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/(.+)
)
The download is handled by an initContainer that runs before your predictor loads. This initContainer image uses our Storage.py library to download the files. However it is also possible for you to override the initContainer with your own custom container to download any files from custom resources.
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:
- Worked notebook
- SKLearn Server
- XGBoost Server
- Tensorflow Serving
- MLflow Server
- SKLearn Server with MinIO
You can also build and add your own custom inference servers, which can then be used in a similar way as the pre-packaged ones.
If your use case does not fit for a reusable standard server then you can create your own component using our wrappers.
In order to handle credentials you must make available a secret with the environment variables that will be added into the initContainer. For this you need to perform the following actions:
- Understand which environment variables you need to set
- Create a secret containing the environment variables
- Provide the Seldon Core Controller or Seldon Deployment with the name of the secret
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 initContainer.
- AWS_ACCESS_KEY_ID
- AWS_SECRET_ACCESS_KEY
- AWS_ENDPOINT_URL
- USE_SSL
- AWS_ACCESS_KEY_ID
- AWS_SECRET_ACCESS_KEY
- AWS_ENDPOINT_URL
- USE_SSL
- AZ_TENANT_ID
- AZ_CLIENT_ID
- AZ_CLIENT_SECRET
- AZ_SUBSCRIPTION_ID
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.
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
You can also create the secret with the following command:
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 create a Secret
object from command line by setting the exact environment variables:
and you can read more about interacting with Secret
object.
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
.
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
executor:
defaultEnvSecretRefName: seldon-core-init-container-secret
# ... other variables
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
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.
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