redisThere are three clustering methods: master-slave replication, sentinel mode, and clustering.
1. Master-slave copy
Master-slave replication refers to letting one server copy the data of another server so that the data of both parties is saved consistently. The copied server is the master server and the copied server is the slave server. When the data of the master server changes, the master server will notify the slave server to save the consistency of the data.
In Redis, use the slaveof command to perform master-slave replication.
(I) Master-slave copy implementation of old version
In the old version, Redis's replication function was divided into two steps: synchronization and command propagation.
Synchronous operationUsed to update the data state of the slave server to the data state of the master server.
Command propagation operationIt is used to notify the slave service to modify the corresponding data when the data of the master server is modified, so as to achieve data consistency.
After sending the slave of the server to the master server:
1.Slave server to master serversendsyncOrder, the main server willimplementbgsaveOrder, in the backgroundGenerate aRDBFile and use oneBuffer recordAll writing commands from now on.
2. Receive and load RDB files from the server,Update your own data. After the update is completed,The main server will send commands in the buffer, execute these commands on the slave server, and the data of the master and slave server will be saved consistently.
3.After the master and slave server save consistently, when the master server's data is modified, the slave server will be executed.Command propagation operation, that is, the modification command is sent to the slave server so that the data will continue to be saved consistently.
defect:
The implementation of the master-slave replication in the old version has a major flaw:After the slave server is disconnected and reconnected, the synchronization operation will be performed again between the master and slave servers, that is, all data from the master server will be retransmitted to the slave server.. But in fact, after reconnection, there may be only a small amount of data between the master and slave servers, and most of the storage is consistent. This makes a lot of data invalid and extremely inefficient.
(II) The new version of master-slave copy implementation
In order to solve the problem of master-slave replication efficiency, Redis started with 2.8 and usedpsyncOrderReplace sync.
There are two modes of the psync command: full resynchronization and partial resynchronization.
Complete resynchronizationUsed for initial replication, consistent with synchronization operations.
Partial resynchronizationUsed for replication after disconnection and reconnection.
Implementation of partial resynchronization:
Partial resynchronization,When the slave server is disconnected and reconnected, if the conditions allow, the master server can transfer theWrite commands executed during server disconnectionSend it to the slave server. As long as the slave server receives and executes commands, the data can be updated, and the master server can achieve data consistency.
Partial resynchronization has been usedCopy offsetandCopy the squeezing bufferTo achieve it.
Copy offset:Both the master and slave server each maintain a replication offset. The master server sends it to the slave server every timeNWhen one byte of data is added to your copy offsetN, every time the server accepts the master serverNWhen one byte of data is added to your copy offsetN. By comparing the replication offsets of the master and slave server, it is easy to know whether the master and slave server data is consistent.
After the server is disconnected,Difference in replication offset between master and slave servers, the master server can send the corresponding data to the slave server, that isData modified from server after disconnection。
The data sent by the master server to the slave server existsCopy backlog buffermiddle.
Copy backlog bufferIt's oneFixed lengthFirst-in-first-out queue, Due to its fixed length, when the queue is full, when the data is enqueued, the data at the front of the queue must be dequeued first, that is, copy the backlog buffer valueKeep fixed sizedata.
When the master server performs a command propagation operation, it not only sends the write command to the slave server, but also writes it to the replication backlog buffer, and records the copy offset of each byte of data in the buffer.. After reconnecting from the server, if the copy offset from the serverGreater thanThe copy offset of the buffer squadron head will bePartial resynchronizationto copy in a way; if the copy offset from the serverLess thanThe copy offset of the buffer squadron head will beComplete resynchronizationto copy.
two,RedisSentinel mode (Sentinel）
(I) Sentinel Mode
Sentinel mode is RedisHigh availabilitySolution:Through one or moresentinelcompositionsentinelsystem,Monitor any multiple master serversand theAll from the serverand in monitoredThe main server is offlineAfter that, automatically add one under the main serverThe sub-server is upgraded to a new primary server, and then the command is executed by the new master server.
The Sentinel system has three main functions:
Monitoring）： SentinelThe master and slave servers will be checked at a certain frequency whether they are operating normally.
Notification）: When a problem occurs with a monitored Redis server, Sentinel can send notifications to administrators or other applications through the API.
Automatic failure migration）: When a master server does not work properly, Sentinel will start an automatic failure migration operation. It will upgrade one of the slave servers of the failed master server to a new master server, and change the other slave servers of the failed master server to replicate the new master server; when the client tries to connect to the failed master server, the cluster will also return the address of the new master server to the client, so that the cluster can use the new master server instead of the failed server.
(II) Monitoring method
Every 10 seconds, each sentry node will send info commands to the master and slave nodes to obtainLatest topology。
Every 2s, each sentry node passesSubscribe to ConnectTowardsServer ChannelSend the judgment of the sentry node for the main node and the information of the current sentry node.,at the same timeEachSentinelThe nodes will alsosubscriptionThe channel, andReceived information。Notice:Sentinel is connectingMaster or slave serverWhen theCommand Connection and Subscription Connection, but connect to otherSentinelWhen nodes are created onlyCommand connection; This is because Sentinel connects through the server's subscriptionDiscover newSentieneland update othersSentinelThe status of, and for known Sentinels, communication can be communicated directly using command connections.
Every 1s, Each sentry node will send a ping command to the master node, slave node, and other sentry nodes to perform a heartbeat detection (heartbeat detection mechanism) to confirm whether these nodes are currently reachable.
