Hi team,

Thanks for developing TransferQueue — I’m working on a multi-process inference system and your framework is very relevant.

Quick question: In your architecture, have you compared using the Ray object store (ray.put() / ray.get() for shared memory IPC) versus socket/IPC (or storage + metadata) transport for moving large tensor data between processes (same node or across nodes)?

• Do you have benchmark data or evidence showing one approach is better than the other in your usage scenario?
• Which mechanism does TransferQueue currently assume or recommend (shared-memory via Ray, or socket/IPC)?
• If I’m building a system that transfers large batches across worker processes, would you recommend starting with Ray’s object store or custom IPC/sockets, given complexity vs performance trade-offs?

Thanks for your advice!
Best,