A distributed consensus ensures a consensus of data among nodes in a distributed system or reaches an agreement on a proposal. This topic may be very familiar to any technicians that work with distributed systems such as HDFS, MQ, ZooKeeper, Kafka, Redis, and Elasticsearch. With the rapid development and the increasing complexity of distributed networks, developers have always been exploring possible solutions to solve this persistent problem in both theory and practice.
Next, with the rise of blockchain technology, especially public blockchains in open networks and private blockchains in permissioned networks, this consensus problem has once again received much attention and needs to be considered from a new perspective.
In this article, we will look into the problems and challenges of distributed consensus and corresponding consensus algorithms. We will also briefly analyze the applicability and limitations of these consensus algorithms and discuss the combination of these traditional consensus algorithms and new blockchain technologies. Later, this article concentrates on the consensus algorithm and mechanism in the public blockchain field from the perspective of the reliability of human beings. This article also considers the association between the distributed consensus algorithms in traditional computer science and the consensus mechanism in blockchain and shows how new consensus ideas can be seen in the public blockchain field.
To fully understand distributed consensus, we need to first build an understanding of the features of a distributed network. What are the main features and characteristics of a distributed network? Or what are some possible problems involved with a distributed network? Let's look into some of these questions in this section of this article.
First, let's consider cash faults. A crash fault in a distributed network often may be related to one of the following issues:
These above problems are common in distributed systems. They are essentially the inevitable risks caused by unreliable and unstable physical hardware in distributed systems. For example, networks, or communication channels, cannot always be stable and reliable. Disks on physical machines or CPUs will not always be of good condition. Therefore, it is safe to say that crash faults are the most basic and common type of faults to be solved in distributed systems.
The crash faults are based on a simple assumption: Either nodes do not work or respond normally, or although they work and respond normally, they cannot implement inconsistency, that is to say, being idle is okay for them, but they cannot commit some errors. Malicious nodes in networks may change and forge data at any time, making it harder to solve the consensus problem. These troublemaking problems that may change and forge data or response information are often refer Byzantine faults. The crash fault is called a non-Byzantine fault.
Byzantine originated from Lamport's paper. It is no exaggeration to say that Byzantine fault tolerance (BFT) is the most complex and rigorous tolerance model. By analogy, some number of generals plan an attack on a castle together and each general can choose to start the attack or retreat. However, to successfully take the castle, all the generals must act synchronously. Next, given that the generals are too far from each other to use direct communication, messengers are used to carry messages. However, messages are not reliable. They may successfully delivery messages after a very long time, fail to deliver messages or even change with messages. The generals may not be reliable, either, for example, one of them may be a traitor who does not act in accordance with the plan. The messengers in this story represent communication channels in distributed networks and the generals represent nodes.
The most critical problem that distributed consensus algorithms need to solve is the question of how to implement the certainty and consensus, so that reliable consensus results are returned across the entire distributed network, which can be full of risks and uncertainties. Naturally, it is relatively easy to solve crash faults. Algorithms used to solve this type of faults are called crash fault tolerance (CFT) algorithms or non-Byzantine fault tolerance algorithms. Byzantine faults may cause unauthorized changes, have higher complexity and are more difficult to solve. Algorithms for solving these problems are called Byzantine fault tolerance algorithms.
What is the boundary between the two types of fault tolerance algorithms? In what scenarios do these two types of faults occur? Is it really necessary to take unauthorized changes into consideration? The answers to these questions may depend on the actual network environments and business scenarios.
Generally, if a system is in a reliable internal network, we only need to consider the issue of crash fault tolerance (CFT). For example, for distributed components in many companies such as distributed storage, message queue, and distributed services, we only need to consider CFT. The reasons for this is as follows: The entire enterprise network is closed and protected by multiple firewalls, making external access and attacks unlikely. Individual nodes are deployed in a unified manner, and it is very unlikely that the machines and running software gets changed without proper authorization. The distributed network is relatively "pure" at this point, and we only need to pay special attention to the communication network and machine hardware. We need to consider network latency and instability as well as downtime and faults that machines may experience at any time.
Then, there's Byzantine fault tolerance (BFT), which related to the entire distributed network being evaluated in a larger environment. In addition to physical hardware, it is also necessary to take some "man-made" factors. After all, it is specific persons instead of machines that perform misconduct. Assume that a distributed network is relatively open, for example, a private network of tens of companies in a specific industry. Or assume a completely open network, for example, a network that anyone has access to. Node machines and software on these machines are deployed by individual companies or individuals themselves. If the benefit is tempting enough, a person may launch DDoS attacks on one of these nodes, making authorized, often malicious changes to software code and to the code execution logic, or even the data that is persisted on disks in the network. In such case, we face bigger challenges. In addition to the unreliable communication networks and machine hardware, we need to consider and deal with the "troublemakers" in the system.
To solve these problems encountered in actual scenarios, many computational scientists have conducted lots of theoretical studies. These theoretical studies may seem too abstract and cumbersome for engineering technicians, and some of these studies are about boring mathematical issues. However, these theories can provide significant guidance on how to solve these problems. Also, these theories show what the theoretical limitation of possible solutions are and which direction can be explored and which direction cannot work. Standing on the shoulders of giants, we do not have to spend all the energy making a "perpetual motion machine". Since most of you have some knowledge of these theories, let's just do a brief review.
As early as 1985, Fisher, Lynch, and Paterson published the impossibility theorem of distributed consensus. Earlier, we also have shown that a natural and important problem of fault-tolerant cooperative computing cannot be solved in a totally asynchronous model of computation. That is to say, in asynchronous networks, it is impossible to implement consensus algorithms that tolerate even a single node fault. Building on this, Byzantine failures are not taken into consideration in this theorem. It is also assumed that the network is very stable and that all messages are delivered correctly and exactly once. In this paper, we show the surprising result that no completely asynchronous consensus protocol can tolerate even a single unannounced process death. We do not consider Byzantine failures, and we assume that the message system is reliable—it delivers all messages correctly and exactly once.
