Ray and Microservices


This is perhaps quite a broad question and I can drill down a bit more as part of the ensuing discussion.

What would be an approach to integrate a ML product that is cloud native and has a microservice architecture with the compute cluster managed by Ray?

I am trying to figure out integration points for such a project. More specific:

  1. Can you deploy services in the cloud (same VPC) that can communicate with the worker nodes or can communication only happen to the head node from the outside
  2. Is there a pattern where by microservice logic can be represented as an actor in the ray workflow that can then be serialized (state of the actor) when the cluster (ML training lifecycle is over) is shutdown.

Appreciate pointers or existing integrations that I can look at.


Hi Raghu, have you looked into Ray Serve - Ray Serve: Scalable and Programmable Serving — Ray 2.3.0 ?