Mvcc and bufferpool (VI)

Database notes

• Mysql Index bottom layer （ One ）
• Explain Tools and index optimization （ Two ）
• SQL stay MySQL Execution process in （ 3、 ... and ）
• MySQL Index optimization （ Four ）
• MySQL Transactions and locks （ 5、 ... and ）
• MVCC and BufferPool（ 6、 ... and ）

MVCC

MVCC = undo + read-view

Service for isolation

Definition ：

MVCC： Multi version concurrency control mechanism , stay Read committed and repeatable At the isolation level MVCC Mechanism

characteristic ：

Not through locking , To achieve isolation between transactions

Realization principle

undo Log version chain

Definition ：

undo Log version means A row of data is modified by multiple transactions in turn , **MySQL The data before each operation will be retained .** And use trx_id【 Business id】 and roll_pointer 【 rollback pointer 】 take undo Logs are concatenated , Form a historical version chain .

Special ：

In case of deletion

• The latest data in the version chain will be copied
• And then trx_id Modify to delete trx_id
• In this article In the recorded header information delete_flag The flag bit is set to true

characteristic ：

• For multiple transactions , There is only one log version chain
• Business id And rollback pointer Are hidden fields

Consistency view read-view

Definition ：

When the transaction is on , Execute any query SQL, Generates the of the current transaction Consistency view

form ：

• View array
  • from All current uncommitted transactions id form
• Maximum transaction created id

{ [ Not submitted id1, Not submitted id2, Not submitted id3, Not submitted id4] max_id }

Be careful ：

Business id stay == Execute the first DML When the sentence is , Generation ,== Not in begin/start transaction Generate .

Version chain alignment rules

readview：{ [200],300 }

The rules ：

1. Execute query statement
2. The version chain is searched from top to bottom , If delete_flag == true, Don't return data . conversely ：
   • If trx_id Fall in the The green part (trx_id < min_id), Indicates that the version is generated by a committed transaction , so
   • If trx_id Fall in the The red part (trx_id > max_id), Indicates that the version is generated by uncommitted transactions , invisible
   • If trx_id Fall in the yellow (min_id < trx_id < max_id)
     • If trx_id non-existent View array （[200]） in , Indicates that the version is generated by a committed transaction , so
     • If trx_id There is View array （[200]） in , Indicates that the version is generated by uncommitted transactions , invisible

MVCC Read submitted

Realized MVCC Mechanism

characteristic ：

• read-view The array changes dynamically with each query , Realized Sure Read committed transaction modifications Characteristics of

MVCC Repeatable

Realized MVCC Mechanism

characteristic ：

• ==read-view After the array is generated, it is fixed ,== Achieved isolation .

BufferPool

BufferPool stay Engine layer from InnoDB Realization , be based on LRU Algorithm

Why design BufferPool？

Because no BufferPool, One request requires one disk Random, speaking, reading and writing I/O Update disk data , Poor performance ;

Although design BufferPool, There will be reading and writing to log files , but The log file is Sequential reading and writing I/O, Performance is better than random read / write I/O Much higher

One sql Implementation （ Service layer ）

One DMLsql Implementation （ Engine layer ）

1. From disk ibd Load cached data in the file
2. Write the old value to undo Log files 【 Version chain 、 Transaction commit failed, rollback 】
3. to update BufferPool Data in
4. Write redo Log cache
5. Write redo Log files 【 Transaction submitted successfully , Recover in case of system downtime BufferPool data 】
6. Write bin-log Log files 【 Transaction submitted successfully , Recover disk data in case of system downtime 】
7. Write commit Mark to redo Log files 【 Make sure redo and bin-log Uniformity 】
8. Background thread Random take BufferPool Modified data In pages Write to disk ibd file