Roadmaps for 1.4 release

[Jeffwan]

Training Operator 1.3 operator has been released for few months. It's time to build wishlist for the next release. Community is collecting roadmaps for kubeflow 1.5 release as well (Jan or Feb?). kubeflow/community#535

I think for the next release, we can put more time to build a decent elastic training story.

• @gaocegege is working on PyTorch parts and the large PR has been merged.
• On the other hand, mpi-operator v1 has been integrated into training-operator and we can enrich the elastic work (expose arbitrary worker to scale in instead of just operating the numbers) based on what @zw0610 did in the past.
• If we do have additional time, we can revisit tensor flow elastic training story.

There's some more meaningful tasks like GenericJob which can support flexible framework is nice to have. Supporting different gang definition maybe something worth to explore as well. Feel free to brainstorm the ideas and we can summarize a roadmap and then recruit contributors and release managers.

/cc @kubeflow/wg-training-leads

[terrytangyuan]

Notes from today's community meeting:

• Training Operator ROADMAP for Kubeflow 1.5: Training Operator Roadmap presentation at Kubeflow Community meeting on Nov 23, 30 or Dec 7 community#535
  • We will present it at the next community meeting.
• Draft roadmap and initial thoughts Roadmaps for 1.4 release #1496
  • Create draft presentation based on above issue
  • Support MPI Operator v1 in universal Training Operator
    • V1 MPI controller - completed but need to double check with contributors whether manifests are ready for KF1.5
    • V2 MPI controller - not migrated to common operator yet Migrate v2 MPI operator to the unified operator #1479
• Better test coverage:
  • Enable tests for all frameworks #1311
  • Improve test coverage  #1497

[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.