3.6

Deployments

Enterprise Applicable to Enterprise Edition
Deploy Applicable to Deploy Edition

Deployments feature that users are able to deploy trained models as services and to manage these deployments. While model services are online, administrators are able to monitor the status of model service such service health and resources usage information; moreover, users are able to check deployment history and log.

PrimeHub provides two ways of Model Deployment, one is to deploy a model image directly which has wrapped a model file into, the other is to specify a model file and a pre-packages server image as a deployment, under the hood, PrimeHub serves it as a service.

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.

StatusDeployedFailedDeployingStopped

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.

InfoDescription
TitleDeployment name
EndpointWhere the model is deployed and serve.
MetadataHovering over View to display the metadata info
Last UpdatedLast 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.

Deployment Details

Since Model URI is introduced, there are two ways for model deployments. One way is to deploy a ready-to-serve packaged model image which is packaged by user already directly using Model Image field; the other way is to assign a model file using the Model URI field with a specified pre-packaged image using Model Image field, under the hood, the model file will be packaged into images.

Environment Variables

Add key-value pairs of environmental variables.

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

Metadata

Add key-value pairs of metadata.

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

Resources

  • 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.

Deploy

  • Update Message: Leave a message here for each update.

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

InfoDescription
StatusThe status of the deployment.
MessageThe message related to the deployment if any.
EndpointWhere the model is deployed and serve.
Creation TimeTimestamps
Last UpdatedTimestamps
Model ImageThe image which can be served directly or acted as the pre-packaged ML base image.
Image Pull SecretThe secret is used for pulling the image.
Model URIThe location where the trained model is stored.
DescriptionThe description which users input during the creation.
Instance TypeThe resources allocation which is requested for the deployment.
ReplicasThe replicas of the deployment.
Access TypePublic or Private
Run an ExampleReplace ${YOUR_DATA} with real data in the Curl query sample to test the service; the sample varies according to Public or Private access.
  • Metadata table
  • Environment Variables table: disclose values by clicking the eye icon.

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.

InfoDescription
UserWho triggered this deployment
Deployment Stoppedtrue or false
Model ImageThe model image url
Model URIThe model stored location
ReplicasThe amount of replicas
GroupThe deployment's owner group
Instance TypeThe instance type is selected
TimestampLast updated time
DescriptionUser input description
Access TypePublic or Private
ClientsClients if Access Type is private
  • Metadata table
  • Environment Variables table: disclose values by clicking the eye icon.

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.

  1. Accessing the Grafana from User Portal.

  2. Selecting PrimeHub / Model Deployments dashboard, it lists deployment which are providing services.

  3. 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.

License warning

when used model amount > granted model amount + 10%, a warning message appears, furthermore, Create Deployment is disabled.

Please contact your system administrator for assistance to upgrade your license to run more models.

To learn the current PrimeHub license information, please see PrimeHub License.

← Manage and Deploy a ModelConcepts →