[Discussion] Arena project roadmap

[johnugeorge]

There was a discussion in the last community meeting regarding the future of Arena project in the Kubeflow Github org. Currently, the project is active with contributions, but the roadmap/goals are not documented.

Two options were discussed.

1. Add it as an official Kubeflow component as in https://www.kubeflow.org/docs/components/ . This will require a new Working Group(or join existing WG) with validated release for every official Kubeflow release, community support, documentation, website, roadmap etc.
2. Move Arena project to a separate Github repository outside Kubeflow.

We would like to know more on roadmap, active contributors and organizations who are willing to support.

cc @Syulin7 who is a lead contributor in the Arena project

[johnugeorge]

/cc @jbottum

[Syulin7]

Thanks @johnugeorge , I will discuss arena roadmap with contributors.

[terrytangyuan]

I had a discussion with Arena team. Here's an update:

1. Arena team has put together a 2024 roadmap document here: https://github.com/kubeflow/arena/blob/master/ROADMAP.md
2. The OWNERS file has also been updated to include the active maintainers from multiple companies: https://github.com/kubeflow/arena/blob/master/OWNERS
3. Arena team is interested in staying in Kubeflow and establishing WG CLI.

[ChenYi015]

Hi @johnugeorge, I am one of the maintainers of arena. We would like to propose a new CLI WG to sponsor arena and I am going to move this forward. I will raise a draft WG CLI PR to the community recently. Any guidance will be appreciated.

[andreyvelich]

Thank you for this @ChenYi015! Do we know the scope for Arena project ? Is it only to trigger model training jobs ? For example, we can expand the scope of Kubeflow Training WG.

[terrytangyuan]

Looks like there are commands for serving as well: https://github.com/kubeflow/arena/tree/master/pkg/commands/serving

[ChenYi015]

Do we know the scope for Arena project ? Is it only to trigger model training jobs ?

@andreyvelich Arena can also be used to submit serving jobs and do model management.

• For training jobs, supported types are tfjob, pytorchjob, mpijob, etjob, sparkjob, deepspeedjob, volcanojob, horovodjob. Ray jobs will also be supported in the recent future.
• For serving jobs, supported types are tensorflow, triton, kserve, seldon, custom.
• For model management, arena can be used to manage MLflow registered models and model versions.

For example, we can expand the scope of Kubeflow Training WG.

Since arena can do many other stuff besides submitting training jobs, I think it would be more appropriate to put it under a WG like CLI WG.