Isolation level RR, gap lock, unreal reading

Table structure and data , Which field c There is a general index on .

Here's the scene ：

In the transaction A in ,T1 The data read at any time is id=10,15,20 Three pieces of data , stay T4 The data read at any time is id=10,12,15,20, The number of data pieces read twice is different . This is the phenomenon of unreal reading .

innodb in RR Unreal reading at isolation level

It's just the scene , But it is in RR Execute at the isolation level ：

In the transaction A in ,T1 The data read at any time is id=10,15,20 Three pieces of data , stay T4 The data read at any time is id=10,15,20, The content read twice is the same .

The reason is the same , Because RR There are... Under the isolation level MVCC Mechanism , In the use of begin When starting a transaction , Carry out article 1 sql When , Create a consistent read view (read-view), Subsequent transactions are read based on this consistency view , So I can't see the changes of other transactions .

Look at the following scene ：

Will find , In the transaction A in ,T1 The data read at any time is id=10,15,20 Three pieces of data , stay T4 The data read at any time is id=10,12,15,20, The number of data pieces read twice is different .

This is because in the mysql in , All unlocked reads are consistent reads , All locked reads are current reads . The current read is not a read based on the consistency view of the current transaction , It is the latest submitted data read .

therefore , about innodb Come on , The isolation level is RR When , Current reading will produce unreal reading .

The problem of unreal reading

Here's the scene ：

From business A The angle of , stay T2 Always query c=10, Only id=10 This line of data , So change this line of data , Set up d=0, But in the end id=10、12 Two lines of data d All are 0, Semantically , Easy to cause confusion .

Except for that , The bigger problem of unreal reading is the inconsistency of data .

Here's the scene ,binlog Format =statement：

After the execution , Master database data ：

binlog

insert into t (id,c,d) values(12,12,12)
update t set d=0 where c>=10 and c<15

Will find ,binlog Recorded in the sql Execute in sequence and actual main library sql The order of is different , When replaying from the Library binlog When ,

There will be inconsistency between master and slave data ,id=12 This data is different in master database and slave database . If mysql For some reason, it needs to be based on binlog Data recovery , The recovered data is also inconsistent with the beginning .

ps： So , If you use mysql The segregated sector is CR, that binlog The format must be ROW.

 

RR The isolation level solves the unreal reading problem -- Clearance lock

Phantom read problem refers to the fact that the same two reads in a transaction read different data ,innodb The solution is to lock and block .

The same scenario as above ：

Business A stay T2 When it comes to execution , Would be right id=10 This line is locked up , And right c=(5,10),(10,15) This gap is shoulder locked , When a transaction B stay T3 When inserting , Due to clearance lock conflict , Will block waiting , Wait for business A Post commit transaction B In order to proceed .

This is the time ,binlog The recorded statements are in the same order as the actual execution of the main library ：
 

update t set d=0 where c>=10 and c<15

insert into t (id,c,d) values(12,12,12)

Does this seem to be at the isolation level RR Next ” perfect “ Solved the unreal reading problem ？

Look at the scene ：

stay T2 Read the following at any time ：

stay T5 Read the following at any time ：

Two reads of the same condition , Read different data pieces , This situation can only be solved by the serial isolation level .

So , Under the segregated sector , It solves the unreal reading problem of some scenes , Ensure the correctness of the data , But because the read in the transaction needs to see the modification of this transaction , There are also some unreal scenes that cannot be solved . To sum up ：

1. stay RR Below grade , Because of MVCC, Read the consistent read view based on the current transaction , You can't see the changes of other transactions ( Other things have changed , Nothing has changed , Is visible ), To some extent, it solves the unreal reading problem of some scenes .ps： That's exactly what it is. mvcc, Realize the lock free reading of consistent reading .
2. Introduce clearance lock , To ensure the binlog Format =statement Under the circumstances ,binlog It's also true . Due to the existence of clearance lock , Blocking the insertion of data in the gap , For the unreal reading problem of current reading , The gap lock can ensure that there is no unreal reading .

in addition , Why is gap lock not a mechanism to solve unreal reading , No, just right insert It works , Yes update It's the same ：

This will also be blocked , Although for data , Business B Not inserting data , But for indexes c Come on , Is in the (10,15) A data is inserted into this gap , Because there is a gap lock in this gap , It will block .

Clearance lock

innodb The data of is organized by index , The data is placed on the primary key index tree ：

If there are other secondary indexes , For example, in the field c An index is added ： There will be another one B+ Trees , The leaf node stores the primary key .

Then we call it clearance lock , Whose gap is it ？ For example, enter the following statement , What is the situation of locking ：

update t set d=0 where c=10;

Give a conclusion first ：

1. The gap is loaded on the index , The index used is added to the index .
2. For modifying class operations , Row locks must be added to the primary key index tree .

therefore , Lock diagram ：

You can verify ：

The initial data are as follows ：

The implementation is as follows sql：

T3 Business B It won't block , Successful implementation , The reason is that the gap lock is added to the index c On , There is no gap on the primary key .

If the transaction B Execution is ：insert into  t (id,c,d) values(14,10,10), Will be blocked . In fact, as long as here c Column Values in (10,15) Between , Will be blocked .

So , Gap refers to the index column used , Is the index on this column .