Kubeflow Trainer v2.2 Release

[andreyvelich]

Creating this issue to track Kubeflow Trainer v2.2 release.
Currently, we target mid February to prepare the first release candidate.

Milestone for the v2.2 release: https://github.com/kubeflow/trainer/milestone/7

• Support JAX Runtimes #2442
• Track TrainJob progress and expose training metrics #2779
• feat: Flux Framework as a plugin for HPC and MPI bootstrap #2841
• Extend the PodTemplateOverrides structure with "manager" field #2856
• TTL for TrainJobs #2899
• Support for Tensor Caching in Kubeflow Data Cache #3173
• Support XGBoost/LightGBM runtime and examples #2598

Please let me know if I miss any pending large features that we should complete.

cc @kubeflow/kubeflow-trainer-team @kubeflow/kubeflow-sdk-team @robert-bell @jaiakash @milinddethe15 @vsoch @akshaychitneni @kannon92 @mimowo @kaisoz @stivanov-intercom @XploY04 @nqvuong1998

[andreyvelich]

/area release

[Krishna-kg732]

Hi Andrey, quick question on the features list for this release — should we explicitly add XGBoost support there as well?

Given the ongoing work around Trainer/TrainJob improvements, it might be useful to call it out if it’s in scope. Happy to help clarify or update the list if needed.

[andreyvelich]

Sure, that would be awesome to include XGBoost runtime support into Trainer v2.2 release!
Let's see if we have time to finish the implementation.

[Goku2099]

Just sharing a data point from local testing in case it’s useful for the v2.2 release.
I rendered and linted all manifests under manifests/overlays locally (kustomize + kube-linter) and didn’t see any lint issues.
Didn’t find anything release-blocking on my side, but happy to check other areas if needed.

[katara-Jayprakash]

/assign

[katara-Jayprakash]

/unassign

[andreyvelich]

/milestone v2.2

[Goku2099]

@andreyvelich Tracking issue for Tensor Caching in Data Cache: #3173