This message was deleted.

on_deployment*

Hey, we would like to understand your use case a little bit more. What is the exact reason that you want to share data between processes? this is generally not recommended unless one know exactly what one is doing

Although we have found a workaround but to mention cases : 1. We wanted to download model from s3 once and store it at the start of the service and load the model. Currently every worker is downloading again and loading the model. 2. We wanted logging to be done in one file for every worker, for that we wanted to make a multiprocessing queue which can synchronize all workers to send the logs in one file. Currently since we can't write this before spawning processes, so we are unable to make one queue overall for the service.

BentoML has a idiomatic way in handling this which is to use BentoML models. the models will be packed together with the docker image. When the runners process starts it simply reads from the local bento store

Yes, that we have explored but we wanted to download model at runtime so didn't worked out like that.

I see. what are your current workaround?

Not a workaround exactly, but letting every worker to download their own model and load it.