Introduction to distributed transactions (Seata)

One 、 brief introduction

1、 Local and distributed transactions

1.1 Business

Database transactions ( abbreviation ： Business ,Transaction) Refers to the database ⾏ In process ⼀ A logical unit , from ⼀ A limited sequence of database operations .
Transactions have the following four characteristics , It's used to be called ACID characteristic ：
primary ⼦ sex (Atomicity)： Business as ⼀ The whole is held ⾏, The operations on the database contained in it are either all performed ⾏, Or none of them ⾏.
⼀ Sexual nature (Consistency)： Transactions should ensure that the state of the database changes from ⼀ individual ⼀ Cause the state to change to another ⼀ individual ⼀ To state .⼀ To state means that the data in the database should be full ⾜ Integrity constraints . besides ,⼀ There's more to sex ⼀ Layer semantics , The intermediate state of a transaction cannot be observed ( This layer of semantics also says that it should belong to the original ⼦ sex ).
Isolation, (Isolation)： Multiple transactions execute concurrently ⾏ when ,⼀ The execution of a transaction ⾏ It shall not affect the execution of other affairs ⾏, Just like this ⼀ An operation is being performed by the database ⾏⼀ sample .
persistence (Durability)： Changes made to the database by committed transactions should be persisted in the database . At the end of the transaction , This operation will be irreversible .

1.2 Local transactions

At first , Transactions are limited to single transactions ⼀ Access control of database resources , After the service of Architecture , The concept of transactions extends to services . If will ⼀ A list ⼀ Service operation of as ⼀ One transaction , Then the whole service operation can only involve ⼀ A list ⼀ Database resources , This type is based on a single service order ⼀ Database resource access transactions , It's called local affairs (Local Transaction).

1.3 Distributed transactions

​ Distributed transaction refers to the participants of transaction 、⽀ Server holding transactions 、 Resource servers and transaction managers are located on different nodes of different distributed systems , And belong to different should ⽤, Distributed transactions need to ensure that either all of these operations succeed , All or nothing . In essence , Distributed transaction is to ensure the data of different databases ⼀ Sexual nature . When ⼀ Service operations access different database resources ,⼜ When you want to have transactional access to them , You need to pick ⽤ Distribute transactions to coordinate all transaction participants .

The figure below reflects this ⼀ A distributed transaction that spans multiple services ：

If will go up ⾯ These two scenarios (⼀ Services can be adjusted ⽤ Multiple database resources , It can also be adjusted ⽤ Other services ) Combined in ⼀ rise , In this regard ⾏ extend , The participants of the whole distributed transaction will form a tree topology as shown in the figure below .

2、 Distributed transaction theory

2.1 CAP Theorem

CAP The theorem is in 1998 California ⼤ Computer scientists who study Eric Brewer （ Egyptian ⾥ g . Brewer ） Put forward , There are three indicators of a distributed system , These three indicators cannot be achieved at the same time .

• Consistency ⼀ Sexual nature
• Availability can ⽤ sex
• Partition tolerance Partition fault tolerance

Partition fault tolerance （Partition tolerance）

Distributed system cluster ,⼀ A broken machine should not affect other machines

⼤ Most distributed systems are distributed in multiple ⼦⽹ Collateral . Every ⼦⽹ Luo is called ⼀ Districts （partition）. Partition fault tolerance means , Interval communication may fail .⽐ Such as ,⼀ Servers are in China , another ⼀ Servers are located in the United States , These are the two areas , There may be ⽆ France communications .

⼀ In general , Partition fault tolerance ⽆ There is no way to avoid , So we can think that CAP Of P Always become ⽴.CAP The theorem tells us , The rest C and A ⽆ At the same time .

can ⽤ sex （Availability）

Just receive ⽤ User's request , The server has to respond ⽤ Users can choose to go to G1 or G2 Initiate read operation . No matter which server , Just receive the request , You have to tell ⽤ Household , What is the v0 still v1, Otherwise, you will be dissatisfied ⾜ can ⽤ sex .

⼀ Sexual nature （Consistency）

⼀ You will be able to read the latest data

Consistency in ⽂ be called ⼀ Sexual nature . intend , Read operation after write operation , The value must be returned .

