redis-6. Redis master-slave replication, cap, Paxos, cluster sharding cluster 01

redis

redis stand-alone

stay redis stand-alone , A single node , Problems with single instances

• A single point of failure （ The physical machine hung up ）
• Limited capacity
• Under great pressure

Solution

AKF

Yes XYZ Three axes , Only one can happen , Multiple dimensions can occur together

AKF Cubes are also called scala cube, It's in 《The Art of Scalability》 First proposed in a Book , It aims to provide a systematic expansion idea .AKF The system expansion is divided into the following three dimensions ：

X Axis ： Directly and horizontally copy the application process to expand the system .
Y Axis ： Split the functions to expand the system .
Z Axis ： Expand the system based on user information .
No matter what redis,mysql still kafka Can be extended based on this concept

adopt akf To solve the problem will extend to other problems
problem 1 Data consistency issues
programme 1： All nodes are blocked , Until all numbers are consistent , But this situation will lead to redis Reduced availability of , More than one side is to solve the availability problem

programme 2： Tolerate the possibility of data loss , Return directly to the host redis ok, Insert standby asynchronously redis, This situation must tolerate data loss

programme 3： How to guarantee not to lose , The middle price can be inserted kafka, Achieve final consistency of data . There are many possibilities ： Before reaching final consistency , Inconsistent data may be obtained

problem 2 If the Lord redis Hang up , Make the whole service unavailable
The main equipment ： The client does not participate in the business , Only as a standby machine
Master-slave ： Master slave copy , You can retrieve data from the library （ Emphasize the concept ）
Generally, it is highly available to the primary user HA, If the Lord redis Hang up , We can do it manually To prepare 1redis Top up , As a host , Let prepare 2redis Continue to be an assistant ,

But ah , brothers , We are fucking programmers , Every time this kind of fault occurs, it is not troublesome , It will be very low And there is still time to recover , This is when the monitoring program appears , But a monitoring program will also hang up , Therefore, the monitor itself should also be a cluster that has become more and more .

Multiple monitors ： Multiple monitors are waiting for a master , If redis Hang up , Monitor to find him , Because there are network delays , How to judge by multiple monitors redis Did you hang up or not ？ Like monitor 1 Sending hung up , monitor 2 Sending hung up , monitor 3 No, I didn't listen , How to deal with this situation ？ Strong consistency cannot be used at this time , Only some machines need to give the result judgment , that ！ How many are there in one part ？

The three monitors produce different results , Create competition . There will be a problem ： Network partition （ Split brain ）, Different clients get different data , External performance is data inconsistency .
If 2 A monitor said redis Hang up ,1 A monitor said redis Still alive , this 2 The forces monitored are 2. The force is 1 There is no decision-making power for monitoring . At this time, the external performance is not an ambiguous state .
therefore , When the faction reaches n/2+1, That is, half way through , You can solve your mental problems . Therefore, the cluster generally uses an odd number of servers .
Why is it an odd number ？
We use it 3 Taiwan and 4 Taiwan comparison , They are only allowed to hang one . but 4 Table ratio 3 It's easier to hang one , and 4 The price is higher . So the first 4 There is no need for a platform .

CAP

A term is used to explain ：
Master-slave : The client can access the master You can also access the backup
The main equipment : The client will only access the primary The standby machine is used to replace the main hook
CAP: Uniformity Usability Partition tolerance Two at most Not all of them

Master slave copy

Redis Use default asynchronous replication , It is characterized by low latency and high performance , It's the vast majority Redis The natural replication pattern of use cases . however , from Redis The server asynchronously confirms its slave master Redis The amount of data received by the server in a cycle .
Conditions are limited , I started three on one host redis

5.0  Previous master-slave replication operations 
127.0.0.1:16379> help SLAVEOF

  SLAVEOF host port
  summary: Make the server a replica of another instance, or promote it as master. Deprecated starting with Redis 5. Use REPLICAOF instead.
  since: 1.0.0
  group: server
5.0  Subsequent master-slave replication operations 
127.0.0.1:16379> help REPLICAOF

  REPLICAOF host port
  summary: Make the server a replica of another instance, or promote it as master.
  since: 5.0.0
  group: server

I'm here at this time redis16380 Set the main database above

REPLICAOF 127.0.0.1 16379

Report errors ！ Thought I was 16379 Added above password . Therefore, it is impossible to link

To configure... Here 16380.conf On the configuration 6379 Password


Repeat the previous link operation

16379.log

16380.log

127.0.0.1:16379> set k1 zjj
OK

By default 16380 Is read-only

127.0.0.1:16380> get k1
"zjj"
127.0.0.1:16380> set k1 aaa
(error) READONLY You can't write against a read only replica.

