hive-gate 1.3

Python interface to MAIA. It can be used as interface to any Kubernetes-based platform.

Deploy Custom Environments in MAIA

To create custom environments and deploy applications in MAIA (including pods, services and ingresses) a Helm chart is available: Hive_Gate.

With the Hive_Gate chart it is possible to deploy any Docker Image as a Pod, expose the required ports as services, mount persistent volumes on the specified locations and optionally create Ingress resources to expose the application to the external traffic using the HTTPS protocol.

To add the chart to Helm, run:

helm repo add hive-gate https://maia-kth.github.io/Hive_Gate/
helm repo update

Custom Helm values

A number of custom parameters can be specified for the Helm chart, including the Docker image to deploy, the port to expose, etc.

The custom configuration is set in a JSON configuration file, following the conventions described below.

General Configuration

Namespace [Required]

Specify the Cluster Namespace where to deploy the resources

{
  "namespace": "NAMESPACE_NAME"
}

Chart Name [Required]

Specify the Helm Chart Release name

{
  "chart_name": "Helm_Chart_name"
}

Docker image [Required]

To specify the Docker image to deploy

{
  "docker_image": "DOCKER_IMAGE"
}

In case of a custom Docker image, provide the docker build context ( the folder path containing the Dockerfile and all the required files), and the corresponding custom image name. The built Docker image is pushed to the private docker registry.

{
  "docker_image": {
    "context_folder": "/PATH/TO/MY/DOCKER_CONTEXT",
    "image_name": "my-custom-image"
  }
}

Requested Resources [Required]

To request resources (RAM,CPU and optionally GPU).

{
  "memory_size" : "REQUESTED_RAM_SIZE",
  "cpu" : "REQUESTED_CPUs"
}

Optionally, to request GPU usage:

{
  "gpu_request" : "NUMBER_OF_GPUs"
}

Services

To specify which ports (and corresponding services) can be reached from outside the pod.

{
  "ports" : {
    "SERVICE_NAME_1": ["PORT_NUMBER"],
    "SERVICE_NAME_2": ["PORT_NUMBER"]
  }
}

The default Service Type is ClusterIP. To expose a service as a type NodePort:

{
  "service_type": "NodePort",
  "ports" : {
    "SERVICE_NAME_1": ["PORT_NUMBER", "NODE_PORT_NUMBER"],
    "SERVICE_NAME_2": ["PORT_NUMBER", "NODE_PORT_NUMBER"]
  }
}

Persistent Volumes

2 different types of persistent volumes are available: hostPath (local folder) and nfs (shared nfs folder). For each of these types, it is possible to request a Persistent Volume via a Persistent Volume Claim, or to directly reference the folder in the host/nfs server to mount.

Direct access:

{
  "host_volume_mounts": 
  {
    "host-folder_1": ["/host/folder/path_1", "/mount/path_1","readOnly"],
    "host-folder_2": ["/host/folder/path_2", "/mount/path_2"]
  },
  "nfs_server": "NFS_SERVER_IP",
  "nfs_volume_mounts": 
  {
    "nfs-folder_1": ["/nfs/folder/path_1", "/mount/path_1","readOnly"],
    "nfs-folder_2": ["/nfs/folder/path_2", "/mount/path_2"]
  }
}

The "readOnly" options can be added to specify the mounted folder as read-only.

Request PVC:

{
  "persistent_volume" : 
  [
    {
      "mountPath": "/mount/path_1",
      "size": "VOLUME_SIZE",
      "access_mode": "ACCESS_TYPE",
      "pvc_type": "STORAGE_CLASS"
    },
    {
      "mountPath": "/mount/path_2",
      "size": "VOLUME_SIZE",
     "access_mode": "ACCESS_TYPE",
     "pvc_type": "STORAGE_CLASS"
    }
  ]
}

"STORAGE_CLASS" can be any of the storage classes available on the cluster:

kubectl get sc

Existing Persistent Volumes

Previously created pv can be mounted into multiple pods (ONLY if the access mode was previously set to **ReadWriteMany **)

{
 "existing_persistent_volume": [
  {
   "name": "EXISTING_PVC_NAME",
   "mountPath": "/mount/path"
  }
 ]
}

Mounted files

Single files can be mounted inside the Pod. First, a ConfigMap including the file is created, and then it is mounted into the Pod.

{
 "mount_files": {
  "file_name": [
   "/local/file/path",
   "/file/mount/path"
  ]
 }
}

Node Selection

To optionally select which node in the cluster to use for deploying the application.

{
  "node_selector": "NODE_NAME"
}

GPU Selection

To optionally select which available GPUs in the cluster to request. type and vram attribute can be specified (only one of them is needed, both can be included). Example: type: "RTX-2070-Super", vram: "8G"

{
  "gpu_selector": {
    "type": "GPU_TYPE",
    "vram": "VRAM_SIZE"
  }
}

Ingress

Used to create an Ingress resources to access the application at the specified port by using an HTTPS address.

IMPORTANT! The specified DNS needs to be active and connected to the cluster DNS (".k8s-maia.com")

IMPORTANT! oauth_url should be explicitly specified, since only oauth-based authenticated users can be authorized through the ingress. Contact the MAIA admin to retrieve this information.

{
 "ingress": {
  "host": "SUBDOMAIN.k8s-maia.com",
  "port": "SERVICE_PORT",
  "oauth_url": "oauth.MY_NAMESPACE"
 }
}

Environment variables

To add environment variables, used during the creation and deployment of the pod (i.e., environment variables to specify for the Docker Image).

{
  "env_variables": 
  {
    "KEY_1": "VAL_1",
    "KEY_2": "VAL_2"
  }
}

Hive Docker Configuration

User info

When deploying Hive-based applications, it is possible to create single-multiple user account in the environment. For each of the users, username, password email, and, optionally, an ssh public key are required. This information is stored inside Secrets:

USER_1_SECRET:
    user: USER_1
    password: pw
    email: [e-mail]
    ssh_publickey [Optional]: "ssh-rsa ..." 

To provide the user information to the Pod:

{
  "user_secret" : 
  [ 
    "user-1-secret",
    "user-2-secret"
  ],
  "user_secret_params" : ["user","password","email", "ssh_publickey"]
}

Configuration File Example

{
  "namespace": "machine-learning",
  "chart_name": "jupyterlab-1-v1",
  "docker_image": "jupyter/scipy-notebook",
  "tag" : "latest",
  "memory_size" : "4Gi",
  "cpu" : "5000m",
  "ports" : {
    "jupyter": [8888]
  },
  "persistent_volume" :
  [
    {
      "mountPath": "/home/jovyan",
      "size": "100Gi",
      "access_mode": "ReadWriteOnce",
      "pvc_type": "local-hostpath"
    }
  ]
}

Deploy Charts

To deploy an Helm Hive Chart, first create a config file according to the specific requirements (as described [above](#Custom Helm values)). Then install the Hive_Gate package running:

pip install hive-gate

Or download the executable file:

[Hive_Gate_deploy_helm_chart (Windows .exe)]

Hive_Gate_deploy_helm_chart (Ubuntu)

Finally:

Hive_Gate_deploy_helm_chart --config-file /PATH/TO/CONFIG/FILE