Prepackaged Model Servers

Seldon provides several prepacked servers you can use to deploy trained models:

For these servers you only need the location of the saved model in a local filestore, Google bucket, S3 bucket, azure or minio. An example manifest with an sklearn server is shown below:

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: sklearn
spec:
  name: iris
  predictors:
    - graph:
        children: []
        implementation: SKLEARN_SERVER
        modelUri: gs://seldon-models/sklearn/iris
        name: classifier
      name: default
      replicas: 1

The modelUri specifies the bucket containing the saved model, in this case gs://seldon-models/sklearn/iris.

modeluri supports the following three object storage providers:

  • Google Cloud Storage (using gs://)

  • S3-compatible (using s3://)

  • Minio-compatible (using s3://)

  • Azure Blob storage (using https://(.+?).blob.core.windows.net/(.+))

Init Containers

Seldon Core uses Init Containers to download model binaries for the prepackaged model servers. We use kfserving’s storage.py library for our Init Containers by defining

storageInitializer:
  image: gcr.io/kfserving/storage-initializer:v0.4.0

in our default helm values. See the Dockerfile and its entrypoint for a detailed reference. You can overwrite this value to specify another default initContainer. See details on requirements bellow

Secrets are injected into the init containers as environmental variables from kubernetes secrets. The default secret name can be defined by setting following helm value

predictiveUnit:
  defaultEnvSecretRefName: ""

Customizing Init Containers

You can specify a custom initContainer image and default secret globally by overwriting the helm values specified in the previous section.

To illustrate how initContainers are used by the prepackaged model servers, consider a Seldon Deployment with volumes, volumeMounts and initContainers injected by the Seldon Core Operator:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: iris
spec:
  predictors:
  - name: default
    replicas: 1
    graph:
      name: classifier
      implementation: SKLEARN_SERVER
      modelUri: s3://sklearn/iris

    componentSpecs:
    - spec:
        volumes:
        - name: classifier-provision-location
          emptyDir: {}

        containers:
        - name: classifier
          volumeMounts:
          - mountPath: /mnt/models
            name: classifier-provision-location
            readOnly: true

        initContainers:
        - name: classifier-model-initializer
          image: gcr.io/kfserving/storage-initializer:v0.4.0
          imagePullPolicy: IfNotPresent
          args:
          - s3://sklearn/iris
          - /mnt/models

          envFrom:
          - secretRef:
              name: seldon-init-container-secret

          volumeMounts:
          - mountPath: /mnt/models
            name: classifier-provision-location

Key observations:

  • Our prepackaged model will expect model binaries to be saved into /mnt/models path

  • Default initContainers name is constructed from {predictiveUnitName}-model-initializer

  • The entrypoint of the container must take two arguments:

    • First representing the models URI

    • Second the desired path where binary should be downloaded to

  • If user would to provide their own initContainer which name matches the above pattern it would be used as provided

Image and secret used by Storage Initializer can be customised per-deployment:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: rclone-sklearn
spec:
  predictors:
  - name: default
    replicas: 1
    graph:
      name: classifier
      implementation: SKLEARN_SERVER
      modelUri: mys3:sklearn/iris
      storageInitializerImage: gcr.io/kfserving/storage-initializer:v0.4.0    # Specify custom image here
      envSecretRefName: seldon-init-container-secret                          # Specify custom secret here

See our example to learn how to write a custom container that uses rclone for cloud storage operations.

Further Customisation for Prepackaged Model Servers

If you want to customize the resources for the server you can add a skeleton Container with the same name to your podSpecs, e.g.

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: sklearn
spec:
  name: iris
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - name: classifier
          resources:
            requests:
              memory: 50Mi
    graph:
      children: []
      implementation: SKLEARN_SERVER
      modelUri: gs://seldon-models/sklearn/iris
      name: classifier
    name: default
    replicas: 1

The image name and other details will be added when this is deployed automatically.

A Kubernetes PersistentVolume can be used instead of a bucket using pvc://.

Next steps:

You can also build and add your own custom inference servers, which can then be used in a similar way as the prepackaged ones.

If your use case does not fit for a reusable standard server then you can create your own component using our wrappers.

Handling Credentials

In order to handle credentials you must make available a secret with the environment variables that will be added into the Init Containers. For this you need to perform the following actions:

  1. Understand which environment variables you need to set

  2. Create a secret containing the environment variables

  3. Provide the Seldon Core Controller or Seldon Deployment with the name of the secret

1. Understand which Environment Variables you need to set

In order to understand what are the environment variables required, you can have a look directly into our Storage.py library that we use in our Init Containers.

