Page Comparison

...

For some algorithms, cross-edge synchronization is required. More details will be provided later.

Design Details

The model operator is supposed to run as a separate binary, fully decoupled from the KubeEdge platform code. It leverages the KubeEdge platform to schedule work on edge nodes. Here is the high level design diagram for the work flow:

Image Added

There are generally two phases for machine learning model development, i.e. training and inference. Model behaviors are quite different depending on whether it is used for training or for inference. So we might as well define two different types of model CRDs:

...

Code Block

language	cpp
title	InferenceModel CRD

// InferenceModelSpec defines the desired state of an InferenceModel.
// Two ways of deployment are provided. The first is through a docker image.
// The second is through a data store, with a manifest of model files.
// If image is provided, manifest and targetVersion will be ignored.
type InferenceModelSpec struct {
	ModelName     string               `json:"modelName"`
	DeployToLayer string               `json:"deployToLayer"`
	FrameworkType string               `json:"fraemworkType,omitemptyframeworkType"`

	// +optional
	NodeSelector  map[string]string    `json:"nodeSelector,omitempty"`

	Image// +optional
	NodeName string `json:"nodeName,omitempty"`

	// +optional
	Image string               `json:"image,omitempty"`

	Manifest// +optional
 	Manifest   []InferenceModelFile `json:"manifest,omitempty"`

	// +optional
	TargetVersion string               `json:"targetVersion,omitempty"`

	Replicas// +optional
	//    *int32              +kubebuilder:validation:Minimum=0
	ServingPort int32 `json:"servingPort"`

	// +optional
	// +kubebuilder:validation:Minimum=0
	Replicas *int32 `json:"replicas,omitempty"`
}

// InferenceModelFile defines an archive file for a single version of the model
type InferenceModelFile struct {
	Version     string `json:"version,omitempty"`
	DownloadURL string `json:"downloadURL,omitempty"`
	Sha256sum   string `json:"sha256sum,omitempty"`
}

// InferenceModelStatus defines the observed state of InferenceModel
type InferenceModelStatus struct {
	URL            string `json:"url,omitempty"`
	ServingVersion string `json:"servingVersion,omitempty"`
}

Code Block

language	yml
title	Sample inferencemodel instance

apiVersion: ai.kubeedge.io/v1alpha1
kind: InferenceModel
metadata:
  name: facialexpression
spec:
  modelName: facialexpression
  deployToLayer: edge
  frameworkType: tensorflow
  image:
  nodeSelector:
    kubernetes.io/hostname: precision-5820
  nodeName: precision-5820
  manifest:
    - version: '3'
      downloadURL: http://192.168.1.13/model_emotion_3.tar.gz
      sha256sum: dec87e2f3c06e60e554acac0b2b80e394c616b0ecdf878fab7f04fd414a66eff
    - version: '4'
      downloadURL: http://192.168.1.13/model_emotion_4.tar.gz
      sha256sum: 108a433a941411217e5d4bf9f43a262d0247a14c35ccbf677f63ba3b46ae6285
  targetVersion: '4'
  servingPort: 8080
  replicas: 1

An instance of the InferenceModel specifies a single serving service for the provided model.

...

Deployment method	Pros	Cons
Docker image	Docker images encapsulate all the runtime dependencies Very flexible as users can build whatever image they want Docker will manage all the life cycles of the "model offloading"	Users have to build the images themselves, write the Dockerfile, build the image and upload to a docker registry Users have to provide a private docker registry if they don't want to use the public dockerhub.
Machine learning model file manifestsfiles manifest	Data scientists directly work with model files. It would be nice if they can just drop their model files somewhere By using a data store, it opens the door for serverless computing	Our framework has to manage the whole life cycles of model files deployment, update, delete, etc.

How can the InferenceModel CRD be used?

Simple machine learning offloading to edge

Just create an instance of InferenceModel with "DeployToLayer == edge"

Joint Inference by edge and cloud

Create three resources:

An instance of InferenceModel to the cloud
An instance of InferenceModel to the edge
A pod running on the edge for serving customer traffic. It contains the logic for deciding whether or not to call cloud model serving API.

Joint inference by device, edge and cloud

We can have three models, with different size and accuracies, running on device, edge, and cloud respectively.

Versions Compared

Old Version 13

New Version Current

Key

Design Details

How can the InferenceModel CRD be used?

Simple machine learning offloading to edge

Joint Inference by edge and cloud

Joint inference by device, edge and cloud