Several abnormal scenarios of things system

• background

The author recently reviewed multi version concurrency control （MVCC）, Here is a brief summary of the relevant contents , Let's start with the isolation level .

• Why talk about isolation , What is the scene ？

First , As a transaction system , If... Per unit time , Only one transaction is running , There is no isolation .

But actually , Real transaction system for actual production , In a unit time, multiple transactions are reading and writing resources at the same time .

Then there must be a concurrent read / write problem for the same resource , If you don't control it, you will make mistakes .

Therefore, the scenario discussed here is the data processing scenario in the multi transaction concurrent environment .

• What problems can isolation solve ？ Why should isolation be graded ？

What kind of transaction system is reasonable , Correct , Or it can accurately abstract the logic of the real world ？ Satisfy ACID characteristic . Satisfy C To be realized I, Therefore, the purpose of isolation is to meet C.

However, it is necessary to answer why isolation should be graded , We must never be satisfied C The abnormal scene of must be disassembled for analysis .

• Several abnormal scenarios of transaction system

1、Lost Update Update missing

What is update loss ？ Two concurrent transactions T1、T2, Write to the same resource without any lock .T1 The result of the change is T2 Transaction tampering , This is the missing update . The point is that there is no means of isolation , Pure concurrency . This scenario , Can not appear in any reasonable application system , It is a fundamental major error .

2、Dirty Read Dirty reading

Definition of dirty reading , Is to read dirty data . First of all, let's clarify what dirty data is .

In a transaction system , Dirty data refers to uncommitted data , Intermediate state data . Then read the data that the transaction did not commit , Is to read dirty data , It's called dirty reading . You can use the following diagram to show ：

This is from the Internet

There are two things T1 T2 At the same time ,T1 To change the a Value , When not submitted ,T2 It reads a By T1 Changed value . If at this time ,T1 Rolled back , that T2 Read the a Value , Is a value that is not considered to exist , Or read the intermediate state of the transaction , This is dirty reading .

We can simply use the method of locking to understand that ： All transactions here are on access to critical resources , Only the read lock is added . It means , Other transactions can read uncommitted data at any time （ Because read locks are compatible ）, But you cannot write data that has been modified （ Because read lock and write lock are incompatible ）. So in the dirty reading scenario , The general default is that , The problem of missing updates has been solved .

3、Un-repeatable Read / Read skew Non repeatability / Reading partial order

Unrepeatable refers to the scene , Business T1 The data read is due to other transactions T2 After modification and submission , When T1 When reading again , There are different results . Let's first look at , The difference between unrepeatable and dirty reading is that ： It can't be read repeatedly , All the results read , Are submitted data , The dirty data is invisible . You can use the following figure to show ：

This is from the Internet

Through the above figure, we can simply understand , To achieve the difference between non repeatability and dirty reading , Is in T2 When a transaction prepares to write data , The read lock is upgraded to write lock , bring T1 Can't read T2 The data value before the transaction is committed , But when T2 After all the locks have been released ,T1 When reading data again , It is already a modified version .

4、Phantom Read Fantasy reading

Unreal reading refers to ,T1 Different points in time during the execution of a transaction , The result sets obtained with the same predicate conditions are different , This is limited to extra data , Not the data . What we need to make clear here is , The difference between unreal reading and unrepeatable reading mentioned above is ： Under the unreal reading scene , There is no problem that can not be repeated , That is, the value of the same primary key is the same in two queries , There are only some more records （ Why not lack data ？）. You can use the following diagram to show ：

This is from the Internet

We can simply understand it as , Business T1 On the resources to be operated , Read only data is locked , Write lock the data to be modified , And they are all released after the transaction is committed . Therefore, other transactions cannot modify the same resource , So in the scenario of unreal reading , Non repeatability does not happen . however T1 But the newly generated data row cannot be locked , Therefore, the difference between unreal reading and unrepeatable degree is reflected here, that is, the result set of the same predicate is different , Generally, the result set of the last time is a superset of the result set of the previous time .

• summary

Disassemble from the above analysis , And understand these abnormal scenarios from the perspective of locking , You can find one Progressive relationship ： When the last problem is solved , Another new problem has been discovered . That's why , The concept of different isolation levels with progressive relationships defined according to these exception scenarios .

In the next chapter < Isolation level of the transaction system > In the analysis of , We can still find , Each isolation level , Solve a level problem . Coming soon ~

KunlunDB The project is open source

【GitHub：】
https://github.com/zettadb

【Gitee：】
https://gitee.com/zettadb

END