Paxos algorithm

Paxos解决的是什么问题

state machine replication

如上所示的系统中，如果客户端要执行某个命令，那么将遵循如下的步骤：

1. 发送命令到其中某一个server
2. server首先将命令记录到log中，然后将命令发送到其他的服务中；其他server同样将其记录到log中
3. 当命令完整的记录到各个server之后，就可以传到state machine去执行了，并将结果返回给客户端

其中，consensus module用来保证log的复制是正确的，也就是paxos要解决的问题。

Paxos算法

The Paxos algorithm, when presented in plain English, is very simple.

共识问题

上述问题可以简化为，有多个服务可以propose value，而共识算法将保证有且仅有一个value会被选中（chosen）。这里面隐含的条件是：

• 如果没有任何value被propose，那么也不应该有任何值被选中
• 值一旦被选中，各个服务应该可以知晓（learn）选中的值

从而可以得出共识算法实现的安全性约束：

Safety requirements

Validity (or non-triviality)

Only a value that has been proposed may be chosen, A process never learns that a value has been chosen unless it actually has been.

Agreement (or consistency, or safety)

Only a single value is chosen, and

Termination (or liveness)

if a value has been chosen, then a process can eventually learn the value.

assumptions

另外，paxos中是不考虑拜占庭问题的，有如下假设成立：

• Agents operate at arbitrary speed, may fail by stopping, and may restart. Since all agents may fail after a value is chosen and then restart, a solution is impossible unless some information can be re- membered by an agent that has failed and restarted.
• Messages can take arbitrarily long to be delivered, can be duplicated, and can be lost, but they are not corrupted.

Choosing a Value

Paxos made simple中，从最简单（但不正确）的方式开始，逐步加强约束，最终得到正确的算法。