AWS Required Variables

  • AWS_ACCESS_KEY_ID

  • AWS_SECRET_ACCESS_KEY

  • AWS_ENDPOINT_URL

  • USE_SSL

Minio Required Variables

  • AWS_ACCESS_KEY_ID

  • AWS_SECRET_ACCESS_KEY

  • AWS_ENDPOINT_URL

  • USE_SSL

Azure Required Variables

  • AZ_TENANT_ID

  • AZ_CLIENT_ID

  • AZ_CLIENT_SECRET

  • AZ_SUBSCRIPTION_ID

Google Cloud Required Variables

Currently for Google Cloud it is required to follow a slightly more complex method given that it requires the secret to be mounted as a file. For this please follow the example at the Google Cloud Section.

If application cretentials are not set, the client will use an Anonymous client.

2. Create a secret containing the environment variables

You can now create a secret, below we show what the env variables would look like for the AWS credentials.

apiVersion: v1
kind: Secret
metadata:
  name: seldon-init-container-secret
type: Opaque
data:
  AWS_ACCESS_KEY_ID: XXXX
  AWS_SECRET_ACCESS_KEY: XXXX
  AWS_ENDPOINT_URL: XXXX
  USE_SSL: XXXX

It is also possible to create a Secret object from the command line:

kubectl create secret generic seldon-init-container-secret \
    --from-literal=AWS_ENDPOINT_URL='XXXX' \
    --from-literal=AWS_ACCESS_KEY_ID='XXXX' \
    --from-literal=AWS_SECRET_ACCESS_KEY='XXXX' \
    --from-literal=USE_SSL=false

You can read the documentation of Kubernetes to learn more about Kubernetes Secrets.

3. Ensure your SeldonDeployment has access to the secret

In order for your SeldonDeployment to know what is the name of the secret, we have to specify the name of the secret we created - in the example above we named the secret seldon-init-container-secret.

Option 1: Default Seldon Core Manager Controller value

You can set a global default when you install Seldon Core through the Helm chart through the values.yaml variable executor.defaultEnvSecretRefName. You can see all the variables available in the Advanced Helm Installation Page.

# ... other variables
predictiveUnit:
  defaultEnvSecretRefName: seldon-init-container-secret
# ... other variables

Option 2: Override through SeldonDeployment config

It is also possible to provide an override value when you deploy your model using the SeldonDeploymen YAML. You can do this through the envSecretRefName value:

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: sklearn
spec:
  name: iris
  predictors:
  - graph:
      children: []
      implementation: SKLEARN_SERVER
      modelUri: s3://seldon-models/sklearn/iris
      envSecretRefName: seldon-init-container-secret
      name: classifier
    name: default
    replicas: 1

Examples

MinIO running inside same Kubernetes cluster

Assuming that you have MinIO instance running on port 9000 avaible at minio.minio-system.svc.cluster.local and you want to reference bucket mymodel you would set

AWS_ENDPOINT_URL=http://minio.minio-system.svc.cluster.local:9000

with modelUri being set as s3://mymodel.

For full example please see this notebook.

Adding Credentials for Google Cloud

Currently the Google Credentials require a file to be set up so the process required involves creation of a service account as outlined below.

You can also create a ServiceAccount and attach a differently formatted Secret to it similar to how kfserving does it. See kfserving documentation on this topic. Supported annotation prefix includes serving.kubeflow.org and machinelearning.seldon.io.

For GCP/GKE, go to gcloud console and create a key as json and export as a file. Then create a secret from the file using:

kubectl create secret generic user-gcp-sa --from-file=gcloud-application-credentials.json=<LOCALFILE>

The file in the secret needs to be called gcloud-application-credentials.json (the name can be configured in the seldon configmap, visible in kubectl get cm -n seldon-system seldon-config -o yaml).

Then create a service account to reference the secret:

apiVersion: v1
kind: ServiceAccount
metadata:
  name: user-gcp-sa
secrets:
  - name: user-gcp-sa

This can then be referenced in the SeldonDeployment manifest by setting serviceAccountName: user-gcp-sa at the same level as m̀odelUri e.g.

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: sklearn
spec:
  name: iris
  predictors:
  - graph:
      children: []
      implementation: SKLEARN_SERVER
      modelUri: gs://seldon-models/sklearn/iris
      serviceAccountName: user-gcp-sa
      name: classifier
    name: default
    replicas: 1