facebookresearch/LIGHT

LIGHT full redeploy checklist

JackUrb opened this issue · 0 comments

LIGHT Redeploy Checklist

Minimum Proof of Concept

  • build the new databases, such that they can be deployed on AWS
  • #297 (Update the main server to write to the EpisodeDB)
  • #298 (update main server to read from the UserDB)
  • #300 (Finish ModelPool abstraction to consolidate model initialization)
  • #302 (Make it possible to connect to hosted models).
  • Deploy new DBs to AWS
  • Deploy model servers
  • Create dummy WorldServer that wraps existing deploy. Deploy as a single instance.
  • Test end-to-end LIGHT with single world server and multiple model servers

Complete Deploy

  • Create LandingServer - split authentication out from WorldServer
  • Create server registry DB
  • Move static files out from deploy servers to scaling bucket
  • Extend WorldServer
    • Add functionality to self-register to the registry
    • Add functionality to remove from registry on graceful shutdown
    • Launch second world server, test e2e quick deploys
  • update all the routing to point to the LandingServer

Add world builder

  • Update the graph builders to point to the new EnvDB
  • Port the LIGHTDatabase out and over
  • Revamp world builder backend to map to EnvDB for loading components and saving graphs
  • Point the world builder to the model servers as well (maybe handled on the LandingServer?)