Refresh your understanding of redis cluster

Redis colony

Redis Cluster adoption   Fragmentation   To share data , It also provides replication and failover capabilities .

1. node

1.1 Node handshake

Redis Each node in the cluster uses CLUSTER MEET Command to connect .

Log in and copy

      
      

       
       CLUSTER MEET <ip> <port>
      
      
      
      

       
       
1.
      
      

To a node node send out CLUSTER MEET command , It can make node Node and ip and port At the specified node   handshake  . After a successful handshake ,node Node adds the target node to node The node in the cluster .

Repeat this operation , Multiple nodes can be in the same cluster .

1.2 Cluster data structure

• clusterNode structure Saves the current state of a node , Including node creation time 、 Node name 、 Node configuration era 、 node IP Address and port number and so on .
• clusterLink structure yes clusterNode A property of , Save the information needed to connect nodes , Like socket descriptors 、 Input buffer, output buffer, etc .
• clusterState structure Saved by each node , The status of the cluster from the perspective of the current node is recorded , For example, whether the cluster is online or offline , How many nodes does the cluster contain , The current configuration era of the cluster and so on .

1.3 CLUSTER MEET Command implementation

Node receiving command A Connect with the target node B Conduct   handshake  .

2. Slot assignment

2.1 Redis Slot in

Redis The cluster stores key value pairs in the database through sharding , The entire database of the cluster is divided into 16384 Slot , Every key in the database belongs here 16384 One of the slots , Each node in the cluster can be processed 0 Or at most 16384 Slot .

When in the database 16384 Each slot has a node processing , Cluster in the   Online status  . If there is no node processing in any slot , So the cluster is in   Offline status  .

By sending to the node CLUSTER ADDSLOTS command , We can assign one or more slots to the node . For example, the following command can change the slot 0~5000 Assign to node 7000 be responsible for ：

Log in and copy

      
      

       
       127.0.0.1:7000> CLUSTER ADDSLOTS 0 1 2 3 4 ... 5000
      
      
      
      

       
       
1.
      
      

2.2 Record the slot assignment information of the node

clusterNode Structural slots Properties and numslot Property records which slots the node is responsible for processing .

Log in and copy

      
      

       
       slotsnumslots
      
      
      
      

       
       
1.
      
      

2.3 Propagate the slot assignment information of the node

Nodes will send their own slots The array is sent by message to the other nodes in the cluster , Tell them which slots they are currently dealing with .

When node A Slave node via message B There the node is received B Of slots Array time , It will be on its own clusterState.nodes Look up the node in the dictionary B Corresponding clusterNode structure , And update .

3. Execute commands in the cluster

The online cluster can execute commands .

When the client sends a command to the node , The node receiving the command will check which slot the database key to be processed by the command belongs to , And check that the slot is assigned to yourself .

Log in and copy

      
      

       
       MOVED
      
      
      
      

       
       
1.
      
      

3.1 Calculate which slot the key belongs to

redis Use the following algorithm to calculate the given key Which slot does it belong to

Log in and copy

      
      

       
       def slot_number(key):return CRC16(key) & 16383
      
      
      
      

       
       
1.
      
      

CRC16(key) Statement calculates key Of CRC-16 The checksum , and & 16383 Calculate a value between 0 and 16383 An integer between as key Slot number of .

CLUSTER KEYSLOT The command is implemented according to the slot allocation algorithm above .

3.2 Determine whether the slot is in the charge of the current node

Check your own slots The corresponding position in the array .

3.3 MOVED error

MOVED The wrong format is ：

Log in and copy

      
      

       
       MOVED <slot> <ip>:<port>
      
      
      
      

       
       
1.
      
      

The client can be based on MOVED error , Turn to the right node .

3.4 The implementation of the node database

In addition to saving key value pairs in data , Nodes also use clusterState The structure of the slots_to_keys Jump table to save the relationship between slot and key .

The score of each node of the hop table is a slot number . The member of each node is a database key .

Use this jump table , It is convenient to batch operate all database keys belonging to some slots .

4. The shard

4.1 Re segmentation process

Redis The re fragmentation operation of the cluster can cause the node assigned to a slot to be reassigned to another node . Re sectioning can be done online , In the process, the cluster does not need to go offline , And can handle command requests normally .

Re slicing by Redis Cluster management software redis-trib Responsible for the execution of .

The re slicing process is shown in the figure above . If the slot to be sliced belongs to multiple nodes , You have to send commands to multiple nodes .

4.2 ASK error

A situation may occur during re segmentation ： The key value pairs belonging to the migrated slot are saved in the source node , Another part of key value pairs is saved in the target node .

At this time, when the client sends a command to the source node , And the database key required to be processed happens to be being migrated ：

• The source node now finds its own data , If you find it, send the command directly .
• If not found, return to the client ASK error , Direct the client to the target node to execute the command .

One ASK The error is shown in the figure below ：

received ASK The wrong client will provide according to the error IP Address and port number , Turn to the target node , Send a message to the target node first ASKING command , Then resend the original command .

ASKING Command can open the of the client sending the command REDIS_ASKING identification . This is useful when the server receives the following commands .

The server that is in the process of re fragmentation judges the status of the client sending the request ASKING Identify whether to open

• If opened , The client's request is a redirection request during slot dispatch , Normal execution .
• If it doesn't open , The client's request is an ordinary command request with addressing error , return MOVED error .

5. Replication and failover

Redis The primary node in the cluster is used for processing slots , The slave node is used to replicate a master node , And take over the copied master node when it goes offline .

Suppose the status between servers in a certain scenario is as follows ：

If the main server 7000 Offline , Then the remaining primary servers are 7001、7002、7003 From 7000 Two slave servers 7004、7005 Select a takeover 7000 Responsible tank . The other slave server will become the slave server of the new master server .

If the follow-up 7000 Back online , It will become 7004 The new master node of .

5.1 Fault detection

Each node in the cluster will regularly send messages to other nodes in the cluster PING news , To check whether the other party is offline . If you receive PING The node of did not return within the specified time PONG news , Will be marked as   Suspected offline  .

Nodes in the cluster have three states from the perspective of other nodes ：  On-line 、 Suspected offline 、 Offline .

When a master node A The master node is informed by the message B Think of the master node C When entering the suspected offline , Master node A It will be on its own clusterState.nodes Find the master node in the dictionary C Corresponding clusterNode structure , And the master node B Of offline reports added to clusterNode Structural fail_reports In the list .

If you're in a cluster , A master node finds that more than half of the master nodes will a master node x Mark as meaning offline , Then he will mark the master node as offline , And broadcast to all other nodes FAIL news .

5.2 Fail over

When a slave node finds that the master node it is copying is offline , The slave node will start to fail over the offline master node ：

1. The primary node that finds the problem first holds an election , Call yourself the new master node . Election rules and Sentinel Middle election leader Sentinel Very similar , They're all based on Raft The leading election method of algorithm .
2. Success is called the new master node
3. If it fails, wait for the next slave node to propose an election .
4. The selected slave node performs SLAVEOF no one command , Becomes the new master node .
5. The new master node takes over all slot assignments of the offline master node .
6. The new master node broadcasts PONG news , Let other nodes know that they have become the master node .
7. The new master node starts to perform the duties of the master node