摘要：Consensus, or state machine replication is a foundational building block of distributed systems and modern cryptography. Consensus in the classical, "permissioned" setting has been extensively studied in the 30 years of distributed systems literature. Recent developments in Bitcoin and other decentralized cryptocurrencies popularized a new form of consensus in a "permissionless" setting, where anyone can join and leave dynamically, and there is no a-priori knowledge of the number of consensus nodes. So far, however, all known permissionless consensus protocols assume network synchrony, i.e., the protocol must know an upper bound of the network's delay, and transactions confirm slower than this a-priori upper bound. We initiate the study of the feasibilities and infeasibilities of achieving responsiveness in permissionless consensus. In a responsive protocol, the transaction confirmation time depends only on the actual network delay, but not on any a-priori known upper bound such as a synchronous round. Classical protocols in the partial synchronous and asynchronous models naturally achieve responsiveness, since the protocol does not even know any delay upper bound. Unfortunately, we show that in the permissionless setting, consensus is impossible in the asynchronous or partially synchronous models. On the positive side, we construct a protocol called Hybrid Consensus by combining classical-style and blockchain-style consensus. Hybrid Consensus shows that responsiveness is nonetheless possible to achieve in permissionless consensus (assuming proof-of-work) when 1) the protocol knows an upper bound on the network delay; 2) we allow a non-responsive warmup period after which transaction confirmation can become responsive; 3) honesty has some stickiness, i.e., it takes a short while for an adversary to corrupt a node or put it to sleep; and 4) less than 1/3 of the nodes are corrupt. We show that all these conditions are in fact necessary - if only one of them is violated, responsiveness would have been impossible. Our work makes a step forward in our understanding of the permissionless model and its differences and relations to classical consensus.
