perf: Reduce native shuffle memory overhead by 50% (#1452)

perf: Reduce native shuffle memory overhead by 50% (#1452) #2825