How are you handling the durable execution state? I built a similar orchestration layer for a book generation pipeline using Celery and Redis, but managing the websocket stability and intermediate state for ComfyUI was a nightmare. Curious if this is a custom implementation on top of Postgres or if you're using something like Temporal.

I am using Temporal, Durable execution state was a big problem for orchestrating a single backend, combining multiple backends meant also transferring artifacts between different servers which as you imagine required durability as well

Any reason why you did not end up using Temporal?