[re understand the communication model] the application of reactor mode in redis and Kafka

Re understand the communication model

Each framework has its own communication model , Used to handle network events . It's just that different frameworks depend on their own priorities , The requirements and implementation of network communication are different .

The development of our network communication mode ：

• Single thread ： Only one request can be processed at a time , Other requests are blocked , Low processing efficiency ;
• Multithreading ： One request, one connection , Create a large number of threads , It brings thread switching and maintenance problems , High system complexity ;
• Thread pool ： Thread reuse , Thread management , But thread pool resources are limited , There are also thread waiting problems .
• Reactor： Modern high-performance Networks IO Pattern , Use the event driven mechanism .

0,Reactor

Reactor Pattern , Reactor mode , It is an event driven network request processing mode , It is implemented in many high-performance network communication frameworks .

stay Reactor In the pattern , Divide the network processing requests into one event , Specifically, it includes reading events 、 Write events 、 Connection event .

In this mode , There are three key components ：reactor、acceptor、handler, Respectively correspond to listening and distributing requests , Connect requests and process requests .

The basic flow ： Client requests are first reactor receive ,reactor Distribute the request according to the request type , The connection request will be distributed to acceptor Components , He always monitors the network status , For read requests and write requests, it is given to handler Handle .

meanwhile , According to the different ways of handling ,reactor Modes can be divided into many kinds ：

single reactror Single thread ： One reactor, One handler;

shortcoming ： Single thread , The processing capacity is slow （ Compared with these three ）,redis6.0 This method was used before , such as keys * It is easy to block

single reactror Multithreading ： One reactor,handler Drop the request to worker Thread pool , Read / write events are handled by the thread pool

Take full advantage of multi-core capabilities ,redis6.0 The multi threading mechanism introduced is this mode , however redis The logic of specific command execution is still single thread .

many reactor Multithreading ：mainReactor be responsible for acceptor Handle ,subReactors be responsible for handler Handle

You can see that the performance is very strong ,kafka,netty Have used this mode .

1,Redis Medium Reactor

redis in reactor The mode is embodied in IO Based on multiplexing , say concretely redis For different os Of IO Multiplexing mechanism realizes different reactor Model , such as mac os Of kqueue; Let's mainly talk about Redis be based on Linux Of epoll Functionally implemented reactor Pattern . Be careful ： Only epoll It adopts the idea of event driven .

that redis How to reuse epoll What about ？

First through a while Loop constantly listening for network events , Implement different processing strategies according to different events . There are three main types of events , Empty event , Continue polling ; Time event , Call the time handler ; Read and write events , You need to call os Of io The multiplexing mechanism performs event capture . Here is a brief mention of multiplexing functions epoll workflow ：

epoll First, through epoll_ctl Function to register a file descriptor , And listen for other file descriptor States , If a fd be ready , The callback mechanism will be adopted callback, Activate this fd And put it in the ready list , Then it is handled by specific threads .

2,Kafka Medium Reactor

kafaka It is also based on Reactor Pattern designed its own network communication framework , Bottom use NIO Medium selector function .

First ,broker There was a stockerserver Will listen and dispatch requests ,acceptor Processing connection requests , Rotate and select one processor Handle this connection . This network thread is handled by parameters num.network.threads=3 control . And then in IO Thread by work Thread specific execution ,num.io.threads=8. After execution , Will reponse Put it in the response queue , And then back to the client . among , Each network thread has its own response queue .

besides ,kafka Added a cache layer ,RequestChannel, Used to cache those delayed requests , Used to handle settings ack=all Of produce request

Some details in the source code ：

#SocketServer, Managed two sets acceptor,processor,RequestChannel; Including data requests data-plane, Control request control-plane;

#KafkaRequestHandlerPool,IO Thread pool , The thread that actually processes the request ,KafkaRequestHandler Match and process different requests .

kafka The request process of is the process of network communication , Let's talk about it again , To handle different requests ,kafka Priority the request . It is divided into

Data request ： Such as produce request , Write messages to disk ;fetch request , Pull messages from disk or cache pages

Control request ： Such as updating the copy and ISR Of leaderAndlsr request , And ordering copies offline stopReplica Request etc.

The benefit of request priority is , Some requests can be processed first , Prevent invalid request processing . such as , When deleting a topic , We will use stopreplica request , So even if it's ack=all Under the circumstances , There is no need to wait ack Complete , You can execute the deletion request , Speed up the process , Avoid invalid operations . The implementation mechanism is , Start the two request processing models mentioned above , And configure different listeners Listen for the request , Then the specified .

Reference link ：

select、poll、epoll Detailed analysis of

reactor Mode introduction and classification

Redis in reactor Mode explanation and source code analysis