Model Deployment (Alpha)

Alpha version. It is subject to changes in the stable release.

Model Deployment feature that users are allowed to manage model deployment endpoints within Group Management with CRUD operations. If the feature is enabled within a group, end-users of the group are allowed to deploy models obeying the restriction of the group quota. While a deployed model serves, administrators are able to monitor the status of model service such service health and resources usage information; moreover, end-users are able to check deployment history and log.

Enable Model Deployment in Group Management

First, we have to turn on Model Deployment in groups.

If the current group's Model Deployment is not enabled, it shows a message

Feature not available - Model Deployment is not enabled for this group. Please contact your administrator to enable it.

Please contact Admin to turn on it or switch to a group which Model Deployment is enabled.

List

This page indicates all of deployed services status.

Status	Deployed	Failed	Deploying	Stopped

Cell

Each cell represents a deployment. Clicking on a cell to view the detail.

Tips

Hovering the cursor over Metadata View of a cell, it lists the metadata view.
Hovering the cursor over Endpoint Link shows the entire link and clicking the link makes a copy to the clipboard.

Info	Description
Title	Deployment name
Group	Deployment's owner group.
Endpoint	Where the model is deployed and serve.
Metadata	Hovering over `View` to display the metadata info
Last Updated	Last updated time.

Search deploy name: Search deployments by a keyword.
Create Deployment button: Clicking for a deployment creation.
Refresh button: Clicking to retrieve the latest statuses of deployments.
Deployed By Me: Check; list deployments deployed by "me".

Create

Confirm if the current group is what you desire; switch the group by the Group: dropdown at the top of the right side.

Environment Settings

InstanceTypes: Select a instance type which allocates resources for the job.
Replicas: How many replicas for the service.

Endpoint

Private Access: The toggle of private access of the endpoint; if enabled, there is Clients tab appearing in detail page.

Deployment Details

Deployment Name: The name of the deployment.
Deployment ID: An unique ID.
Model Image: The image which the model is based on.
Image Pull Secret: a pull secret for the model image if required.
Descriptions: User input description.
Update Message: Leave a message here for each update.

Metadata

We can add key-value pairs.

Name: The name of key.
Value: The value.

Click Deploy button to start the deployment.

One the deployment is triggered, there is a notification popping up, we can click on it to view the deployment detail.

Deployment Detail

Information

Info	Description
Status	The status of the deployment.
Message	The message related to the deployment if any.
Endpoint	Where the model is deployed and serve.
Access Type	Public or Private
Model Image	The image which the model is based on.
Replicas	The replicas of the deployment.
Deployment Name	The name of the deployment.
Group	The owner group.
Instance Type	The resources allocation which is requested for the deployment.
Creation Time	Timestamps
Last Updated	Timestamps
Description	The description which users input during the creation.
Run an Example	Verifying the service by using `curl` querying the service; it is varied with `Public` or `Private` access.

Logs

Replicas: Select a replica.

In Logs tab, the logs of current deployment are displayed here.

Timestamp reflects Universal Time Coordinated (UTC).

Clicking Scroll to Bottom button skips to the bottom of logs.

It displays only logs of latest 2000 lines; clicking the Download button to download the whole log file.

History

In History tab, we can view the detail of historical deployments with each update message.

Clicking on the View link for the detail of each deployment.

Info	Description
User	Who triggered this deployment
Deployment Stopped	true or false
Model Image	The model image url
Replicas	The amount of replicas
Group	The deployment's owner group
Instance Type	The instance type is selected
Timestamp	Last updated time
Description	User input description
Metadata	Metadata mappings table

Clients

This tab appears if Private Access of the endpoint is enabled.

Filling in Client Name and clicking Add client to generate a Client Token.

The token is mandatory for accessing a private endpoint, we can bring it as -u <client-name>:<client-token> in a curl command.

curl -X POST \
    -u <client-name>:<client-token> \
    -d '{"data":{"names":["a","b"],"tensor":{"shape":[2,2],"values":[0,0,1,1]}}}' \
    -H "Content-Type: application/json" \
    https://<primehub>/deployment/<model>/api/<version>/predictions

Client Token displays once only for a short while, please save it. If lost, deleting the record and generate it again.

Tips:

In the predict function of a deployed model, we may want to collect the information of clients who send requests to. We could learn the info from the header of requests, X-Forwarded-User like the example below.

from flask import request as req

...
req.headers.get('X-Forwarded-User') # you can get the client name from the header
...

Update

On a deployment detail page, clicking Update button to enter the deployment editing.

Group、 Deployment name、 Deployment ID are not allowed to be modified in a update once the initial deployment is created.

Delete

Clicking on the cell of the deployment which we want to delete.

On a deployment detail page, clicking Delete button at top-right for the deployment deletion.

Stop

Clicking on the cell of the deployment which we want to stop.

On a deployment detail page, clicking Stop button at top-right to stop the current deployed service.

Monitor

PrimeHub provides a monitoring dashboard based on Seldon Core Analytics on Grafana, we are able to monitor metrics of deployment/model/model version.

Accessing the Grafana from User Portal.
Selecting PrimeHub / Model Deployments dashboard, it lists deployment which are providing services.
Selecting a deployment and monitoring metrics on the board.

Default Metrics:

QPS (Queries Per Second)
Success rate
4xx, error if any
5xx, error is any
Predict QPS
Reward

The reward is interpreted as the proportion of successes in the batch of data samples. Thus this implementation inherently assumes binary rewards for each sample in the batch. The helper function n_success_failures calculates the number of successes and failures given the batch of data samples and the reward. -Reference.
Latency

The board is based on Seldon Core Analytics, it can be advanced customized. Please see the document and code.