Of course, this is only theoretical. It shows the theoretical limitations of solving these problems, but it does not mean that these problems cannot be solved in practice. If we are willing to relax restrictions and make some sacrifice, we can find practical and feasible solutions in engineering.
The most basic prerequisite of the FLP impossibility theorem is the asynchronous network model. What are the characteristics of an asynchronous model and a synchronous model respectively?
Fortunately, our real network environment is more similar to a synchronous model. Therefore, we can determine the maximum timeout based on experience or sampling. For example, you mailed a book to one of your friends. However, after three days, the book is still not delivered to your friend. At this point, your friend may find it hard to determine whether the delivery has been delayed or whether the book has gotten lost during the delivery process. However, if the book is still not delivered to your friend after one month, you and your friend can basically come to the conclusion that the book has gotten lost during delivery. This conclusion is based on our experience and statistics: An item can usually be successfully delivered within one to two weeks. An asynchronous model reflects the worst cases and the extreme situations of the inter-code communication. An asynchronous model has something in common with a synchronous node: a consensus protocol that works in an asynchronous model also works in a synchronous model. A synchronous model has modifications and restrictions on an asynchronous model so that the synchronous model is closer to the real scenarios and it is possible to solve the consensus problem in practice.
In addition, even in an asynchronous network model, FLP does not indicate that consensus is unreachable, just that it's not always reachable in bounded time. In practice, it is still possible to find solutions if the restrictions on the bounded time are relaxed.
According to research into DLS, consensus algorithms can be divided into three major types by network model.
From a different perspective, FLP actually covers three properties of a distributed system: safety, liveness, and fault tolerance.
FLP impossibility means that, in an asynchronous network, no distributed consensus protocol can meet the three properties at the same time. In a distributed system, node failures are almost inevitable. Therefore, fault tolerance must be taken into consideration. FLP impossibility means that any consensus protocol can only have either liveness or safety in addition to fault tolerance. In practice, we can often make some sacrifices. For example, we can sacrifice a certain degree of safety, which means that the system can always reach an agreement quickly but the agreement is not very reliable. We can also sacrifice a certain degree of liveness, which means that the system can reach a very reliable agreement but this process takes too long or an agreement can never be reached due to relentless debate. Fortunately, many practical scenarios show strong robustness, making it very unlikely for events to invalidate a consensus protocol.
In addition, FLP does not exclude randomized algorithms like Las Vegas. Many consensus algorithms adopt the Las Vegas algorithm to avoid the limitations imposed by FLP impossibility on deterministic and asynchronous networks. These non-deterministic consensus algorithms involve the Las Vegas rule: A consensus is always reachable in a network, but the time needed to reach consensus may be unbounded. When these type of algorithms are used, it is likely to reach a consensus decision in each round. The probability (P) of reaching consensus within T seconds will increase exponentially with the increase of T and will get closer and closer to 1. In fact, this method has been adopted by many successful consensus algorithms and is an escape hatch within the scope of the FLP impossibility. The bitcoin consensus mechanism described later in this article also adopts this method.
The well-known Consistency Availability Partition Tolerenace (CAP) theorem states explicitly from a different perspective that "Of three properties of shared-data systems (data consistency, system availability and tolerance to network partitions), one can only achieve two at any given time". CAP is very similar to FLP, but they are not exactly the same. They focus on different perspectives. Even very similar concepts do not have the exactly same meanings. For example:
Theoretically, only two of the CAP parts can be implemented. However, the boundary selection is not binary. The whole spectrum in between is useful, as it mixes different levels of Availability and Consistency usually yields a better result.
In practice, we often need to make some trade-offs based on actual business scenarios. For example:
Of course, these statements are not the final conclusion. Since individual systems are evolving constantly, a correct conclusion today may be not true tomorrow. To become better, a system has to keep exploring scenarios that are suitable and find an optimal balance point.
To process a variety of real and complex problems and challenges in distributed systems, many solutions have been developed based on theoretical guidance. This article does not describe the implementation details and specific differences of these algorithms. Instead, only a general introduction is given to make a comparison on the whole from a broader perspective.
One of the most famous distributed consensus algorithms is Paxos suggested by Lamport, though its complexity is also "notorious". Lamport proposed this creative mechanism that is practical and can be implementable through engineering and can ensure the consistency of distributed systems to the maximum extent. Paxos is widely used in many distributed systems, including Chubby and ZooKeeper. Basic Paxos (single decree, that is, to only agree on a value each time) has two roles: A Proposer can process client request and actively propose a proposal value. An Acceptor passively responds to the information sent by a Proposer, votes on proposals made, and persists values and states during the decision-making process. To simplify the model, the Learner role can be ignored. This does not affect the decision-making in the model.
As shown in the figure, the consensus decision-making process uses the two-phase commit protocol:
The implementability of Paxos in practice is also based on many assumptions and restrictions. Paxos can only be used to process CFT and cannot be used to process Byzantine failures. Therefore, it is a non-Byzantine fault tolerance algorithm. From the perspective of FLP, Paxos implements fault tolerance and safety and abandons liveness (safe but not live). That is to say, this algorithm may never end or reach consensus, though this case is very unlikely. From the perspective of CAP, Paxos only ensures C (consistency) and P (partition tolerance), but weakens the level of availability. To improve the availability of Paxos systems, we can increase the number of Learners.
Despite these shortcomings, Paxos is still reliable, effective and well-tested in practice. In essence, Paxos is a (dominant) distributed consensus protocol in asynchronous systems. The inventor of Chubby even said that "there is only one consensus protocol, and that's Paxos"—all other approaches are just broken versions of Paxos. Why Paxos is effective in practice is that conditions that may affect the liveness and availability of Paxos systems are usually not easily triggered. If those conditions do occur, the impact is also not very unacceptable.
