High burst solution 2

For a variety of reasons , We have encountered some scenarios that require high burst support .

Scenario business requirements ：
1、 High burst
2、 There are a lot of read and write operations , And need immediate response
3、 Stable performance

First step ： Enhance service capabilities - SLB Load Balancer

user C The client operates the data through the back-end program , If the user traffic reaches a certain level at a certain time , So it is difficult for a single server to support such a large amount of traffic . Cause server downtime . There are many reasons for server downtime , Let's make a concrete analysis of
1、 Because the flow is too large , The server bandwidth is full , This causes users to download slowly , Poor experience
Traffic is mutual , When users upload content , The server is downloading , When the server responds to the data to the user, the user downloads it , When the user group is large enough , The traffic is full , As a result, everyone's download speed is limited , This will also lead to a series of problems
Connection timeout ： Because the data is not completely transmitted to the user within a certain time .
Ask to wait ： Bandwidth is full , You need to wait for the server to finish downloading .
CPU Full performance ： The user is requesting , The server is responding , However, the throughput of the server becomes smaller due to the traffic , Lead to a lot of tcp Request to enter the waiting state , The server needs to maintain this wait state （http 3 The second handshake ）, Maintenance status requires consumption cpu
404 Respond to ： The website has been down , reason ： Too much flow -> Connection timeout -> Ask to wait -> cpu Full performance -> 404 Respond to
Increase server bandwidth traffic

2、 The computing power of the machine is not enough
Business logic is too complex 、 A single interface consumes a lot of resources （ Such as uploading pictures 、 Audio etc. ）. Regardless of bandwidth , The computer that processes pictures and audio also needs to be maintained in the request （ You need to accept from the user in memory now C Upload of client , And save it locally , Wait for the request to end before releasing the resource ）
Add server configuration

3、 A single server has been launched , Still unable to support the access of user groups
Then, multiple servers serve users C End group services , Such as load balancing . Specific implementation principles to flip through my other chapters

4、SLB
Real time monitoring of machine performance nodes , When the machine performance reaches a certain threshold , Will be considered to be running out of performance , here SLB A new server instance will be automatically created （ The new server ） And it is attached to its own monitoring system to provide services for users .

The second step ： Increase database capacity - Read / write separation 、 Data fragmentation 、 hot standby 、 Connection pool

Generally, the place that causes the website to crash is the database , The importance of databases goes without saying , First, we can use third-party cloud products （RSD or poloarDB ） Generally, dual machine hot standby is used and implemented by default , We don't need to say more about this .
1、 Connection pool
Both traditional and database access are used to connect to the database , Release immediately after use （php）. This will bring the overhead of establishing a connection when interacting with the database . There is nothing wrong with low-frequency access , But if it is a high burst , This is not friendly , It is better to build a database connection pool to reduce the cost of establishing connections between programs and databases .（ That is, the connection will not be released after the first use , Instead, join the connection pool , The next time you reuse the database , You don't have to re-establish the connection , Directly from the connection pool ）. Such as ali cloud RDS or polarDB It implements this connection pool for us , If you are php, It can be accessed and used as usual , It is a database middleware itself . That is, we have changed from database access to database connection pool middleware

2、 Read / write separation
Most of the high burst scenes are read a lot , If so, there are two options
programme 1：redis colony or Fragmentation
If your business is a lot of access to a single piece of data , Then I suggest you query the data , Distribute this data to multiple redis（redis It's best to work on different physical machines , avoid CPU Resource occupancy conflict ）, And then randomly select from these multiple... When querying redis Get data in , The purpose is to reduce the cost of a single machine redis The pressure of the , But usually one redis That's enough. .

If your business has extensive access to multiple data , And the frequency of each data access is similar , Then make slices , Put the data A among redisA、 data B among redisB、 During the investigation , Go to whatever data you want redis In order to get .

programme 2：mysql colony Or slice
If your business is a lot of access to a single piece of data , Well, first of all, consider mysql The number of connections （mysql Configuration of ） Whether it supports so many traffic accesses . Can I create a connection pool to buffer the number of connections . Next, consider how my table index optimization works , When getting to the user , How much performance is consumed by a single request , How about the database consumption under the test burst .
1、 colony ： Adopt read-write separation mode , The read library obtains from the main library binlog Binary , So as to achieve the master-slave data synchronization function , And users C A large number of requests on the end will get data from the reader , If we create multiple read Libraries , Then there are multiple databases to provide services for users , Thus, the pressure of a single database server is reduced

2、 Fragmentation ： Categorize data , Put them in different databases , Common problems are that the business is split and mounted on different servers, while the database is also migrated to different servers . At this point, accessing data is to obtain different data from different databases

What if your scenario requires your service to respond in a timely manner ？
Common activities such as second kill , Ask you to give a feedback immediately .（ Processing per second 20W request , Regardless of the server , Only database stability is considered ）
Queues cannot be used MQ（ The queue also needs to be polled slowly in the background , In case of high burst , The waiting time may be a little long , Even if there are several more processing processes , There is no guarantee of immediate response ）.
redis I can't hold on （ In case of a great burst ,redis A single 10W It's the top of the sky , And it doesn't support transactions , Secsha needs transaction support , Cannot exceed inventory ）
Personal opinion ： use mysql、 But one mysql. We can allocate the inventory of this second kill commodity to 100 platform mysql database , also mysql The databases are located on different physical machines . user C End in the second kill , There will be a workflow as follows ：
1、 Check what data is still in stock （ This step is not to poll all databases , But in redis Configurations stored in the cluster , This configuration is Database identifier => Is there any stock Such as db0=>1 db1=>0 representative db0 In stock db1 period ）
2、 Randomly select one from the database set with inventory for users to use
3、 After obtaining the database address, the user goes to the database to execute the business program （ Second kill logic ）, Because the database returns no inventory exception （ Out of stock ）, Trigger the inventory closing mechanism （ Go to redis Close the inventory of this database in the cluster ）
4、 Running a task schedule in the background （ Every time 30s Do it once , The task is to check this 100 Which databases still have inventory , Which are not in stock . And keep in real time with redis Database configuration status synchronization for ）

The third step ：CDN Network distribution

Users will download a large number of static resources when browsing , Such as images ,css,js, Audio 、 Video etc. , Don't hesitate , Throw all these contents to CDN To deal with , That is, users no longer access and download these resources from the business server , But to CDN Visit and download these resources ,CDN The essence is network distribution , Better service providers have their own nodes all over the country , And theoretically, there is no upper limit to the performance , You don't have to think about CDN Performance problems of .