fix(kubeflow): tail only first and last node logs#538

Merged

ko3n1g merged 20 commits into

mainfrom

ko3n1g/fix/kubeflow-tail-first-last-rank

Jun 4, 2026

Commits on May 30, 2026

fix(kubeflow): stream only rank 0 + last rank, write all ranks to disk
ko3n1g
committed
fix(kubeflow): resolve last pod via completion-index label + full-history streaming

ko3n1g
and
claude
committed
fix(kubeflow): forward by global rank (node_rank*nproc+local), not pod heuristic

ko3n1g
and
claude
committed
fix(kubeflow): make TrainJob launch idempotent on 409 conflict

ko3n1g
and
claude
committed
fix(kubeflow): reload kube client across cert rotation for long runs

ko3n1g
and
claude
committed
fix(kubeflow): scope code_dir per job to avoid concurrent clobber

ko3n1g
and
claude
committed
fix(kubeflow): unique TrainJob name + forward all ranks (deduped)

ko3n1g
and
claude
committed

Commits on May 31, 2026

Commits on Jun 2, 2026

feat(kubeflow): add copy_to_workspace/copy_from_workspace for arbitrary PVC paths

ko3n1g
and
claude
committed