I worked on a big commercial MMO engine for about a decade. In our implementation, the player runs ahead on the client (client autonomous) but server verified

I worked on a big commercial MMO engine for about a decade. In our implementation, the player runs ahead on the client (client autonomous) but server verified (actions replayed on the server). The new authoritative server position for the player is sent back to the client, and the client replays whatever player movements last made since the response to a movement is heard back from the server, with the player resynchronizing to that point in time where you moved against server objects and playing forward from that point transparently; the client (and server) maintain a short queue of movement history for each moving entity. Thus, if you ran into something on the server but it didn't obstruct your movement much, you would tend to blip much less. The physics framerate was very low compared to graphics framerate, and there would be some degradation in updates received on the client by distance as throttled by the server based on area of interest management. Position updates represented most of the bandwidth of the game. Everything is UDP based with different forms of reliability options on top. NPCs were "server authoritative" and their actions are replayed on the client. Interpenetrations are resolved via rigid body physics resolution on the client if something blips, but the server is ultimately the source of truth (nothing can interpenetrate on the server), so if a rigid body resolution on the client doesn't resolve some condition, the eventual resynchronization of player position from the server would make it happen at some point. It worked out pretty well most of the time; certainly you can construct many scenarios where it goes awfully bad from the perspective of a client (on the server everything is always fine), but we preferred the illusion of immediate feedback/low latency versus this queuing up everything to take place N milliseconds in the future, and we didn't need exact reproducibility between clients, just eventual (and hopefully pretty quick) consistency.