Multiple issues with gke/deploy.sh

[gindeleo]

There are multiple small issues with deploying kubeflow to GKE via [deploy.sh] on HEAD.(https://github.com/kubeflow/kubeflow/blob/master/scripts/gke/deploy.sh)

1. KUBEFLOW_REPO must be an absoulute path to specify a preexisting local kubeflow package. Should this path be absolute as the default is set to kubeflow_repo (which then gets augmented to its abolutute path contrary to the externally set env variable? See:
   https://github.com/kubeflow/kubeflow/blob/master/scripts/gke/deploy.sh#L18

2. If GCFS_INSTANCE is not set but the GCFS instance already exists then the wait here will terminate the script as there is no PID existing. Checking upfront if the instance exist (as mentioned in @jlewi 's comment here should fix that.
   Also, how much time is saved by creating the GCFS instance directly (and in parallel/async) as opposed to via deployment manager?

3. Creating the local variable here: https://github.com/kubeflow/kubeflow/blob/master/scripts/gke/deploy.sh#L149
   raises an error as it's not called in a function.

[jlewi]

deploy.sh is deprecated on master; looks like I forgot to delete it.

You should be using kfctl.sh on master
https://github.com/kubeflow/kubeflow/blob/master/scripts/kfctl.sh

Instructions
https://master.kubeflow.org/docs/started/getting-started-gke/