Due to the complexity of Paxos, Ongaro presented a simpler algorithm in 2014—Raft. Raft is much easier to understand and implement in engineering. This is also the initial objective of Raft. Many easy-to-understand design details have been made on the condition that the functionality is not affected.
The Raft algorithm is a leader-based asymmetric model. A node in a system can only be in one of the three states at any point in time: leader, follower, and candidate. In the initial stage, all the nodes are followers. To become the leader, a node (follower) must become a candidate and launch a round of electoral votes. If the node does not receive enough votes, the node becomes a follower again. However, if it receives a majority of the votes, the node becomes the leader. If the leader encounters failures and finds that a new leader is elected after it recovers from failures, the original leader automatically goes back to the follower state.
Raft also introduces the Term concept to identify expired information in a timely manner. A term is similar to an epoch file in ZooKeeper. A term number increases monotonically over time and at most one leader can be elected in a given term. If the logs have a last entry with different terms, then the log with the later term is more up-to-date.
Raft also introduces the heartbeat packet and timeout. To maintain its authority, an elected leader must continuously send a heartbeat packet to the other nodes in the cluster. If a follower does not receive the heartbeat packet during a given election timeout, the leader is considered to have crashed and the follower changes its status to candidate and starts a leader election.
The leader election in Raft is implemented through the heartbeat and random timeout. Log replication is implemented through strong leadership: The leader receives the client command, appends it to its log and replicates the log to other followers. Raft ensures safety by only allowing a leader to decide whether to commit a log or not.
Election and replication will not be described here in detail. For more information about election and replication in Raft. Note that the leader election and the normal operations orchestrated by the leader are relatively simple. The leader change process is actually a bit more complex in Raft.
However, although the principle/mechanism of Raft is not exactly the same as Paxos, the problems that they solve and the trade-off policies that they adopt can be considered similar. That is to say, Raft can only solve crash faults, emphasizing fault tolerance, safety, and consistency and weakens the level of liveness and availability.
Although many discussions on BFT solutions have been conducted since the Byzantine Generals' Problem raised by Lamport in 1982, many solutions for these problems are inefficient, slow and complex. The situation improved in 1999 when Castro and Liskov presented the Practical Byzantine Fault Tolerance (PBFT) algorithm. PBFT is the first algorithm of its kind with the complexity reduced from the exponential level to the polynomial level. PBFT enables several thousand TPS and feasible solutions to nodes acting maliciously in practice. It is proven that the PBFT algorithm will work normally if the number of malicious nodes in a system is no more than 1/3 of the total nodes.
All the nodes in a PBFT system are ordered sequentially with one node being the leader node and others considered as backup nodes. All the nodes in a system communicate with each other and reach consensus based on the majority principle. Each PBFT consensus round is called a view. The leading node is changed during every view and can be replaced with a protocol called a view change if a certain amount of time has passed without the leading node broadcasting the request. This replica timeout mechanism ensures that the crashed or malicious leader can be detected and that a new view starts by re-electing a new leader.
As shown in the figure, five phases are experienced from the client launching requests to receiving responses. The consensus process adopts the three-phase protocol. The following content briefly describes the five phases:
PBFT implements safety based on an asynchronous network model, but it depends on message timeouts to perform periodic synchronization. Because the leader-plan solution is adopted, the message synchronization process is very fast and the sequential writing is also implemented. However, the leader re-election is difficult. A malicious leader may start to send a message when the time is very close to the timeout window, causing the system to be seriously slow. This disadvantage can be used to launch attacks against the network and make correctly running nodes look incorrect, causing endless leader election.
Compared with Paxos and Raft, PBFT can process more problems: In addition to crash faults, it can process Byzantine problems that may cause troubles and unauthorized changes. However, from the perspective of the trade-off policy adopted in PBFT, PBFT is still similar to Paxos and Raft. From the perspective of FLP, PBFT also emphasizes the fault tolerance and safety and weakens the level of liveness. From the perspective of CAP, PBFT emphasizes the tolerance to network partition faults and consistency and weakens the level of availability.
Despite these shortcomings, PBFT is still effective and feasible in practice if the number of faulty or malicious nodes is no more than 1/3 of the total nodes. PBFT is not the only BFT algorithm. Other BFT-like algorithms are also evolving, for example, Lamport once suggested BFT Paxos (an enhanced version of Paxos) to process Byzantine faults. Recently, the BFT Raft algorithm was presented based on the combination of PBFT and Raft. However, from the perspective of the problem scope and mechanism, these algorithms are still similar to previous ideas and frameworks. Introduction to these algorithms will not be described in this article.
From Paxos, Raft, and PBFT to various variants of Paxos and Raft and new BFT-like algorithms, distributed consensus algorithms have been developing, improving and evolving. Many big companies have also developed distributed consensus algorithms that meet their business scenarios. Although these algorithms are not very perfect, they play an important role in specific business practices. Then what are the application scenarios of these algorithms? What are the limitations of these algorithms?
Non-BFT algorithms such as Paxos and Raft can only process machine hardware faults and cannot handle situations where malicious nodes are present. These non-BFT algorithms can only run in very reliable network environments, for example, the internal network of a company. In such relatively closed networks, access requires strict authorization, ensuring that the identities of individual nodes are known and reliable. This serves to eliminate malicious nodes, and to allow the algorithms to run effectively.
