KEP-3311: Support shared initializer for train jobs
Open
kind/plan-kep
- Dominant language
- Go
- Stars
- 2.2k
- Forks
- 1.1k
- Avg merge
- 3d 22h
- Merged PRs (30d)
- 39
Description
Today each TrainJob initializes its own dataset and model independently. In multi-job workflows (HPO trials, iterative experiments, team-shared datasets), this means redundant
downloads, duplicated PVCs, and duplicated cache clusters — wasting bandwidth, storage, and compute.
This issue tracks proposal for a SharedInitializer plugin that lets multiple TrainJobs share a single initialized data source.
Contributor guide
Assessment
This issue has not been assessed yet.