I think about 20-100 times faster. Not for the whole program, just for the data access part. People love the idea of 'zero copy' but 'zero allocation' is where the real gains are. Memory allocators are complex and sometimes slowish things. I've imagined a library that manages normally cloned data (mostly strings) by owning them in some global singleton object and handing out Arc references to them (actually has to be offsets, not actual references, if you want to persist and restart your code). Mainly so that I don't have to do .clone() and .to_owned() all the fricking time feeling like I'm slowing down the code, and so I can impl Copy on all the structs and as a side effect you could access them directly in LMDB without serialization. That system also has to persist to LMDB.