Integrated architecture of performance and cost: modular architecture

The little friend who heard about the modular architecture for the first time may be a little confused for a moment , I've heard of micro service architecture 、 Grid architecture and so on, but what the hell is the cellular architecture ？

In fact, many are commonly used in our mobile phones APP They are all modular architectures , Similar to Gaode navigation 、 Financial banking APP It has been a modular architecture for many years , We all know navigation APP For the request RT And request stability are very strict , If RT The delay is high , When you have passed a certain intersection , The navigation suddenly tells you that you need to turn .... At this time, are 10000 grass mud horses pouring out of your heart , So this navigation APP You must unload it immediately when you get home

What is a modular architecture ？

So called unit , It refers to a self-contained set that can complete all business operations , This collection contains all the services required by all businesses , And the data assigned to this unit .

If a business has 30 100 million data , If you deploy with a modular architecture , We can disassemble the data into 3 Share , each 10 100 million data , establish 3 A unit , These three units , Business services are exactly the same , But the data stored in the database is different , Namely 10 100 million data .

Here, the unit is split according to the data , Data needs to have separable properties , Such as user ID And other globally unique identification information , Otherwise, the unit cannot be split .

The traditional layered architecture is shown in the figure below ：

• Each layer of service has multiple nodes , The upper layer calls the lower layer with uniform traffic
• Each layer of services can be expanded or reduced separately

  The modular architecture is shown in the figure below ：

• It is also divided into multi-layer structure , But the number of different service nodes in each layer is fixed , The call chain is also fixed
• A closed loop is formed inside a unit

 

  Advantages of modular architecture ：

No matter the traditional distributed architecture or the unitary architecture, there is no perfect design , Any architecture design needs to make corresponding choices according to the changes of use scenarios

Traditional distributed architecture （SOA As a service ） The advantages are obvious , Different layered services can be deployed independently , However, the problem with this architecture is that the service level is split , The database must be divided into databases and tables , Otherwise, the increasing traffic will cause problems in the database , But the problem of the number of connections brought by the service cannot be solved by dividing the database into databases and tables , Eventually, the database connection will explode , If the database is split horizontally, the transformation and intrusion of code will also be huge

As computer hardware is getting better , The era of lower and lower costs , In fact, vertical expansion is also a new idea , Cellular architecture is a kind of vertical extension , However, this vertical expansion is different from the traditional vertical expansion , It belongs to complete logic + Data slicing

Just like cards , If a machine room is under great pressure and cannot be expanded , Then put one card in another computer room for deployment , And if some unitization occurs DOWN On the plane , The impact is also small , This method is very flexible. Whether it is horizontal expansion or multi live deployment in different places, it can be completed immediately

Process comparison ：

In a conventional distributed architecture , When there is a strong demand for bank deposits and withdrawals , Usually, the request will be completed to the database master in a service that specially operates the user account DB The operation of , Then master DB In the synchronization to replica node , Because this scenario requires high data consistency , Therefore, subsequent withdrawals will also be operated to the main DB Not to the replica DB, Although the data security guarantee conforms to , But the performance is too low , Because it is a strong operation on a piece of data in the database , It is bound to be blocked and the experience will be reduced

  If it is a modular architecture , It will become the following process

• This is the simplest unit interaction structure , The data of each unit passes MQ Sync , But this is not the only way to use it in real scenes , Usually, we will use a faster method of small batch data synchronization, that is, the synchronization of memory middleware , However, although this method is fast, it is not suitable for scenarios with a large amount of synchronized data

request Make sure users ID The corresponding routing unit remains unchanged for a period of time , Even if the attribute of the user's relevant latitude changes, the routing unit can remain unchanged for a period of time （ Navigation type APP Unitization will have more detailed logic ）, Then there is an essential difference between this process and the above process ,A Users don't care B Whether the user's account is operated successfully , as long as A If the account operation is successful, then A The account is deducted immediately ,  This unit uses the synchronization mechanism to tell other units , At this point if B To withdraw money ,B If the unit is not synchronized to the latest data B Can't see new money , This is why there is a waiting time for bank transfers

Core elements ：

• Route nearby
• Incremental synchronization of data copies

this 2 If the core elements are completed, then the cellular architecture with many activities in different places can basically run , Closed loop inside the unit + The request route is allocated nearby , this 2 This mechanism ensures that the performance is higher than that of the traditional distributed deployment , But the problems are complexity and original architecture （ Code ） The rectification cost is relatively high

Nearest route scheme ：

To meet the RT requirement , Then the services in the unit should complete the closed loop as much as possible to reduce the interaction with other units , Therefore, the routing strategy must consider this

• User mode routing
• User's common location routing
• Device routing

The usual entry packet routing strategy is still adopted , The mapping relationship of the routing table is user -- Group relationship , Group -- Unit relationship

  After the storage structure is determined , We also need to determine the storage scheme , Because the whole user - Group analysis depends on a large number of call log analysis , When the number of users is also very objective, we have to consider the storage cost , We may immediately think of ：bitmap、db、 The bloon filter

In the routing scenario, cross cell routing is considered , The impact on the user's request is only RT elevated , So as long as the probability is very small, the business is basically acceptable , Therefore, it is more appropriate to choose a Bronte filter , It can support users ID、 equipment ID And other ways of routing , The storage cost and computing performance are good , The probability of routing errors is also acceptable

 

  There is, of course, another situation , After user grouping calculation , No group was found where the user should be , At this time, we need a thorough routing strategy , That's the mold strategy