⼀ Consistency and predictability ⽤ Sexual ⽭ shield
⼀ Consistency and predictability ⽤ sex (A), Why can't it be at the same time ⽴？ The answer is simple , Because communication may fail （ That is, partition fault tolerance ）.
If you promise G2 Of ⼀ Sexual nature , that G1 Must be in writing operation , lock G2 Read and write operations . Only after data synchronization , To reopen reading and writing . Lock in period ,G2 Can't read or write , Nothing to ⽤ sex (CP).
If you promise G2 Yes ⽤ sex , Then it's bound to fail to lock in G2, therefore ⼀ It's impossible to be sexual ⽴(AP).
in summary ,G2 ⽆ At the same time ⼀ Consistency and predictability ⽤ sex . Only ⼀ individual ⽬ mark . If you pursue ⼀ Sexual nature , that ⽆ Method to ensure the scalability of all nodes ⽤ sex ; If we pursue the scalability of all nodes ⽤ sex , Then you can't do it ⼀ Sexual nature .

2.2 Base theory

BASE： Full name ：Basically Available( Basically ⽤),Soft state（ Soft state ）, and Eventually consistent（ Final ⼀ Sexual nature ） Abbreviations of three phrases , Come on ⾃ ebay The architect of .BASE The theory is right CAP in ⼀ Consistency and predictability ⽤ The result of the sex trade-off , It comes from ⼤ Type Interconnection ⽹ Summary of distributed practice , Is based on CAP The theorem evolves gradually ⽽ To the . Its core ⼼ Thought is ：

both ⽆ The law is strong ⼀ Sexual nature （Strong consistency）, But everyone should ⽤ According to ⾃ Personal business characteristics , Mining ⽤ proper ⽅ Formula to make the system reach the final ⼀ Sexual nature （Eventual consistency）.

Basically Available( Basically ⽤)

understand : Allow service degradation or allow response time to be affected ⼀ Fixed loss

What is basic ⽤ Well ？ Suppose the system , There was an unforeseen breakdown , But you can still ⽤, phase ⽐ A more normal system ⽽⾔：

1. Loss of response time ： Normal search engines 0.5 Seconds is returned to ⽤ Household results ,⽽ Basically ⽤ The search engine can be found in 1 Seconds for ⽤ Return results .
2. Loss of function ： stay ⼀ E-commerce ⽹ On the site , Under normal circumstances ,⽤ You can successfully complete every ⼀ Orders , But here we are ⼤ Promotion period , In order to protect the stability of the shopping system , Some consumers may be guided to ⼀ Downgrades ⻚⾯.

Soft state（ Soft state ）
understand : Appears when data synchronization is allowed ⼀ Time delay
What is soft state ？ Relative to the original ⼦ sex ⽽⾔, Data copies of multiple nodes are required ⼀ To , This is a ⼀ Kind of “ Hard state ”.
Soft state means ： Allow data in the system to have intermediate states , It is considered that this state does not affect the overall availability of the system ⽤ sex , That is to say, the system is allowed to have data delay in data copies of multiple different nodes .
Eventually consistent（ Final ⼀ Sexual nature ）

understand : after ⼀ After synchronizing data for a period of time , Finally, we can achieve ⼀ individual ⼀ The state of being .

The system can ensure that there is no other new update operation , The data finally ⼀ It is certain to achieve ⼀ The state of being , Therefore, all client data access to the system can finally get the latest value .

3、 Distributed transaction solutions

3.1 be based on XA Two phase submission of the agreement

You know XA Distributed transactions in the specification include AP,RM,TM form ：

Among them ⽤ Program (Application Program , abbreviation AP)： AP Define transaction boundaries （ Define transaction start and end ） And access the resources within the transaction boundary .

Explorer (Resource Manager, abbreviation RM)： Rm Manage computer shared resources , Many software can access these resources , Resources include ⽐ Such as a database 、⽂ Piece system 、 Printer server, etc .

Transaction manager (Transaction Manager , abbreviation TM)： Responsible for managing the overall affairs , Assign transactions only ⼀ identification , Monitor the execution of transactions ⾏ speed of progress , And responsible for the submission of affairs 、 Roll back 、 Failure recovery, etc .

