[PROPOSAL] In-memory Arrow Data Cache Solution for Distributed Training on Kubernetes

[Akshay Chitneni]

Hi Kubeflow community,

We are working on an in-memory cache solution based on Apache Arrow and DataFusion for ML workloads optimizing data loading of tabular data (e.g. Apache Iceberg tables) for distirbuted training on Kubernetes. We have an intitial proposal on the design of cache and integration with kubeflow trainer v2. Please share your feedback, and we will soon transfer this to a KEP.

We are also speaking at Kubeflow Summit 2025 in London next week and would love to meet and discuss if you are at the conference - https://sched.co/1u5iQ

Proposal doc: https://docs.google.com/document/d/1xj3K6bOT4f0EPiC4zwr2OsbRkROjIA8u_TBRZRVxrHI

Regards,

Akshay