I'm going to modify 16381 follow 16379, At this time, verify all the above problems , In the case of following ,redis Followers will flushall My own library

127.0.0.1:16381> set k2 aaa
OK
127.0.0.1:16381> get k2
"aaa"
127.0.0.1:16381> 
127.0.0.1:16381> REPLICAOF 127.0.0.1 16379
OK
127.0.0.1:16381> 
127.0.0.1:16381> get k1
"zjj"
127.0.0.1:16381> get k2
(nil)

At this time I shutdown 16381

127.0.0.1:16381> SHUTDOWN

16379 newspaper 16381 The link doesn't go up , At this time, we will redis 16381 Restart up , that 16379 How to synchronize your data

stay redis 16381 Continue writing in case of hang up 16379

127.0.0.1:16379> set k2 aaa
OK
127.0.0.1:16379> set k3 bbb
OK

Restart 16381 redis-server ./6381.conf --replicaof 127.0.0.1 6379
After reboot , It will still synchronize the increment during the period of suspension .

127.0.0.1:16381> keys *
1) "k1"
2) "k3"
3) "k2"

• masterauth < master-password >

Set the password

• replica-serve-stale-data yes

When in the process of synchronization , Or slave and master The connection between them has been broken ,replica Can I still provide read requests ？
By default ,replica You can't provide a request , Go straight back to error, But when the following configuration is turned on , The old data can be used to respond to the request .

• replica-read-only yes

By default ,slave Is read-only , Prevent misoperation from writing bad data , But if you have to be writable , It can also be opened through the following configuration .

• repl-diskless-sync no

about master End ：
The following configuration can be enabled , After opening master Will no longer be generated rdb file , Directly encode the memory data and send it to replica

• repl-backlog-size 1mb

Incremental replication
Is the buffer configured as large as possible ？
If the configuration is too large , such as 10G, The actual data in memory may only be 100M, When replica and master The connection between them has been broken long enough , then replica It's reconnected master, Sent a offset To come over , And then it started 10G The transmission of commands ,
In fact, if it is full synchronization at this time , You may only need 100M The transmission of the data of . According to this situation , The buffer should not be too large .

• min-replicas-to-write 3
• min-replicas-max-lag 10

If replica and master The connection between is broken ,master How does it work ？ It's indifference ？
By default ,replica and master After disconnection ,master I don't do anything , It's just a replica, after all ,master Don't care . But it can be configured through the above , Give Way master care about .
This configuration is represented in 10 At least... In seconds 3 individual replica On-line , and master The connection is normal , otherwise master Write operations from the client are not accepted ,master During this time it became read-only , This will allow replica As much as possible and master The gap between the two countries has narrowed . But it also affects master Normal request for .
Detailed Links ：https://blog.51cto.com/happytree007/2586641

Sentinel sentry

Redis Of Sentinel The system is used to manage multiple Redis The server （instance）, The system performs the following three tasks ：

• monitor （Monitoring）： Sentinel Constantly check whether your master and slave servers are working properly .
• remind （Notification）： When someone is being monitored Redis When there's a problem with the server , Sentinel Can pass API Send notifications to administrators or other applications .
• Automatic failover （Automatic failover）： When a primary server doesn't work , Sentinel An automatic failover operation will start , It will upgrade one of the failed master servers from the server to the new master server , And let the other slave servers of the failed master server copy the new master server ; When a client tries to connect to a failed primary server , The cluster will also return the address of the new primary server to the client , So that the cluster can use the new master server instead of the failed server .

at present Sentinel System is Redis Of unstable Part of a branch , You must go to Redis Project Github page Clone a copy of unstable The score is , Then compile to get Sentinel System .
Sentinel The program can be compiled in src Found in the document , It's a name for redis-sentinel The program .
You can also use the methods described in the next section , Give Way redis-server The program runs in Sentinel In mode .
in addition , A new version of Sentinel It's already included in Redis 2.8.0 Release file of version .

redis The installation directory src It can be found below

Take a look at our system default program directory

[[email protected] redis]  redis-sentinel 26379.conf 

port 26379
##  At present Sentinel Node monitoring  127.0.0.1:6379  This master node  , 1  Represents the minimum required to determine that the master node failed 2 individual Sentinel Node node agreement 
sentinel monitor mymaster 127.0.0.1 16379 2

# sentinel author-pass Define the password for the service ,mymaster It's the service name ,123456 yes Redis Server password 
# sentinel auth-pass <master-name> <password>
sentinel auth-pass mymaster 123456

##  Every Sentinel Nodes are scheduled PING Command to judge Redis Data nodes and the rest Sentinel Whether the node is reachable , If exceeded 10000 MS and no reply , It is unreachable 
sentinel down-after-milliseconds mymaster 10000

You can refer to ：https://www.jianshu.com/p/06ab9daf921d

redis-sentinel 26379.conf 
redis-sentinel 26380.conf 
redis-sentinel 26381.conf 

take salve The library is promoted to the main library

slaveof no one

 127.0.0.1:16379> SHUTDOWN

here 26379,26380 Listen for service changes （ Be careful , Wait for the network or heartbeat 1 About seconds ）

This moment reminds me