⼆ Phase agreement :
The first ⼀ Stage TM： All required RM Prepare to submit the corresponding transaction points ⽀, inquiry RM Is there any way ⼒ Ensure the successful submission of transactions ⽀,RM according to ⾃⼰ The situation of , If you judge ⾃⼰ Into the ⾏ Of ⼯ Work can be submitted , That's right ⼯ Make content into ⾏ Persistence , And give TM receipt OK; If not, give it TM The receipt of NO.RM After sending a negative reply and rolling back the already ⼯ After work , You can discard this transaction ⽀ Information. .
The first ⼆ Stage TM： According to the stage 1 each RM prepare Result , Decide whether to commit or roll back the transaction . If all RM all prepare success , that TM Inform all RM Into the ⾏ Submit ; If there is RM prepare receipt NO Words , be TM Inform all RM Roll back ⾃⼰ Affairs of ⽀.
That is to say TM And RM The agreement is submitted in two stages ⾏ Interaction .

advantage ： Try to ensure the strength of the data ⼀ Cause , Suitable for strong data ⼀ The request is very ⾼ Key areas .（ In fact, it can't be 100% Strong guarantee ⼀ Cause ）
shortcoming ： The implementation is complex , Sacrifice, but ⽤ sex , The impact on performance is relatively ⼤, Not suitable for ⾼ Concurrent ⾼ Performance scenarios .

3.2 TCC Compensation mechanism

TCC It's actually picking ⽤ Compensation mechanism , Its core ⼼ Thought is ： For each operation , You have to register ⼀ Corresponding confirmation and compensation （ revoke ） operation . It's divided into three stages ：

• Try The stage is mainly to test the business system and reserve resources .
• Confirm The phase is mainly to confirm and submit the business system ,Try Stage execution ⾏ Succeed and start to implement ⾏ Confirm When the stage , Default Confirm There is no mistake in the stage . namely ： as long as Try success ,Confirm⼀ Success is certain .
• Cancel The stage is mainly in business execution ⾏ error , Execute when rollback is required ⾏ Our business is cancelled , Reserve resource release .

A Want to B Transfer accounts , Ideas ⼤ It's all about ：

 We have ⼀ A local ⽅ Law ,⾥⾯ In turn ⽤
1、⾸ First in  Try  Stage , First adjust ⽤ Remote access ⼝ hold  B and  A The money is frozen .
2、 stay  Confirm  Stage , Of board ⾏ Remote dispatching ⽤ Transfer operation , The transfer succeeded in ⾏ thaw .
3、 If the first 2 Step execution ⾏ success , Then the transfer is successful , If the first ⼆ Step execution ⾏ Failure , Then tune ⽤ Remote freezing connection ⼝ Corresponding thawing ⽅ Law  (Cancel).

advantage ： phase ⽐ Two-phase commit , can ⽤ sex ⽐ Stronger

shortcoming ： Data ⼀ Poor consistency ⼀ some .TCC Belong to should ⽤ Layer of ⼀ Kind of compensation ⽅ type , So programmers need to write a lot of compensation code in the implementation , stay ⼀ In some scenes ,⼀ Some business processes may ⽤TCC Not very well defined and handled .

3.3 The message is ultimately consistent

The news finally ⼀ Consistency should be made by the industry ⽤ The most , Its core ⼼ The idea is to split distributed transactions into local transactions ⾏ Handle , This kind of thinking comes from ebay. We can start from the bottom ⾯ See... In the flow chart of ⼀ Some details ：

The basic idea is ：
news ⽣ production ⽅, Additional construction is required ⼀ A message table , And record the message sending status . The message table and business data should be in ⼀ One transaction ⾥ Submit , That is to say, they are going to ⼀ A database ⾥⾯. Then the message will pass MQ Consumption sent to message ⽅. If the message fails to be sent , Will enter ⾏ Try sending again .
News consumption ⽅, This message needs to be processed , And finish ⾃⼰ Business logic of . At this time, if the local transaction is successful , Indicates that the process has been successful , If the processing fails , Then it will try again ⾏. If it is business ⾯ The failure of the , You can give ⽣ production ⽅ send out ⼀ Business compensation messages , notice ⽣ production ⽅ Into the ⾏ Rollback and other operations .
⽣ production ⽅ And consumption ⽅ Scan local message table regularly , Send messages that have not been processed or failed ⼀ All over . If there are reliable ⾃ Dynamic reconciliation
Replenishment logic , such ⽅ The case is still ⾮ Chang Shi ⽤ Of .
advantage ： ⼀ Kind of ⾮ Often classic implementation , Avoid distributed transactions , Finally ⼀ Sexual nature .
shortcoming ： The message table is coupled to the business system , If there is no encapsulated solution ⽅ case , There will be a lot of chores to deal with .

