for access Jenkins UI. To Unlock Jenkins, please execute the following commands:

sudo docker exec -ti serving_grounding_dino-jenkins bash
cat /var/jenkins_home/secrets/initialAdminPassword

Copy the password and you can access Jenkins UI.

• Connect to GitHub repo to Jenkins using Webhook
• Add Github credential to Jenkins (select appropriate scopes for the personal access token)
• Install the Kubernetes, Docker, Docker Pineline, GCloud SDK Plugins at Manage Jenkins/Plugins. After successful installation, restart the Jenkins container in your Compute Engine instance:

sudo docker restart serving_grounding_dino-jenkins

• Add Dockerhub credential to Jenkins at Manage Jenkins/Credentials
• Set up a connection to GKE by adding the cluster certificate key at Manage Jenkins/Clouds.

kubectl create clusterrolebinding model-serving-admin-binding \
  --clusterrole=admin \
  --serviceaccount=model-serving:default \
  --namespace=model-serving

kubectl create clusterrolebinding anonymous-admin-binding \
  --clusterrole=admin \
  --user=system:anonymous \
  --namespace=model-serving

• If everything is setup correctly, then the result should look like this

• Run the following command to test the API:

python client.py

• If you want to test the API on you own image:

python client -u [YOUR_API_URL] -p [YOUR_PROMPT] -i [YOUR_IMAGE_PATH]