Distributed consensus mechanism

Want to make APP Same thing as WeChat , Can run small programs smoothly ？ | Experience will send you to Xinjiang 、 Huawei 、 Cherry keyboard ！>>>

• C/S Pattern

1. Let's start with the simplest two computers , One is the server , One is the client , namely C/S Pattern . The client wants to manipulate the data on the server node .
2. Since the client wants to manipulate the data on the server , Network communication is necessary , dispatch orders . But we can't guarantee that every message sent by the client will be received by the server .
3. So when the server receives the message , A confirmation message will be sent to the client , If the client does not receive the confirmation message sent by the server , The client will resend the same message command .
4. But the time when the client sends a message command to the server may be different . The time when the server sends the confirmation message to the client is also different .
5. This will result in different order of message commands received between multiple servers and multiple clients . Between different servers , Their states are different .
6. For a group of servers , If all servers execute a sequence of commands in the same order ( This set of commands may be sent by multiple clients ), This group of servers implements state replication . Because it's easier for a single server to do this , We usually only send message commands to a single server , A single server must also execute a sequence of commands in a sequence .
7. When a single server receives a command , It can forward this command to other servers . When it receives confirmation from other servers , It will inform the client that the command has been successfully executed .
8. A single server is prone to downtime , If we don't use a single server to forward commands , It's a two-phase commit , A client tries to get all the server locks first , If you get all the server locks , Start sending commands . In this way, all servers will not be able to receive commands from other clients , Ensure the order consistency of all servers , After the client sends the command, it releases all the server locks . If this client can't get all the server's locks , Release the lock of the server you get , Wait for a while , Try to get all the server locks again . Of course, to save time , We believe that the client can start to execute the command after getting half of the server's lock , Because one client gets half of the server's locks , Then other clients can't get most of the server's locks , In this way, clients who only get a small number of locks have to release the locks .

• Paxos

When we adopt a two-stage proposal , Instead of a single server forwarding , When multiple clients attempt to acquire the lock of most servers at the same time , What's going to happen ？ Whether clients have to release all the locks they acquire , To avoid deadlock . Or the client gets part of the lock and hangs up ？

How about if we let this lock expire ？ We call these expired locks tickets . In this way, the server that has not been released will not be locked all the time . First, the client requests a ticket from all servers , Tickets are sent from the server to the client , When the client receives half of the tickets sent from the server , The ticket will be sent to all servers along with the command . If the server finds that the ticket received is the latest one sent out by itself , And it's not overdue , It will give the client a positive feedback . The client received positive feedback from more than half of the server , Will notify All the servers Start executing orders . But the server hasn't been informed by the client to execute the command , It will also continue to accept ticket requests from other clients , And will send an up-to-date ticket to other clients . When other clients get half the votes , New tickets and your own commands will be sent to all servers . At this time, if some servers receive this command and ticket , The previous client that received more than half of the server's positive feedback informed all servers to start executing the command , At this time, the commands of all servers are inconsistent .