# AWS-ECR-EKS MODEL DEPLOYMENT

## Description
Demo Pipeline for model deployment using ECR for container registry and EKS clusters serving the application.

## Use Case
The repository goal is to provide a quick way to deploy a PyTorch model for Fashion MNIST Classification to Kubernetes in an AWS EKS cluster.

## Set up

### Creating EKS cluster with runway

Before starting the pipeline for the first time, if you don't have an existing EKS cluster, it is necessary to set up a cluster on AWS correctly. You could do it manually, but we recommend the [terraform-aws-eks](https://github.azc.ext.hp.com/runway/terraform-aws-eks) runway project. Our tests are made using the "[eks-with-new-vpc](https://github.azc.ext.hp.com/runway/terraform-aws-eks/tree/master/examples/eks-with-new-vpc)" example.

With the cluster created, it is needed to log the credentials in kubeclt to permit a role ARN to alter the cluster. To log in, use the command:

aws eks update-kubeconfig --name EKS_CLUSTER_NAME --region YOUR_AWS_REGION

Now alter the configmap to enable the role to make modifications:

kubectl edit -n kube-system configmap/aws-auth

Add the credentials like so:

apiVersion: v1
data:
  mapRoles: |
  ...
  
  - groups:
      - system:masters
    rolearn: arn:aws:iam::YOUR_IAM_ROLE/role
    username: default
    
  ...
  

Now, your IAM role can create the services with codeway.

### Changes to run the codeway template

It is essential to change some values at [codeway.yaml](./codeway.yaml) to suit your needs: roleARN , awsAccountNumber , externalId, awsRegion, and the eks cluster name. Those values are located at "registryParameters", the second "parameters" indentation, and the first bash command.

          ...
      
      registryParameters:
        - roleARN: 'YOUR_ROLE_ARN'
          awsAccountNumber: 'YOUR_AWS_ACCOUNT_NUMBER'
          externalId: 'YOUR_EXTERNALID'
          awsRegion: 'YOUR_AWS_REGION'
          
          ...
          
        parameters:
          roleArn: 'YOUR_ROLE_ARN'
          externalId: 'YOUR_EXTERNALID'

      - bash: aws eks update-kubeconfig --name EKS_CLUSTER_NAME --region YOUR_AWS_REGION
          
          ...

### Inference
To send a request to the server is necessary to pass encoded images from Fashion MNIST as a parameter in an HTTP request. To encode the images, use the code in [encode_img.py](./encode_img.py). There are three example images in the [examples.txt](./exemple_inference_imgs.txt) file.

The server address is the external-IP that will take a few minutes to be updated, so to check it, you need to run:

kubectl get svc

It will show:

NAME              TYPE           CLUSTER-IP      EXTERNAL-IP   PORT(S)          AGE
example-service   LoadBalancer   10.100.157.50   EXTERNALIP     5000:30729/TCP   3s

To compose the link, do the following:

EXTERNAL-IP + ":5000?data=" + encoded image

example: "afe1bd276f20141ba96388a78259f638-1063200090.us-east-1.elb.amazonaws.com:5000?data=AAAAAAAAAA..."

### Maintained by:
* Heitor de Castro Felix - [e-mail]
* Filipe Figueredo Monteiro - [e-mail]
* Paulo de Oliveira Guedes - [e-mail]
* Davi Monteiro Paiva - [e-mail]
* Felipe de Melo Battisti - [e-mail]

## Roadmap
We want to integrate the pipeline to use models from the databricks model registry.

## Limitations
Currently, the pipeline does not have CI/CD components, the execution of the first iteration is automatic but not automatically updated by triggers.