Alex

Happy to contribute to this.

As a single data point, I prefer option 2. Internally, we have a `DaskClusterConfig` class with a `spec` method that just returns a dictionary in the format that `custom_cluster_spec` expects....

Here's what the `user` section of my `~/.kube/config` file looks like: ```yaml users: - name: user: auth-provider: config: client-id: client-secret: extra-scopes: id-token: idp-issuer-url: name: ``` In particular, please note the...

We use OIDC. The tokens are relatively long-lived and we provide an out of band auth CLI for when they expire. I'm not sure how unique our setup is, so...

Just found out that our refresh tokens are prohibited by design for security reasons. This will prevent us from using the SDK with the operator, unfortunately.

[Here's](https://github.com/kubeflow/pipelines/blob/master/backend/src/v2/compiler/argocompiler/argo.go#L192C11-L192C32) where the stuff that shouldn't happen happens.

In addition, if an identical component is used twice, e.g.: ```python @dsl.pipeline def pipeline(): for i in range(2): task = comp() ``` Each task gets a **dedicated annotation** even though...

Given the above, our proposed solution is to: 1. Find a better place to store the component logic that can still be referenced via `{{}}` but doesn't run into the...

We're going to try to store the component logic as (deduplicated) workflow parameters instead of annotations. This should resolve the 256 KB annotation limit and dramatically reduce workflow manifest size....

The new pre-commit hooks run black [here](https://github.com/kubeflow/training-operator/pull/2184/files#diff-63a9c44a44acf85fea213a857769990937107cf072831e1a26808cfde9d096b9R19-R29). We haven't added a linter hook yet, but plan to. Would you like to implement that, @Ygnas?