we’re definitely cpu bound afaict - and 10m of single thread perf may outweigh the overhead of serialization if optimized?