There's decades of academic work in the form of "impossibility" results showing that handling byzantine fault tolerance in distributed consensus is alot more computationally expensive than omission/crash fault tolerance. The following (FLP impossibility result) being one of the most famous ones: