What is cerebral fissure?

The so-called brain crack problem is the high availability of multi machine hot standby HA In the system , When the heartbeat of two nodes suddenly breaks , The dispute is classified as two independent individuals , Due to the loss of contact with each other , They all think that the other party has failed , Therefore, they will compete for each other's resources , This is the brain crack problem

Take a chestnut ：

A and B Two nodes clustered as a dual machine , Each holds a subset of the data of the cluster   a and b , At this time , Communication between the two machines suddenly failed ,A Think B Hang up ,B Think A Hang up and appear ：

（1） if A Have b Data backup , be A Complete operation data ,B Also complete operation data , This is likely to cause two nodes to read and write shared data at the same time, causing data damage

（2）A、B Each has only a、b data , Then either the two nodes cannot be started , Or start by dividing up the complete shared data

The cause of the brain crack problem ：

1. Network problems -> Abnormal network problems cause physical separation of clusters , Cause brain crack

2. Node load -> if master Node load is too high , May cause master The node stops responding , Thus leaving the cluster , The cluster reselects the master , After resuming the response, there is a brain crack problem

Solutions ：

1. Try to deploy the cluster in the same Intranet environment , So as to ensure the reliability of communication between nodes

2.master Node and data Node separation , Guarantee master Node responsiveness

（ adopt node.master ：true    And  node.data：false  To decide whether to become master Qualification of nodes ）

 Redis Split brain

Explain the role of sentinels ：

1. Send information every other event to check whether the main server is still alive , If an uncontrollable cause causes the main server to die , It will affect the whole cluster
2. So the sentinel will choose one of the slave servers to be the master server Based on the above environment , At this time, the network environment fluctuated, resulting in the death of the main server , At this time, the sentinel will start the election
3. After the election, a slave server becomes the master server , And just at this time, the original master server , It's restarted , But at this time, it can only become a slave server

  summary ：

After the original master server receives the information from the client , You are connected before synchronizing to the slave server , But after restart, it changes from master server to slave server , Unable to synchronize data , So this part of the data is lost , This is the brain crack problem

resolvent  

Add the following configuration to the configuration file

min-slaves-to-write 1
min-slaves-max-lag 10

The first parameter indicates the connection to master At least slave Number （ Determine how many slave servers , Sending information only when the requirements are met avoids writing data after the main server is disconnected ）

The second parameter represents slave Connect to master Maximum delay time （ Reduce the synchronization interval , The synchronization is completed before the loss of connection ）

ES Cleft brain problem

Active election mechanism ：

1. elasticsearch Once the cluster is established , One will be elected master, Everything else is slave node . But the specific operation , Each node provides write and read operations . That is to say , No matter which node you write to , This data will also be allocated to all nodes in the cluster
2. There is a situation where a node hangs , If it is slave The node hangs up , Well, first of all , Will the data be lost ？ Can't . If you turn it on replicate, Then this data must be backed up on other machines . The backup partition on other nodes will be automatically upgraded to the primary partition of this partition data . It should be noted here that there will be a short period of time yellow State time

When the slave nodes find that they cannot connect with the master node , Then they will decide to elect another node as the master node

• Yes 5 Taiwan machine ,3 The desk is in a machine room ,2 It's in another machine room , When the connection between the two computer rooms is broken , The nodes of each computer room will meet by themselves , Select a master node . At this time, there are two master nodes , When the connection between the computer rooms is restored , At this time, there will be data conflicts

  The solution is to set parameters ：

discovery.zen.minimum_master_nodes