Kubeflow "Unboxing" - Part 5 - Installing 0.3.4 with kfctl.sh

So, let's uninstall all the pods we already have and try installing 0.3.4 with the instructions at https://master.kubeflow.org/docs/started/getting-started/.

Before starting the above, I wanted to return everything in my cluster to its pristine state. You might recall that our last attempt wound up creating everything in the default namespace. So, I went in and told it to delete all the pods. They deleted, then came immediately back. It turns out that in order to get rid of these, you have to delete their deployments first. So, I used "kubectl get deployments" to get a list of all the deployments and then deleted them one-by-one -- to avoid deleting my basic networking deployment from earlier.

This got rid of all the pods except tf-hub-0. After fighting this for 15 minutes, I found that it is created by a StatefulSet. So, I got the name of that stateful set using kubectl and deleted it, then deleted that last pod. Now I am back to running no pods.

---

So, running the commands at the link above seems to produce a veritable plethora of pods in the "kubeflow" namespace, but there are some gotchas along the way to flag up.

1) You need to do an export KUBEFLOW_src=directory_that_does_not_yet_exist first

2) You need to do an export KFAPP=different_directory_that_does_not_yet_exist second

3) When you run ${KUBEFLOW_SRC}/scripts/kfctl.sh init ${KFAPP} --platform none, you will get an error about "unrecognized option --platform"... just ignore this

4) If you get an error about a directory already exiting when you run ${KUBEFLOW_SRC}/scripts/kfctl.sh generate k8s, you can NOT ignore it - you will have to delete that directory and re-run this command. This is basically indicating that you still have an kubeflow installation directory from our previous attempt on your filesystem.