Question Does GetCollisionWorldFromTick() includes ghosts?

Ghosts are part of the physics world and the ghost colliders positions we backup is the last simulated physics frame.
Here is where problem starts, since without modification of the physics loop, server and client are moving and predicting objects differently.

The physics update is execute always after the GhostSimulationGroup. That means, whatever the ghosts does inside each of the prediction tick, is not reflect on the physics until the end of the prediction update.
On the client then, since ghost can predict "partial ticks" (ServerTickFragment < 1) if the physics does not runs inside the prediction loop and is aware of that (skip those updates) the position of the rigid body and its velocity may be simulated in a different way in respect to server.

All of that can cause some sort of problem with the collision history.
On the server, the data stored is always the result of full simulation update (so GhostPrediction + Physics with the frame delta time). And that is fine, even though you technically always raycast to the collider positions "in the past frame" kinda.
On the client, the data in the history buffer can be slightly in different position for the same ServerTick in presence of partial updates. Physics is unaware that the updated position for the ghost is only partial (delta time for the prediction step < delta frame time) and it still run a full physics loop.

Predicted object position (and data) are also restored from the last full predicted tick (GhostPredictionHistorySystem) when running the next partial prediction, but the collision world doesn't;
So when you check the raycast again the other player, bounding volumes might be way off.

At least that's my guess. A fact the tell me this may be the case is that:

witch suggests that on the server the collision checks are working as expect. Am I correct ?

I'm also assuming that:

• since you have your own velocity, your ghosts are all Kinematic rigid body (so physics is not involved in the collision detection) and you never set the PhysicsVelocity component (it is always 0). In case you do the latter, you must be aware that DOTSPhysics, contrarily to PhysicsX, will still move the Kinematic object (so integration still take place) but collision are not handled.
• since you mentioned collisionHistory.GetCollisionWorldFromTick, you are using the raycast inside the GhostPredictionSystemGroup and not inside the FixedStepSimulationSystemGroup.

For ticks that aren't the final tick you can probably raycast in the collision world with collisionHistory.GetCollisionWorldFromTick() (it should have stored good data).
But for the IsFinalPredictionTick if the ServerTickFraction < 1, you may want to try to raycast again the previous tick (PredictedTick - 1, that is the last full tick stored) instead.

In the upcoming version of NetCode we added some support (while not perfect yet) for running physics inside the prediction loop (with the correct ticking logic). That may definitively help solving many headache.

We also have some internal demo that use player prediction and physics (by colliding with other predicted objects and players in the scene) and collisions kinda works.
But we/I didn't try with raycast though, so I will play around that a little with the new setup

 

Thanks to you to prompt the problem out actually!
It would be great if you can tell me more about how you check for those collision or provide a small sample that use the same logic of yours.
So I will able to investigate a little more on my end