Two 、 Distributed transactions seata

1、 brief introduction

Seata（ Original name Fescar） Yeah ⾥18 Open source framework for distributed transactions .Fescar Open source has a great impact on the field of distributed transaction framework ⼤. do

Open source ⼤ Household ,Fescar Come on ⾃ o ⾥ Of GTS, Experienced good ⼏ Secondary double ⼗⼀ The test of ,⼀ Open source has attracted much attention . later Fescar Renamed as

Seata.

Fescar Although it is ⼆ Phase commit protocol for distributed transactions , But it solved XA Of ⼀ There are some shortcomings :

• A single point of the problem ： although ⽬ front Fescar(0.4.2) Or just server Of , however Fescar Officer, ⽅ It is expected to be in 0.5.x Launched in HA-Cluster, Then we can solve the single point problem .

• Synchronous blocking ： Fescar Of ⼆ Stage , Its second ⼀ During the phase, the local transaction has been committed to release resources , Don't like XA There will be two more prepare and commit Stage resources are locked , also Fescar,commit Is asynchronous operation , It also improves performance ⼀⼤ The key .

• Data not available ⼀ Cause ： If there's a part of commit Failure , that fescar-server According to the current transaction mode and sub ⽀ The result of the return state of the transaction comes in ⾏ Different retry strategies . also fescar The local transaction of will be in ⼀ When the stage goes ⾏ Submit , In fact, just look at the database commit The database was already ⼀ To .

• Can only ⽤ Yu Dan ⼀ database ： Fescar There are two modes ,AT and TCC. stay AT In mode, transaction resources can be any ⽀ a ACID The database of , stay TCC There is no limit to transaction resources in mode , It can be cache , It can be ⽂ Pieces of , It can be others and so on . Of course, these two modes can also be mixed ⽤.

meanwhile Fescar Also kept close 0 Business ⼊ Advantages of invasion , Just a simple configuration Fescar Data proxy and add a comment , Add ⼀ individual Undolog surface , We can achieve what we want ⽬ Of .

2、 Realization principle

Fescar take ⼀ Local transactions as ⼀ Distributed transactions are divided into ⽀, So if ⼲ Local transactions distributed in different microservices together form ⼀ A global transaction , The structure is as follows .

TM： Global transaction manager , On the label fescar The server of distributed transactions is turned on , And send the global transaction to TC Transaction control side management

TC： In transaction control ⼼, Control the commit or rollback of global transactions . This component needs to be unique ⽴ Deployment maintenance ,⽬ The first one ⽀ Stand alone version , The subsequent iteration plan will have a cluster version

RM： Explorer , Mainly responsible for sub ⽀ Escalation of transactions , Management of local transactions

⼀ This paragraph briefly describes the implementation process ：

 Service start ⽅ Initiate a global transaction and register to TC.
 In tune ⽤ Collaborative services , The transaction of collaborative service is divided into ⽀ The transaction will complete the phase first ⼀ Commit or roll back transactions for , and ⽣ Transaction rollback undo_log⽇ Records ,
 At the same time, register the current collaboration service to TC And report its transaction status , Merge into the same ⼀ In the global transactions of businesses .
 If there is no problem at this time, continue ⼀ Coordination of collaborative services ⽤, Points of any collaborative services during ⽀ Transaction rollback , You'll be informed TC,TC After notifying the global transaction that all contained have been completed ⼀ Stage submitted sub ⽀ Transaction rollback .
 If all points ⽀ Everything is normal , Finally, return to the global transaction initiation ⽅ when , Will also inform TC,TC After notifying all the points contained in the global transaction ⽀ Delete rollback ⽇ Records .
 In this process, in order to solve the problems of write isolation and degree isolation, it will involve TC Managed global lock .

3、Fescar Pattern

Fescar The implementation of distributed transactions provides 2 Patterns ,AT Patterns and TCC Pattern ：

3.1 TCC Pattern

TCC Pattern ： TCC Compensation mechanism , Cause ⼀ Definite invasion ⼊, It is difficult to realize ⼤, such ⽅ The formula is not recommended , however TCC The pattern is characterized by performance ⾼.
TCC The code of the mode part is as follows ： You can see clinging ⾏ Transaction rollback , All need to be implemented according to different stages ⾏ State judgment , Invade ⼊ Business code .