BFT algorithms do have very strict requirements on the network environment. Even if malicious nodes are present, the entire system is still secure only if the malicious nodes are no more than 1/3 of the total nodes. However, this brings about new problems. How do you know exactly the number of malicious nodes in the network? What is the proportion of the malicious nodes to the total nodes? If access to the network requires permissions, it is relatively easy to solve this problem. For example, in a private network consisting of 10 affiliates, only the 10 authorized companies can access the network. Even if some companies (less than 3 companies) act maliciously and attempt to change data without authorization, the whole system is still reliable and secure. In such a permissioned network, the number of nodes that may act maliciously has been estimated. When some nodes do act maliciously, their real identities can be quickly located. This indirectly improves the safety of the network.
However, BFT algorithms may cause problems in a permissionless (open permission, without permission control) open network. If a distributed network is open and can be accessed by anyone and the cost of the network access is low, it is unknown how many potentially malicious nodes may be in the network. Even when some nodes act maliciously, determining their identities is also a difficult issue. A typical attack scenario is a Sybil attack where attackers can forge identities to control a large number of nodes and then the entire distributed network.
In addition, the biggest limitation of BFT algorithms is that they can only coordinate a small number of nodes (for example, no more than 100 nodes). If the nodes are in the thousands, the system shows very poor performance or even fails to reach consensus, affecting the liveness and availability of the system. You may have noticed that multicast is required throughout the three-phase protocol of PBFT: In the pre-prepare phase, the primary node broadcasts the request to all the secondary nodes; in the prepare phase, the secondary nodes broadcast to all other nodes; in the commit phase, individual nodes (primary and secondary) broadcast to all other nodes. From this process, we can know that the number of communication times is the number of nodes squared. When a system has a large number of nodes, this broadcast mechanism will be a disaster. The system is almost unable to reach consensus in a short period of time.
From the preceding content, we can draw a conclusion that traditional distributed consensus algorithms such as Paxos, Raft, and PBFT are generally applicable to reliable distributed networks that require permission control and have a small number of nodes.
In fact, these traditional consensus algorithms also gain new vigor in the age of blockchain: They are further understood and used. These consensus algorithms are widely used in private blockchain scenarios, where the network environment is relatively reliable. The application of a private blockchain is prospective due to the following characteristics:
In the current phase, private blockchains can be considered a good choice to achieve quick solution implementation and solve pain points in a specific industry. The application of private blockchains in the blockchain industry also indicates the further exploration of future blockchain development. Because authorization is required to join a private blockchain, a certain level of trust has been established in advance, and the network environment is relatively reliable. The probability of malicious behaviors and attacks in the network is extremely low and even if they should occur, responsibilities can be easily and quickly determined. Therefore, traditional consensus algorithms can also be applied in these scenarios. See the following examples:
Can the entire system reach consensus within a limited period of time if the network is completely open and permissionless and can be accessed by anyone at any time? How can we coordinate all the nodes in a network, for example, containing 10 thousand nodes instead of only tens of nodes?
Before answering the preceding questions, you actually need to ask yourself the following questions: Why does a network need to be fully open and permissionless? What kind of scenarios will require 10 thousand nodes? Does this node requirement really exist in actual scenarios? The answers to these questions are directly related to public blockchains in the blockchain. To answer these questions, we need to review the goals of distributed systems.
Why do we need a distributed system? This question is not difficult to answer. Generally, a distributed system can increase fault tolerance. After all, a distributed system depends on many different nodes, and the probability of failures occurring on all the nodes at the same time is much lower than that on a single node. In addition, a distributed system also supports attach resistance. Attacking or destroying many nodes is much more difficult than attacking a single node.
However, the preceding content is still limited to the physical hardware. These advantages can reduce the probability of failures in the physical machine hardware. However, man-made factors are not taken into consideration. If a system is important enough (for example, an electronic money system), in addition to machine failures, we need pay more attention to man-made factors. Will the person who deploys nodes intentionally act maliciously? How can we prevent corruption and collusion among nodes in the system?
Vitalik Buterin, the founder of Ethereum, mentioned the significance of decentralization, as shown in the following figure. A traditional distributed system implements architectural decentralization from the perspective of fault tolerance and attack resistance (how many physical machines are in the system and on how many machines can the system allow simultaneous failures). Similarly, now we need to consider how to implement political decentralization and collusion resistance. How many people or organizations ultimately control nodes in the system? How can we prevent corruption and collusion? If we say that a traditional distributed system focuses on the reliability of networks and machine hardware, then what we need to consider now is the "reliability": Can we find an effective technique to prevent man-made malicious actions? How can we ensure that the majority of the nodes in an important network are not maliciously controlled by one person or organization?
It is worth mentioning that this question is still highly controversial. Many people have never thought of corruption and collusion or think that it is completely unnecessary to consider them. Perhaps they think that there is nothing they can do with this technical problem. After all, this question exists far away from the real world that we live in. We live in a world where centralized platforms receive a good reputation, provide credit endorsement and control all the rules and processes. For example, few people worry that banks will intentionally make false accounting and misappropriate their assets stored in the banks. After all, banks are generally considered reliable. If banks are not reliable, people may not be able to carry out any business activity.
However, banks being reliable is just our assumption. We are left with the options "trust" and "suspect" and we have to choose the former because we cannot carry out business activities and the economic development will also stagnate if we do not trust banks. However, no practical methods can prove that banks are very reliable.
If reliability is actually necessary and meaningful, can you find a solution to make this world more reliable? Can you prove that a stranger that you're doing business with is reliable instead of having to believe that stranger is reliable? Don't trust; please verify. You do not need to trust that stranger and do not have to trust that stranger, either. You just need to verify that stranger.
To solve this problem, all people must be equal. Everyone can participate in the decision-making process equally and freely. Everyone can freely enter and exit the "council". This is actually technical democracy, which includes the following technical elements: The network must be permissionless and anyone can enter or exit the network at any time; nodes must be peers and can communicate directly; no intermediary or centralized authority exists (completely peer-to-peer); each node may become a bookkeeper.
