GoogleCloudPlatform / kubeflow-distribution

Blueprints for Deploying Kubeflow on Google Cloud Platform and Anthos
Apache License 2.0
80 stars 63 forks source link

Seldon not available in KF 1.1 #146

Open jal06 opened 4 years ago

jal06 commented 4 years ago

I deployed Kubeflow 1.1 on GCP following the documentation

Then I wanted to run the simple example from this documentation

When creating the SeldonDeployment, I get the error below : error: unable to recognize "seldon-test.yaml": no matches for kind "SeldonDeployment" in version "machinelearning.seldon.io/v1"

If I search for seldon objects, nothing is found kubectl get all --all-namespaces | grep seldon

kubectl get all --all-namespaces | grep Seldon

I create an issue in the repo gcp-blueprints because it looks that seldon is not deployed when using blueprints for Kubeflow

or did I miss something ?

Bobgy commented 4 years ago

Is there a way to install Seldom in this existing cluster?

jal06 commented 4 years ago

I was expecting that Seldon should be installed in Kubeflow 1.1 because

Maybe I could try to install Seldon as explained here But I am afraid that it could be painful to maintain later if we install Seldon in the same cluster as Kubeflow.

That's why if Seldon is not deployed when using blueprints for Kubeflow, a better solution should be to deploy Seldon in another standalone K8S cluster. WDYT ?

Bobgy commented 4 years ago

/cc @jlewi Can you share some context about this change?

jlewi commented 4 years ago

Does it make sense to install both KFServing and Seldon by default?

In general, I think we need to move in the direction of making it more composable and letting people select the applications they want.

jal06 commented 4 years ago

@jlewi, I don't know yet which serving tool I could use. I tried to use the serving tools which are proposed in the documentation .

I tried TFServing and I had some problems (error 404 when sending a prediction request : issue 5396 ) I tried KFServing and I had also a problem (error 404 when sending a prediction request : issue 1199 ) That's the reason why I wanted then to try Seldon.

jlewi commented 4 years ago

@jal06 I think not including Seldon by default is working as intended. If you want to use Seldon with KF 1.1 then I think you can follow Seldon's instructions to install Seldon in your cluster.

If you point us at the docs we can try to get them updated.