Is there a substitution of multiprocessing.Manager in Ray to share general python objects like dict and list with a server process?
My main purpose is to share a big dictionary between ray workers without copying the dictionary to each worker repeatedly, which relates to the issue#6979.
Hey, we don’t have a built in version of the multiprocessing Manager API, but you can achieve this using Ray primitives.
As you already alluded to, there are some performance subtleties (I’d recommend reading this: Serialization — Ray v2.0.0.dev0).
Could you provide more details about your workflow? (Maybe using Actor Pools could help with performance too Using Actors — Ray v2.0.0.dev0)