Because the network is permissionless, fully open, transparent and democratic, the number of participating nodes may be very large and the probability of malicious nodes is also very high. Then, how do we coordinate the behaviors of nodes through a certain mechanism to ensure the consistency of the entire system, in this permissionless distributed network environment with a large number of nodes and a high possibility of malicious action? As mentioned before, consensus algorithms cannot do this. We need to seek new solutions.
In addition, decentralization may be the most controversial in the blockchain field. Some people think that decentralization is the value of blockchain, and the soul and premise of the existence of public blockchain, and the degree of decentralization of the system should be ensured as much as possible. While, others think that full decentralization is too ideal and unlikely to be achieved, so weak centralization or multi-centralization should be considered in combination with the actual situation, while giving consideration to the efficiency. Regardless of the value judgment, purely from a technical perspective, the higher the degree of decentralization, the higher the security of the system. Therefore, in the public blockchain system design, the degree of decentralization of the system should be ensured as much as possible. However, in conjunction with Vitalik Buterin's interpretation of the meaning of decentralization, in the process of pursuing decentralization, we should not stay at seemingly decentralized, but should comprehensively consider all dimensions of decentralization, and make the necessary trade-off according to the actual situation.
The innovative solution to distributed consensus in open networks is the Proof-of-Work (PoW) mechanism in Bitcoin.
On October 31, 2008, Satoshi Nakamoto published the Bitcoin white paper "Bitcoin: A Peer-to-Peer Electronic Cash System", which miraculously provides a creative solution to such problems, making it possible to coordinate thousands of nodes in a complex network environment. In fact, Satoshi Nakamoto did not publish the white paper to solve this technical problem. On the contrary, Satoshi Nakamoto thinks bigger. He creatively invented Bitcoin, a completely peer-to-peer electronic cash system, to eliminate the trusted third-party middlemen that traditional payment needs to rely on. In the process of implementing the system, the problem of consistency among multiple nodes in the open network happened to be solved. It can also be said that, the core problem Bitcoin solves is the Double Spending problem of electronic currency in peer-to-peer networks. However, the Bitcoin implementation mechanism is not only about distributed network technology, but also combines cryptography, economics, game theory, and other ideas, and achieves consistency among nodes in a non-deterministic probability mode. Therefore, simply calling it an algorithm is no longer able to accurately express its meaning. It may be more appropriate to call it a consensus mechanism, because its implementation indeed relies on a complete set of policies and systems. Here, we will not elaborate much on the ideological significance and implementation details of Bitcoin, but focus only on the implementation of its consensus mechanism.
A bitcoin is actually an electronic chain of digital signatures. The coin owner can transfer the coin by signing the hash value of the previous transaction and the public key of the next owner, and adding these to the end of the coin. The payee verifies the chain formed by the coin owner by verifying the signature. However, the problem is that the payee cannot verify whether the coins received have not been double spent, that is, the owner may have transferred the same coins to two people. Therefore, we need a mechanism for the payee to ensure that the previous coin owner did not transfer the coins to others before that. To ensure this, the only way is to let everyone know all transactions. In the absence of a trusted third party, to achieve this, all transactions must be broadcast to everyone. Therefore, we need a system where all participants agree on the order in which they receive the coins to form a unique sequence record history. This is actually a problem of distributed consensus.
The solution provided by the Bitcoin system is using a timestamp server composed of all nodes. The timestamp server timestamps the hash of transaction blocks, and broadcasts it. Each timestamp contains the previous timestamp in its hash, forming a chain, with each additional timestamp reinforcing the ones before it. To implement a distributed timestamp server in a peer-to-peer network, Bitcoin uses the Proof-of-Work (PoW) mechanism. When a hash operation is performed, PoW needs to find a certain value so that the first few digits of the overall hash value are all zero, and the average workload increases exponentially as the number of zero digits increases. In addition, the hash has no rules. To ensure that the first few digits of the hash are zero, the only way is to force random trial and error over and over again. Once sufficient CPU computational power is consumed and the hash value that meets the condition is found, the block cannot be changed, unless CPU is used again to redo it.
In addition, PoW solves most decision-making problems. In Bitcoin, the longest chain represents most decisions. If most of the computational power is controlled by honest nodes, the honest chain will grow rapidly and surpass other chains. If an attacker wants to change a previous block without proper authorization, the attacker must redo the PoW tasks of the corresponding block and all its subsequent blocks, and then catch up with and surpass honest nodes. This is very difficult. Mathematically, it is not difficult to prove that with the increase of the number of nodes later, the probability for slower attackers to catch up with the honest nodes decreases exponentially. Generally speaking, it is almost impossible to catch up after 6 blocks. In addition, the difficulty of PoW tasks is not fixed, but it is dynamically adjusted using the Moving Average method, which mainly takes into account the increase of the hardware computing rate and the increase or decrease of the number of miners. If the computing rate is fast, the difficulty increases, the computing rate is slow, and the difficulty decreases. Through dynamic adjustment of the difficulty, the block generation time of Bitcoin is roughly stabilized at about 10 minutes.
The entire network runs as follows:
Details of the transaction and mining are not described in detail here. But, in short, what we can say is that, in Bitcoin, the longest chain of information always prevails. If a node finds a chain longer than itself, it will automatically switch to the longest chain to work.
Since the cost of PoW is high, how can we encourage everyone to contribute their computational power and become nodes to ensure the safety of the entire Bitcoin network? Bitcoin provides two incentive policies:
These incentive policies also implicitly encourage the node to be honest. If a greedy attacker has more than half of the CPU computational power, he has to make a choice: whether to change the transaction records and transfer back the bitcoins he has spent, or to dig for new coins and earn service fees honestly? It is very likely that it is more advantageous to mine honestly. After all, the coins that he can earn are more than the total coins earned by all other nodes. While, destroying the Bitcoin system will also undermine the effectiveness of his own wealth. If the bitcoin is no longer reliable, the value will collapse quickly. In addition, an attacker is not as free to manipulate, change without authorization, or forge transaction records as one might think. All he can do is steal back the bitcoins he has recently spent.
