Weibo comments on high-performance and high availability architecture design (module 5 of architecture practice camp)

One 、 Scenario calculation estimate

Business background

according to 《 Microblogging 2020 User development report 》, Weibo in 2020 year 9 Live every month 5.11 Billion , Diurnal activity 2.24 Billion , Therefore, the following assessment is based on daily activities 2.5 Consider the magnitude of billion .

Comment on

Suppose that the average person sends 5 comments , The daily sending volume of microblog comments is about 2.5*5=12.5 Billion bars .

Most people use Weibo in the morning 8

:

00~9

:

00 spot , At noon, 12

:

00~13

:

00, evening 20:00~22:00, Assume that the proportion of the total amount of microblogging in these time periods is 60%, Then this 4 The average number of microblog comments in an hour  TPS  The calculation is as follows ：

12.5 Billion  * 60% / (4 * 3600) ≈ 50 K/s.

See the comments

The estimated number of times to watch Weibo is 250 Billion , Because comments support paging and folding , Therefore, it is possible to request the comment list many times while viewing the comment , Assume here that the average is 5 Time , It is estimated that the number of requests can reach 1250 100 million times , Most people's time period of watching microblog basically coincides with that of sending microblog , So the average number of microblog comments  QPS  The calculation is as follows ：1250 Billion  * 60% / (4*3600) = 5000K/s.

Two 、 High performance architecture design

Comment on

【 Business characteristic analysis 】

Commenting is like microblogging , Is a typical write operation , So you can't use cache , Load balancing can be used , But the importance and timeliness are weaker than that of microblogging , Allow certain delay and data repetition or loss , So you can use write buffer .

【 Architecture analysis 】

Over 100 million users , You should use a multi-level load balancing architecture , Cover  DNS -> F5 -> Nginx ->  Multilevel load balancing of gateway , And use more powerful message queues on the storage side .

【 Architecture design 】

1.  Load balancing algorithm selection

When posting microblog, it depends on login status , The login status is generally saved in the distributed cache , So when commenting , Send the request to any server , Choose here “ polling ” perhaps “ Random ” Algorithm .

2.  Write buffer selection

Use Kafka, The buffer capacity is not limited , write in Kafka And then directly return success to the user , Then write to the storage asynchronously .

3.  Estimate the number of business servers

Commenting can be split into two asynchronous processes: request processing and data writing ：（1） There are several key processes involved in processing requests ： Content review （ Rely on the audit system ）、 Write data to storage （ Mainly Kaka）, Due to the use of a more powerful Kafka, It is estimated that a service can process 1000 A request , complete 50K/s Of  TPS, need 50 Servers ;（2） For data writing , It mainly depends on the writing of relational database and cache , Generally, it can also reach 1000 Left and right TPS. So in general ,50 platform

The server

that will do , Plus a certain amount of reserve ,55 This server is almost .

See the comments

【 Business characteristic analysis 】

Reading comments is a typical reading scene , Suitable for cache architecture , At the same time, due to the large number of requests , Load balancing architecture also needs .

【 Architecture analysis 】

1.  Over 100 million users , You should use a multi-level load balancing architecture ;

2.  The requested quantity reaches 1250 Billion , You should use a multi-level cache architecture , In especial  CDN  cache , Is the core of cache design .

【 Architecture design 】

1. Load balancing algorithm selection

Visitors can directly read microblog comments , Therefore, sending a request to any server can , Choose here “ polling ” perhaps “ Random ” Algorithm .

2.  Estimate the number of business servers

hypothesis  CDN  Capable of carrying 90% Of users , So the rest 10% The request to read comments enters the system , Ask for  QPS  by 5000K/s * 10% = 500K/s, Suppose that the processing capacity of a single business server is 1000/s, Then the number of machines is 500 platform , according to 20% Reserved quantity of , The final number of machines is 600 platform .

3、 ... and 、 The overall architecture

Multiple levels are responsible for balancing the overall architecture

Look at the comment's multi-level cache architecture

Four 、 Hot event high availability architecture design