What to do about example TFJob YAML specs?

[jlewi]

We currently have some example YAML specs here.

• Do we still need these?
• Should we turn them into ksonnet templates to facilitate running them as part of our E2E tests (see Use ksonnet to easily define TFJobs to be run as tests #374)?

[gaocegege]

I think some examples are outdated and we should remove them. Actually I think we have no enough contributors to maintain multiple examples. Maybe one is enough.

[wbuchwalter]

Also, this examples won't necessary work on all clusters. For example tf_job_gpu.yaml can only run on cluster with NVIDIA device plugin installed, and won't run on clusters still using alpha.kubernetes.io/nvidia-gpu resource such as Azure.

So we could either:

• Use ksonnet to parametrize this with a cloud param.
• Create different directories for different environments.

I like the second option for two reasons:

• I think examples should be very easy to read, and IMHO a simple yaml file is much easier to read/understand than a ksonnet templates.
• This would allow for environment specific examples, such as using Azure Files to save checkpoints on Azure, using S3 on AWS etc.

However like @gaocegege mentioned, this is more work.

[lovejoy]

https://github.com/kubeflow/tf-operator/tree/master/examples/crd https://github.com/kubeflow/tf-operator/tree/master/examples/v1alpha2/dist-mnist are useful. And I think there are many people don't use ksonnet. So it is better if there are some example yaml file.

[jlewi]

True; but we could use ks show to automatically publish our ksonnet examples as YAMl manifests. The advantage of doing it that way is

1. We have E2E tests for our ksonnet prototypes
2. Those ksonnet prototypes are kept up to date.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.