@Override
@GlobalTransactional
public boolean transfer(final String from, final String to, final double amount) {
    
    // Deduction participant ,⼀ Stage execution ⾏
    boolean ret = firstTccAction.prepareMinus(null, from, amount);
    if (!ret) {
    
        // Deduction participant ,⼀ Phase failure ;  Rollback local and distributed transactions 
        throw new RuntimeException(" account number :[" + from + "]  Withholding failed ");
    }
    // Plus participants ,⼀ Stage execution ⾏
    ret = secondTccAction.prepareAdd(null, to, amount);
    if (!ret) {
    
        throw new RuntimeException(" account number :[" + to + "]  Advance collection failed ");
    }
    System.out.println(String.format("transfer amount[%s] from [%s] to [%s] finish.", String.valueOf(amount), from, to));
    return true;
}

3.2 AT Pattern

AT Pattern ： Focus on more DB Data accessed ⼀ Sexual nature , Come true ⽐ Simpler , Invasion of business ⼊ a ⼩, But the performance is not TCC⾼, This mode is recommended ⼤ Home agent ⽤.
AT The code of the mode part is as follows ： There is no need to pay attention to ⾏ state , Invasion of business code ⼊ a ⼩.

@GlobalTransactional(timeoutMills = 300000, name = "dubbo-demo-tx")
public void purchase(String userId, String commodityCode, int orderCount) {
    
    LOGGER.info("purchase begin ... xid: " + RootContext.getXID());
    storageService.deduct(commodityCode, orderCount);
    orderService.create(userId, commodityCode, orderCount);
    throw new RuntimeException("AT  Pattern hair ⽣ abnormal , Roll back the transaction ");
}

3.3 AT Pattern design ideas

AT Kernel of mode ⼼ It's for business ⽆ Invade ⼊, yes ⼀ Improved two-stage submission .

The first stage ：

nucleus ⼼ It's about business sql Into the ⾏ analysis , convert to undolog, Two phase commit often locks resources until ⼆ Stage the actual commit or rollback operation ,⽽ With rollback ⽇ After Zhi , In the first ⼀ Phase release lock on resource , Reduced lock range , carry ⾼ efficiency , Even if the ⼆ Stage hair ⽣ Exceptions require rollback , Just find the right undolog The corresponding data in and inversely parsed into sql To achieve rollback ⽬ Of .Seata Through proxy data sources, the business is sql The hold of ⾏ It can be interpreted as undolog To update business data at the same time ⼊ library , To achieve the business ⽆ Invade ⼊ The effect of .

The second stage ：

If the resolution is submitted as a whole , At this point ⽀ The transaction has now been committed , There is no need for synchronous coordination （ Just need asynchronous cleanup rollback ⽇ Records ）,Phase2 Sure ⾮ Often finish quickly .

If the resolution is a global rollback ,RM Received a rollback request from the coordinator , adopt XID and Branch ID Find the corresponding rollback ⽇ Records , By rolling back records ⽣ Update in reverse SQL And hold ⾏, To complete points ⽀ Roll back of .

4、 Official information

Official website
GitHub Address

https://seata.io/zh-cn/
https://github.com/seata/seata

5、 Quick start

pom.xml

<dependency>
    <groupId>com.alibaba.cloud</groupId>
    <artifactId>spring-cloud-starter-alibaba-seata</artifactId>
    <version>${spring-cloud-alibaba.version}</version>
</dependency>

Code ：

package com.work.order.service;

import com.work.order.model.Order;
import io.seata.spring.annotation.GlobalTransactional;
import org.springframework.stereotype.Service;
import org.springframework.transaction.annotation.Transactional;

import java.math.BigDecimal;
@Service
public class OrderService {
    

    @GlobalTransactional
    @Transactional(rollbackFor = Exception.class)
    public void placeOrder(String userId, String commodityCode, Integer count) {
    
        BigDecimal orderMoney = new BigDecimal(count).multiply(new BigDecimal(5));
        Order order = new Order().setUserId(userId).setCommodityCode(commodityCode).setCount(count).setMoney(
            orderMoney);
        orderDAO.insert(order);
        stockFeignClient.deduct(commodityCode, count);

    }
}