TypeScript: dynamic-state snapshot for rollback netcode (takeSnapshot is ~19x larger than a rollback needs)

[jounih]

Motivation

For rollback netcode, takeSnapshot() is called every simulated frame — the state has to be captured before anyone knows whether a correction will arrive. That makes its size and cost the dominant fixed per-frame charge, and it is currently serializing far more than a rollback needs.

Measured with @dimforge/rapier2d-deterministic-compat 0.20.0 on a contact-rich scene:

bodies	takeSnapshot()	bytes/body	time
50	50.2 KB	~1.0 KB	0.035 ms
200	219.9 KB	~1.1 KB	0.079 ms
800	988.7 KB	~1.2 KB	0.404 ms

At 800 bodies that is ~1 MB serialized every frame, and it dominates allocation churn (~1.07 MB/tick, ~96% of it snapshot copies — around 64 MB/s of GC traffic at 60 Hz).

The dynamic state a rollback actually needs is small: position, rotation, linear and angular velocity, and sleep state — roughly 64 bytes/body, or ~19× less than the current whole-world encoding. Colliders, shapes, mass properties and the broad-phase structure are unchanged between two frames of the same world and are re-serialized every time.

The obstacle, as I understand it

Restoring only positions and velocities is not sound today, and I want to be explicit that I understand why rather than asking for something unsafe. The TGS solver warm-starts from cached contact impulses; a world restored without them re-simulates along a slightly different trajectory than the original pass, which in a lockstep/rollback setting is a divergence between the peer that rolled back and the peer that did not. That is presumably exactly why the snapshot is whole-world.

So the request is really one of:

1. A dynamic-state snapshot that includes the warm-start cache — positions, velocities, sleep state, and the contact impulse cache — but omits the static topology (shapes, colliders, mass properties, broad-phase tree) that a caller can guarantee is unchanged. Presumably with a validation handle so a mismatched restore fails loudly rather than corrupting.
2. Or a documented way to disable warm starting, so callers who accept the solver-quality cost can use a positions-and-velocities snapshot safely. IntegrationParameters currently exposes numSolverIterations, lengthUnit and predictionDistance through the TS bindings, but nothing for warm starting, so this cannot be traded off from JS today.
3. Or a snapshot API that writes into a caller-provided buffer, which would at least remove the per-frame allocation even if the size stayed the same.

Any of the three would help; (1) is the one that would make rollback cheap.

Prior art

Jolt exposes SaveState/RestoreState with a StateRecorderFilter so applications can omit state they know is unchanged — static or inactive bodies in particular — explicitly for network rollback. Something in that spirit for Rapier would fit rollback netcode very well.

I appreciate that whole-world snapshots are the conservative and obviously-correct design; this is a request for an opt-in fast path where the caller can uphold the extra invariants, not a change to the default. Happy to help test or benchmark.