Great topic to kick off. The core challenge with replay systems in competitive titles always comes back to determinism—your entire replay pipeline collapses if the sim isn’t bit-identical across runs. Back in the arcades-to-early-PS2 days, you’d serialize the full CPU registers, memory map, and hardware registers into a “save state,” but replay systems took a different path: they record input frames plus periodic delta snapshots rather than full snapshots every frame, which cuts replay file size from gigabytes down to megabytes. The real killer for competitive titles is floating-point non-determinism across different GPU/CPU revisions—if your shaders or physics use IEEE 754 floats without fixed-point constraints, you get frame drift that makes rollback netcode like GGPO-based match reconstruction impossible. That’s exactly why modern titles that want reliable replay-driven anti-cheat and fair match review (think of the review systems in fighting games and tactical shooters) either lock to deterministic fixed-point math or run everything on a single reference hardware profile.
For next-gen architecture, I’d push toward hybrid snapshotting with delta compression layered on a deterministic lockstep sim, plus storing replay metadata as a lightweight index so spectators can scrub to any frame without decoding the whole file. On the hardware side, NVMe Gen5 speeds mean you’re no longer bottlenecked by write throughput when recording high-tick-rate (256+ tick) matches, so the design constraint shifts