Bitcoin has been running steadily for 10 years without any organization or group to maintain it, relying only on voluntary maintenance of community volunteers. No major problems have ever occurred during this period. This is a miracle and is sufficient to prove the effectiveness of the consensus mechanism behind Bitcoin. Why can Bitcoin do this? Why is the consensus mechanism behind Bitcoin so effective? The Bitnodes data shows that the number of Bitcoin nodes exceeds 10 thousand (Bitcoin has many types of nodes, and the number may be different for nodes with different calibers. Only full nodes are considered here). Why can Bitcoin coordinate tens of thousands of nodes in a permissionless network environment?
In my humble opinion, the following reasons may apply:
The Bitcoin consensus mechanism is no longer confined to the level of distributed algorithms, but contains more ideas, such as economics, game theory, and probability theory. Therefore, it may be more appropriate to call it the consensus mechanism. However, we can still put the Bitcoin PoW consensus mechanism into the framework of consistency. From the perspective of FLP and CAP:
On the whole, the Bitcoin PoW consensus mechanism has achieved a good compromise and trade-offs under the limit of FLP and CAP. In practice, it does provide a feasible solution to the distributed consensus problem in open and complex networks. The stable and reliable operation of Bitcoin over the past decade has also proven this point.
In addition, Bitcoin PoW has been rigorously analyzed and proved in a study by Miller and LaViola to have the following characteristics:
The PoW mechanism is not only reliable in practice, but also can be tested theoretically The PoW mechanism uses the synchronous model and random probability to circumvent the impossibility theorem of FLP deterministic asynchronous model. Compared with the complexity of PBFT algorithm O(n2), the scalability of PoW independent of network size has great advantages: The more nodes, the safer the system. And, the efficiency of the system has not decreased.
What is the magic of the PoW mechanism? In fact, as you may be aware, the idea of PoW is not profound, and in fact it was not first proposed by Satoshi Nakamoto. As early as 1993, this idea was proposed to combat junk mail (Pricing via Processing or Combatting Junk Mail). However, it was not widely used until Satoshi Nakamoto created Bitcoin. The essence of PoW lies in intentionally creating obstacles and increasing the cost of participants, so as to minimize the malicious attempts of participants. For example, the requester is required to do some additional work to detect DDoS attacks and junk mail. For another example, it is very common that a verification code need to be entered when logging on to a website, which is also to increase the login cost and prevent the website from being attacked. The core features of such tasks are non-symmetric: For service requesters, it must be difficult to complete the task. For service providers, the verification task must be simple and fast. For Bitcoin PoW, it is non-symmetrical: It requires a lot of computational power to find a nonce (random number) that makes the hash conform to the condition, after continuous trial and error, while verifying whether the found nonce conforms to the condition only requires a simple hash operation verification.
Bitcoin PoW is essentially one-CPU-one-vote. Why CPU instead of IP address? This is still based on the difficulty of the task. If one-IP-one-vote is used, the system can be easily controlled by people with a large number of IP addresses (such as IP providers). Relatively speaking, at least at the time when ASIC or FPGA is unavailable, the CPU was still relatively expensive hardware, and it was not easy to have a large amount of computational power (CPU + Electric power). This actually implicitly provides a real-world anchor for the value of Bitcoin: The virtual currency system finds the value anchor of the real physical world through computational power, although it seems to many people that this consumption of computational power is meaningless and a waste of energy.
Many people are thinking about how to reduce Bitcoin mining costs. This kind of thinking certainly has positive significance. The cost of PoW needs to be appropriate: If the difficulty and the cost are both too high, it indeed will waste more energy, but the security of the Bitcoin network has also been improved. If the difficulty and the cost are both too low, the purpose of preventing attacks will not be achieved, and the security of the Bitcoin network will be reduced. This is actually a trade-off issue and also a subjective value judgment, depending on the public's understanding and positioning of Bitcoin. Value judgment is always full of subjective prejudice. At present, the debate on Bitcoin is so great, because the public has not yet reached a consensus and is yet to come up with a common vision for the future of Bitcoin.
In short, Bitcoin PoW is a complete set of mechanisms, including technical trade-offs, and economic and gaming considerations, which jointly ensure the security and reliability of the Bitcoin network.
Everything is not perfect, and the PoW mechanism also has its limitations without exception. In fact, we can know a thing or two from the many criticisms of Bitcoin. Generally, the PoW mechanism is considered to have the following limitations:
Among these new solutions, Proof-of-Stake (PoS) undoubtedly attracts the most attention. Also facing the consistency issues in open and complex networks, it proposes a brand-new solution.
In 2011, a user named QuantumMechanic took the lead in putting forward the idea of proof-of-stake in BitcoinTalk Forum. After that, the idea has been continuously developed and perfected, and has been trusted by more and more people.
The basic concepts of PoS are as follows:
It is easy to see that PoS also adopts ideas from economics and game theory, and ensures the security and reliability of the network through incentive policies and punishment mechanisms.
The PoS protocol is still in line with the conclusions of the traditional Byzantine Fault Tolerance (BFT) algorithm. Currently, the research on PoS can be divided into two main lines: a line for the synchronous network model, and another line for the partial asynchronous network model. The chain-based PoS algorithm is almost always dependent on the synchronization network model, and its effectiveness and security can be strictly proved like the PoW algorithm. For reference, see this document.
In addition, from the CAP perspective, the chain-based PoS algorithm is similar to the PoW algorithm, which also achieves fault tolerance as much as possible, and availability is more ensured between availability and consistency.
If the traditional consensus algorithms (Paxos, Raft, and PBFT) implement deterministic finality or consistency, then PoS is similar to PoW, and seeks probabilistic eventual consistency instead. From the perspective of traditional CAP, this is actually a weakening of consistency. However, from the perspective of practical feasibility, it is also a brand new thinking and breakthrough.