(III) Determine whether the main server is offline
There are two steps to determine whether the main server is offline: subjectively offline and objectively offline.
Subjectively offline：singleSentinelI believe that the server has been offline.
Objectively offline：MultipleSentinelI believe that the server has been offline.
Subjectively offline:
When an instance (including master-slave nodes, sentinel nodes)A certain timeforSentinelThe reply isInvalid reply(All replies except for the three replies +PONG, -LOADING, and -MASTERDOWN, and no replies), the Sentinel will mark the instance as subjectively offline.
Note: Since the timeout time set by each Sentinel is different, some Sentinels may believe that the server has been subjectively offline, but some Sentinels believe that the server has not been subjectively offline, that is, it is still in a state of survival.
Objectively offline (Only the main server will make objective offline judgments）：
When Sentinel transfers a master serverAfter being judged as subjective,, willOthers that also monitor the main serverSentinelMake an inquiry, see if they think the main server has entered the offline state (can be subjective or objective). When SentinelreceiveA certain quantityofA sure replyAfter that, the server will be markedIdentify asObjectively offlineand the main serverPerform a failover operation。
Note: The number of positive replies set by each Sentinel is different, so some Sentinels may believe that the server has been objectively offline, while some Sentinels may believe that the server has not been objectively offline, and some Sentinels may even believe that the server is still alive.But only one Sentinel It is believed that a master server has entered an objective offline state. An automatic failure migration operation will be performed on the failed master server.
(IV) Election leaderSentinel
When a master server is judged to be subjectively offline, all Sentinels monitoring the master server will negotiate, elect a leader Sentinel, and he will perform a failure migration operation on the offline master server.
The principle of election isFirst come first served: Every Sentinel that determines that the main server is an objective offline will want other Sentinels to send requests and require them to become the leader Sentinel. As long as a Sentinel receives an election request sent by other Sentinels, it will elect it as the leader Sentinel and reject all subsequent election requests.
(V) Failover
After electing the leader Sentinel, the leader Sentinel will perform a failover operation on the offline primary server, which consists of three steps:
1) From the slave server that has offline the master server, select a slave server and convert it into the master server.
2) Change all slave servers under the offline master server to replicate the new master server.
3) Set the offline master server as the slave of the new master server. When it goes online again, it will become the slave of the new master server.
three,RedisCluster
Redis cluster passesThe way to shardTo save data, all data in the cluster isDivided into16384Slots (slot）, Each data in the database belongs to one of these 16384 slots, and each node in the cluster can handle 0-16384 slots. Which slot each data belongs to is derived through the hash algorithm, andOnly when16384Only when there are nodes in each slot are processed will the cluster be in the online state; as long as there is a slot without nodes in the process, the cluster will be in the offline state.
(I) Slot assignment
EachNodeA Slot segment is evenly allocated, corresponding to 0-16384. Slot cannot be repeated or missing, otherwise the object will be stored repeatedly or cannot be stored. And all nodes will notify other nodes of the Slot segment they are responsible for.
(II) Command execution
When all 16384 slots in the database are assigned, the cluster is online. When the client sends instructions to the node, the node receiving the command will determine whether the data operated in the instruction is in the slot where it is responsible. If it does not belong, the node will return a MOVED error, directing the client to turn the command to the correct node. After receiving the MOVED error, the client will send the command again to the correct node.
(III) Resort
Once a node is offline or joined, data migration will be carried out according to Slot. For example, if Node1 is disconnected, the 0-5640 Slots will be evenly distributed to Node2 and Node3. Since the Slots maintained by Node2 and Node3 will not be redistributed on themselves, the migration process will not affect the use of the 5641-16384Slot segment.
The judgment that a node is offline is that more than half of the nodes in the cluster believe that the node has been offline.
(IV) Master and Slave Node
In a cluster, each node can be a master-slave node (composed of one master server and multiple slave servers), which ensures high availability of the cluster.
If a node in the cluster is a master-slave node, when the master node fails, the Slot segments copied by the master node will not be resliced, but instead a node will be elected from the slave node as the new master node. The election rules are: the slave node asks other master nodes in the cluster. As long as more than half of the cluster nodes agree, the slave node is elected as the master node and continues to execute the command.
Note: In cluster mode, nodes can only use db0 databases, while in stand-alone situations, we can use 16 databases (db0 ~ db15).
Four,Three typesRedisPros and cons of structure
(I) Master-slave copy
advantage:
1) Reading and writing are separated, copied and written by the master node, and the slave node is responsible for reading, which is suitable for high concurrency.
2) The ability of concurrent reading can be improved by increasing the slave nodes.
shortcoming:
1) The master node has limited writing ability, and the writing concurrency of the master-slave replication structure depends on the writing of the master node.performance。
2) If the automatic fault tolerance and recovery function is not available, the downtime of the host and slave will cause some read and write requests to fail.
(II) Sentinel Mode
advantage:
1) Sentinel mode is based on the master-slave mode. All the advantages of the master-slave mode are available.
2) The master and slave can automatically switch, the system is more robust and usability is higher.
shortcoming:
1) There are also cases of insufficient write performance.
(III) Cluster
advantage:
1) Like Sentinel mode, it has the advantages of high availability and separation of read and write.
2) Each node stores different data, and each node can read and write. Solved the problem of low read performance in other structures.
shortcoming:
1) Each node is composed of a server cluster (one master and multiple slave nodes), which may cause waste.