Release v0.7.0 requirements

[danielvegamyhre]

Targeting release in mid November, 2024.

Must haves

• Add global replica count label/annotation to support multislice training workloads with different templates #676
• Support for JobSet Preemption #682
• Blocking restart behavior #684

Nice to haves

• Investigate ways to validate Job templates at JobSet validation time #422
• Investigate restarting pods in-place as a mechanism for faster failure recovery #467
• Investigate using Elastic Index Job in exclusive placement implementation #482
• Comprehensive example task running training workload on GPUs using JobSet #429
• Comprehensive example task for running multislice TPU workloads with JobSet (and JobSet + Kueue) #428
• Add concept doc on how exclusive placement works #458

[ahg-g]

Another must have: #682

[danielvegamyhre]

Another must have: #682

Sounds good, added.

[kannon92]

I really think #463 could be ready for this release.

[danielvegamyhre]

I really think #463 could be ready for this release.

Agreed - added it.

[kannon92]

I think we should remove #463 and #482.

We should consider elasticity as a general KEP and maybe these two issues can be considered.

[danielvegamyhre]

I think we should remove #463 and #482.

We should consider elasticity as a general KEP and maybe these two issues can be considered.

Removed #463 (see my comment here regarding a KEP for elasticity). However, #482 isn't related to supporting use cases requiring elasticity, it's just changing an implementation detail for an existing API (exclusive job placement per topology domain) to improve efficiency, so I've left it in.

[danielvegamyhre]

Added #684 to goals for 0.7.0, cc @nstogner

[ahg-g]

I would like to move ahead with the release without #672 so that #672 can take its time in discussions. Any objections?

[danielvegamyhre]

I would like to move ahead with the release without #672 so that #672 can take its time in discussions. Any objections?

Yeah I agree, there are some fundamental design decisions that are still being heavily discussed. If we need to do a release to ship other features (e.g., RestartStrategy) to users soon, then we can aim to include #672 in the next release instead. I'll go ahead and remove it.

[ahg-g]

Are you able to actually do the release? :)

[danielvegamyhre]

Are you able to actually do the release? :)

I can probably do it this weekend. Has anyone manually tested the changes #682 (preemption) and #684 (blocking restarts)?

[ahg-g]

Yes, we tested both.

[danielvegamyhre]

Release 0.7 published