From the design strategy of PoS, it can be divided into two types. For reference see, this paper.
The idea of PoS is not complex, and what is more easily criticized is precisely the system that is similar to the real world and obtains income according to the proportion of capital contribution. People are already wary of the Matthew Effect in the real world. This system leads to the result that the rich get richer and the poor get poorer: People with more tokens will have more opportunities to become the validator, thus participating in the network and earning more.
However, the views on this issue are very controversial, and many people have put forward completely different views, believing that PoS is fairer and more conducive to countering the centralization trend than PoW. The main reason is that PoW mining relies on physical hardware and electric power resources in the real world, which easily leads to Economies of Scale. Companies that buy 10,000 mining machines have more bargaining power than individuals who buy 1 mining machine, and even can develop their own mining machines with lower costs. And, mines with 10,000 mining machines have higher bargaining power over electricity charges, and can be migrated near power stations in countries and regions with low electricity charges, or even build their own power stations with lower costs. The consequence is that the larger the organization, the lower the comprehensive mining cost, which is exactly what has really happened in the real world. In contrast, PoS does not need to rely on hardware in the real world, and has no Economies of Scale advantage. If price manipulation is not taken into account, the price of 1 coin and the price of 10,000 coins increase linearly. From this perspective, PoS may be fairer and more conducive to decentralization.
Another concern for PoS is its security. After all, PoS no longer performs complex CPU operations like PoW to prove itself. In PoW, if an attack is to be launched, 51% of the computational power needs to be controlled (recently, some studies have found that only 25% of the computational power is required to make the attack possible), which means that most mining machines and computational power resources are required. In PoS, if the whole system is to be controlled, 51% of all tokens are required. Which one is safer? In fact, it is not easy to say, but from a real-world example, if the Bitcoin algorithm is switched to PoS, it takes about half of the Bitcoin market value to control the Bitcoin system, which is about 40-160 billion US dollars (the price range of a bitcoin: 5,000-20,000 US dollars). This figure is far higher than the cost of mining machines. It is almost impossible to launch an attack with such a large amount of money. In this regard, PoS may be safer.
In addition, due to the low cost of PoS deployment (low hardware requirements), tokens can easily be forked in the real world, resulting in a pile of altcoins. This problem does not exist in PoW. PoW relies on hardware for mining, so it is easy to change a parameter of Bitcoin. However, if you really want to run it, it requires a lot of computational power and the support of a large number of miners. For example, forking Bitcoin Cash from Bitcoin has experienced twists and turns. PoS has no such concern at all. Anyone can download the open source code and change it at will. By winning the support of a few nodes, they can claim that they have created a brand new token. For example, dozens or hundreds of altcoins can be easily forked from EOS (a token name), each of which claims to be unique. This is indeed a fact, but it is not easy to say whether it is good or bad.
The most important part of the PoS mechanism is the selection mechanism of the validator or creator in the next block. Who will be the lucky one? The above-mentioned selection according to the proportion of account funds and the probability is actually the simplest way. This method is indeed easy to lead the rich to obtain the income once and for all, thus compromising the enthusiasm of other participants in the network. Currently, many ideas are available to ameliorate this problem, among which the more interesting one is the coin age-based method. When the creator is selected, not only the amount of funds, but also the coin age will be considered. The so-called coin age refers to the retention time of the coins on an account. For example, if 1 coin is transferred to a specified account for 10 days, it can be considered that the coin age is 10, and the coin age will be recomputed from 0 every time the account in which the coin is located changes. In this way, it is possible to restrict the nodes with large amount of funds from becoming the creator frequently. For example, it can be set that only nodes with a coin age of 30 will have the chance to become the creator, and the coin age will be cleared immediately after the node becomes the creator. This actually limits the interests of large participants, and provides more opportunities for small and medium-sized participants.
The famous solution improved base on PoS is Delegated Proof-of-Stake (DPoS), in which the proxy delegation mechanism is adopted. In DPoS, it is no longer possible for all nodes to become the creator. Instead, the nodes vote for each other, and only the nodes with the highest votes can participate in the block creation process. The details are as follows:
It is not difficult to find that DPoS has ensured the extensive participation of nodes as much as possible by introducing the voting mechanism. And, the system is as efficient as possible by limiting the number of validators (generally 21-101). Despite the great controversy, DPoS is still a feasible solution, and more and more blockchain systems are also trying to improve and explore it.
In the public blockchain, many projects adopt the PoS mechanism. The famous ones are:
Ethereum 2.0 layers and phases
In fact, the rise of PoS mechanism is not only due to its own features, such as low cost, high efficiency and decentralization, but also because it opens the door to a wider array of techniques that use game-theoretic mechanism design in order to better discourage centralized cartels from forming and, if they do form, from acting in ways that are harmful to the network.
With the rapid development of blockchain, especially the public blockchain, in recent years, other Proof of X? Such mechanisms are also emerging. From the many mechanisms above, we can see the shadows of the PoS concept, that is, how to design a system from the economic and game theory perspectives to ensure decentralization, security and efficiency as much as possible. The following is a brief description of these mechanisms:
It can be seen that although the Proof-of-X? Mechanisms emerge one after another and are different from each other, the core essential problem to be solved is the same: Who will be the lucky one to keep accounts? These mechanisms just adopt different policies to develop game rules, so that nodes can prove themselves as fairly as possible, and the lucky ones can be chosen fairly. All these policies, include CPU computational power, number of tokens held, storage space size, random wait time, number of tokens burned, node activity, and node contribution, are all explorations for consistency issues in open networks under specific scenarios.
