Build a highly available database

1、 A master-slave mode

1.1、 Why set master-slave synchronization

• Read / write separation , Reduce the burden of database . In the development work , Sometimes I meet someone sql Statement needs lock table , Resulting in temporary inability to use the read service , This will affect the existing business , Use master-slave replication , Let the main database be responsible for writing , Read from the library , such , Even if the main database appears to lock the table , The normal operation of the business can also be ensured by reading from the database
• Database real-time backup . When a node in the system fails , It's easy to fail over ( Master slave switch ).
• High availability
• With the increase of business access in the system , If it is a stand-alone deployment database , It will lead to I/O Access frequency is too high . With master-slave replication , Add multiple data storage nodes , Distribute the load across multiple slave nodes , Reduce the cost of stand-alone disk I/O Frequency of visits , Improve the... Of a single machine I/O performance .

1.2、 Master-slave form

• One master, many followers — Read / write separation , Provide the concurrency capability of the cluster ( In general , The scenario of reading data is redundant to the scenario of writing business )
• Multi master and one slave — The slave database is mainly used for database backup
• Double master replication
• Cascade replication

1.3、 Realization principle

Follow the two-phase commit protocol , The main process is as follows ：

• Update of main database SQL(update、insert、delete) Written to binlog
• Initiate connection from library , Connect to the main library .
• At this time, the main library creates a binlog dump thread , hold bin log Content sent from library .
• After starting from the library , Create a I/O Threads , Read from the main library bin log Content and write to relay log
• From the library will also create a SQL Threads , from relay log Read the contents inside , from ExecMasterLog_Pos The location starts executing the read update event , Write the update to slave Of db

1.4、 Data high availability solution

• Dual active standby ： Two machines A and B,A Give priority to the library , Read and write ,B For backup , Backup data only . If A Library failure ,B The library becomes the main library and is responsible for reading and writing . After fixing the fault ,A Become a standby warehouse , Main library B Synchronize data to the standby database A
• A master from ： Two machines A and B,A Give priority to the library , Read and write ,B For from the library , Read the data . If A Library failure ,B The library becomes the main library and is responsible for reading and writing . After fixing the fault ,A Become a slave , Main library B Synchronize data to slave Library A.
• One master, many followers ： One master library and multiple slave libraries ,A Give priority to the library , Read and write ,B、C、D For from the library , Read the data . If A Library failure ,B The library becomes the main library and is responsible for reading and writing ,C、D Responsible for reading . After fixing the fault ,A Also become from the library , Main library B Synchronize data to slave Library A.
• MariaDB Synchronize multiple hosts ： There is an agent layer to realize load balancing , Multiple databases can be read and written at the same time ; Various databases can be connected through Galera Replication Method for data synchronization , Theoretically, the data of each library is completely consistent .
• database middleware ：mycat Fragmentation storage , Each partition is configured with a master-slave cluster .

1.5、 Comparison between one master multi-slave mode and cascade replication mode

​ In one master multi slave mode , The transmission delay between master and slave will be smaller , But there will be too many binlog dump Threads , It will aggravate the main database IO load .

​ If cascading replication is used , There is only one master dump Threads , Do not read from the master binlog The content of , Only responsible for the master's binlog Transfer to slave Library . Set the slave tables and all kinds to BLACKHOLE engine , The data generated from the main application log is thrown into the black hole , You only need to apply the master database from the master database binlog When the binlog, So set log_slave_updates=on, Open from the master binlog, Only in this way can the logs generated from the master application be recorded binlog Transferred to the （ second level ） From the library .

2、 Sub database and sub table

As the business grows ,MySQL More and more data will be saved in . here , Databases can easily become a bottleneck in system performance , Stand alone storage capacity 、IO、CPU Processing power is limited , When the amount of data in a single table reaches 1000W or 100G in the future , The operation of adding, deleting, changing and querying database tables is faced with the problem of significant performance degradation . This is to consider the method of database and table optimization .

2.1 vertical partitioning

2.1.1 Vertical sub database

Split the system into multiple businesses , Each business uses its own separate database .

2.1.2 Vertical sub table

​ Vertical table splitting is based on the table fields in the database . There may be some tables with many fields in the business , Some fields in the table are long , We only need these long fields occasionally , At this point, we can consider splitting the table vertically .

​ Some of the less commonly used , But the fields with large length are taken out and put into another table .

​ MySQL The bottom layer is stored through data pages , A record takes up too much space, which leads to page spread , Cause additional performance overhead . In addition, the database loads data into memory in behavioral units , In this way, the length of the fields in the table is shorter and the access frequency is higher , Memory can load more data , More hits , Reduced disk IO, Thus, the database performance is improved .

2.2 horizontal partitioning

2.2.1 Sub table in the warehouse

The sub tables in the library are in the same DB On , Split a table into multiple tables according to certain conditions . Tables can be divided according to the stored date （ The data stored in January is 1 A watch , The data stored in February is a table ……）, It can also be stored according to id Sequence numbers are divided into tables （1-1000000 by 1 A watch ,1000001-2000000 by 1 A watch ……）. The specific expression shall be subject to the business , On demand sub table . The chart below shows the daily schedule .

2.2.2 Sub database and sub table

Splitting databases and tables means splitting tables , And split into different machines .

For example, Tencent cloud DCDB This is the way to deal with it . You can specify the of a table shardKey, Then on shardKey take hash, according to hash Values put data in different databases , It can solve the bottleneck problem of single machine physical resources .

There are two kinds of Middleware in the market , That is, Dangdang sharding-jdbc Ali mycat, The differences between the two are as follows ：

3、 Distributed cache

3.1 Optimization principle

After receiving the query request , Let's query the cache first , Determine whether there is data in the cache , Data is directly returned to the application , If you do not query the database again , And load it into the cache , This greatly reduces the number of visits to the database , Naturally, it also improves database performance .

Be careful ： After the introduction of distributed cache, the system needs to consider how to deal with cache penetration 、 Cache breakdown and cache avalanche .

3.2 High performance

Suppose the scene , A request to come , Save the database query results and return them to the user , Assumed time consuming 600ms. But that may not change in the next few hours , Or it can be changed without immediate feedback to users . So what to do at this time ？

Then consider using caching , One key Corresponding to one value, Next time someone checks , No more spending 600ms from mysql Query in , Directly from the cache , Through one key Find out one value,2ms Get it done . Performance improvement 300 times .

That is to say, for some results that need complex operation and time-consuming , And make sure it doesn't change much later , But there are a lot of read requests , Then directly put the results of the query in the cache , Just read the cache directly later .

3.3 High concurrency

Suppose the scene , Traffic access peak one second request 1 Ten thousand times , about mysql Come on , The load is quite large ,mysql It is likely to be down or the efficiency of the query will be greatly reduced . At this point, if you cache , Cache a lot of data , No more searching mysql, Read directly from the cache , Can greatly increase the amount of concurrency .

4 an account of happenings after the event being told

Subsequent updates mycat Sub database and sub table , Distributed cache redis 's blog post