From PoW to PoS, to Proof of "X (Everything you can think)", the issue of consistency in the permissionless network has been under exploration. The connotation of "consistency" is also changing, from how to prevent the faults of network and machine hardware to ensure the data consistency between network nodes, to how to prevent the malicious actions of people in the open network to ensure the true consistency of data between nodes. It can be said that we have moved from the reliability of hardware to the "reliability of people," and public blockchain technology is also regarded as the "trust machine." However, the issue of "reliability of people" is too complicated, and even beyond the technical scope. What can be done at the present stage is far from ensuring the "reliability of people." However, in most cases, it is still at the stage of people's trust in machines and in protocols. Fortunately, we finally took this step and began to face up to this thorny problem to explore innovative solutions.
The world is full of uncertainty, and the same is true for computer science. Since the emergence of computers, we have to face the uncertainty of machine hardware: problems that may arise from unexpected failures. And, since the rise of the Internet, we have to face the uncertainty of the network: the possible latency, disorder, and loss of communication messages. The most natural solution to the uncertainty problems is redundancy. A large number of nodes are used to ensure the overall security of the system, to avoid the single point of failure (SPOF) and enhance fault tolerance and attack defense capabilities. It is on this basis that large-scale distributed networks are booming. The way to find certain certainty between uncertain networks and nodes, and coordinate the consistency among many nodes, is exactly the problem that distributed consensus algorithms need to solve. The CFT algorithms capable of dealing with fault type errors include the most classic Paxos algorithm and the simpler Raft algorithm. The effectiveness of the algorithm can be guaranteed when the normal nodes in the network exceed half. These algorithms are usually used in closed networks with trusted environment to coordinate the consistency between several to dozens of nodes, such as distributed storage, distributed service protocols, and distributed message systems within the company. In addition, they can also be applied to a private chain network composed of a few organizations that require authorization to access.
However, it is not only the network and the machine itself that are uncertain, but also the behavior of people controlling nodes in the network. The way to ensure the consistency of the distributed network under the condition that the attackers may change the data or attack the network without authorization, is exactly the problem that the BFT algorithm needs to solve. The most common BFT algorithm is the PBFT algorithm. The effectiveness of the algorithm can be guaranteed when the normal nodes in the network exceed 1/3. Even so, the ability of PBFT to deal with malicious behaviors in the network is still limited, and its performance will also decrease significantly as the number of nodes in the network increases. These limitations also lead to the PBFT algorithm can only be used in the permissioned network with more reliable environment, to coordinate the consistency between several to dozens of nodes, such as in the private chain scenario.
However, in a permissionless open network, the uncertainty problem is more severe, especially the uncertainty of the behavior of people behind the network nodes. The way to prevent the controllers in the network from forming a magnate through corruption and collusion, thus controlling more than half of the nodes in the network and achieving the purpose of controlling, damaging and attacking the network, is the problem to be solved in the open network. From this perspective, consistency in the open network also implies the premise of security: It requires not only that a consensus can be reached among nodes, but also that the consensus is indeed formed by the true expression of many node controllers. To achieve this consistency and security, it is not only necessary to realize the structural decentralization of physical hardware nodes, but also to ensure the decentralization of the actual controllers behind the nodes as much as possible. For this, it is necessary to: ensure that anyone can deploy and run the network protocol at any time to become a node in the network, and can access the network at any time; ensure peer-to-peer communication between nodes, without any centralized control nodes; and ensure that the roles of nodes are completely equivalent, and all nodes can participate in accounting fairly according to the rules. The way to coordinate the behaviors between tens of thousands of nodes in an open network to ensure the consistency and security of the network, is the problem to be solved by the public blockchain consensus mechanism. Among them, the most typical is the PoW consensus mechanism initiated by Bitcoin, and the subsequent PoS consensus mechanism. These consensus mechanisms are no longer limited to the technical consistency, but more of the ideas of economics and game theory are introduced to ensure the consistency and security of the network as much as possible from the perspective of economics and game theory.
From the consistency in the closed distributed network environment, to the consistency in the permissioned private chain scenario, and to the consensus mechanism in the permissionless public blockchain open network environment, the problems are becoming more and more complex, and the challenges are becoming more and more severe. From a purely technical perspective, the research on consensus is the same strain. These consensus algorithms or consensus mechanisms are also restricted by the FLP impossibility and CAP theorem in the traditional distributed consensus theory research. Paxos, Raft, and PBFT all emphasize fault tolerance and safety/consistency, and weaken liveness and availability. PoW and PoS consider the problem from a brand-new perspective, ensuring fault tolerance, liveness and availability as much as possible, abandoning the pursuit of certainty for safety and consistency, and only pursuing the eventual safety and consistency in a probabilistic way.
In addition, the thinking on consensus is constantly deepening, from the simple data consistency between nodes to the emphasis on the consensus and identification between people behind the nodes, and from ensuring the reliability of the network and hardware to ensuring the reliability of the people behind the nodes that make up the network as much as possible. Although the reliability between people is very complicated and goes beyond the pure technical scope, it is gratifying that we are already on the road, and the ongoing innovative and active exploration in this field will certainly make the world more reliable.
Alibaba Clouder - September 26, 2018
Alibaba Clouder - September 2, 2019
Alibaba Clouder - December 4, 2018
Alibaba Clouder - May 31, 2018
Alibaba Clouder - June 11, 2018
Alibaba Clouder - August 12, 2019
BaaS provides an enterprise-level platform service based on leading blockchain technologies, which helps you build a trusted cloud infrastructure.Learn More
Easy-to-use high-performance virtual machines with data transfer plan starting from $2.50 a monthLearn More
Link IoT Edge allows for the management of millions of edge nodes by extending the capabilities of the cloud, thus providing users with services at the nearest location.Learn More
High Performance Computing (HPC) and AI technology helps scientific research institutions to perform viral gene sequencing, conduct new drug research and development, and shorten the research and development cycle.Learn More